Citation
A study in road geometry accuracy

Material Information

Title:
A study in road geometry accuracy
Creator:
Van Sickle, Jan L
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
xxiii, 440 leaves : ; 28 cm

Subjects

Subjects / Keywords:
Roads -- Design and construction ( lcsh )
Driver assistance systems ( lcsh )
Driver assistance systems ( fast )
Roads -- Design and construction ( fast )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Bibliography:
Includes bibliographical references (leaves 431-440).
Thesis:
Civil engineering
General Note:
Department of Civil Engineering
Statement of Responsibility:
by Jan L. Van Sickle.

Record Information

Source Institution:
|University of Colorado Denver
Holding Location:
|Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
438857102 ( OCLC )
ocn438857102
Classification:
LD1193.E53 2009d V36 ( lcc )

Full Text
A STUDY IN ROAD GEOMETRY ACCURACY
by
Jan L. Van Sickle
B.F.A., University of Nebraska, 1972
M.Eng., University of Colorado Denver, 2006
A thesis submitted to the
University of Colorado Denver
in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
Civil Engineering
2009


This thesis for the Doctor of Philosophy
degree by
Jan L. Van Sickle
has been approved
by
Kumar Navulur


Van Sickle, Jan Lee (Ph.D., Civil Engineering)
A Study in Road Geometry Accuracy
Thesis directed by Professor Lynn Johnson
ABSTRACT
The subject of this study is the mapping that supports Advanced Driver Assistance
System, ADAS and the verification of its positional accuracy. There are many
components of ADAS, but its foundation is digital base mapping built on three-
dimensional road geometry. This information directly supports in-vehicle subsystems
by presenting a model of the vehicles environment. It is also the model on which the
predictive capabilities anticipated by ADAS rely. There are a variety of them. They
include: adaptive cruise control, curve anticipation, predictive lighting, transmission
assistance, collision avoidance, intelligent speed adaptation, stability control, lane
departure warning and more. All depend on the positions of road centerlines, curve
radii, slope and etc. These features must be correct within close limits on the digital
base mapping for ADAS to be successful. ADAS is at an intersection of commerce
and societal good such mapping will mark a step toward reducing the leading cause of
death for people ages 2 to 34 motor vehicle crashes.
Signed
Lynn Johnson


ACKNOWLEDGEMENT
I wish to acknowledge the assistance of Tele Atlas, i-cubed, Stantec, Trimble and
Intermap in the development of this dissertation. I want to thank Tele Atlas for their
permission to use of the portions of the data and illustrations noted in §3.0
Background on Methods and § 5.0 Background on Methods that were developed
during our work in Orange County CA. Jay Clark deserves a special thanks for all his
advice and assistance. Krzysztof Miksa was also instrumental in the work as was
Shingo Ikeda. I also want to thank i-cubed for their assistance in the block
adjustment and orthorectification described in § 6.0 Bundle Block Adjustment and
Orthorectification. I particularly want to acknowledge Mick Garret and Karl Perrault
for their help. And I wish to acknowledge the help of Geoffrey Kirk of Trimble.
Stantec Consulting Inc. generously provided a grant for the LiDAR data collection
that was necessary for the data used in § 7.0 Validation. I want to thank Curt
Chapman, Spencer OBryan, Motaz Mostafa, Gustavo Rios, John Koepeke, Paul
Allen and Ramon Cortez particularly. I also want to acknowledge Keith Tennant of
Intermap for his kind permission to use some of the illustration noted in §7.1.1 Four
Accuracy Trials.
Finally I wish to thank my advisor, Dr. Lynn Johnson for his contributions and
support for my research. I also wish to thank all the members of my committee for
their valuable participation and insights.


TABLE OF CONTENTS
Figures......................................................................xi
Tables......................................................................xix
Chapter
1. Introduction...............................................................1
1.1 Purpose of the Dissertation..............................................2
1.1.1 The Scope of the Mapping..............................................3
1.1.2 Current Specifications................................................3
1.1.3 Future Specifications.................................................4
1.2 Rationale................................................................5
1.2.1 Safety................................................................6
1.2.2 Accuracy..............................................................8
1.3 Originality and Scope of the Dissertation................................9
2. Hypotheses................................................................13
2.1 First Hypothesis: Horizontal Absolute Accuracy..........................13
2.1.1 Hypothesis.............................................................13
2.1.2 Methodology............................................................13
2.2 Second Hypothesis: Horizontal Relative Accuracy........................15
v


2.2.1 Hypothesis
15
2.2.2 Methodology...........................................................15
2.3 Third Hypothesis: Vertical Absolute Accuracy..........................16
2.3.1 Hypothesis............................................................16
2.3.2 Methodology...........................................................17
2.4 Fourth Hypothesis: Vertical Relative Accuracy..........................17
2.4.1 Hypothesis............................................................17
2.4.2 Methodology...........................................................18
2.5 Fifth Hypothesis: Statistical Analysis of Road Geometry Accuracy......18
2.5.1 Hypothesis............................................................18
2.5.2 Methodology...........................................................19
3. Background on Methods....................................................23
3.1 Experimental Design....................................................23
3.1.1 AOI...................................................................23
3.1.2 Imagery...............................................................24
3.1.3 Bundle Block Adjustment...............................................26
3.1.4 Orthorectification....................................................29
3.1.5 GCPs..................................................................31
3.1.6 CPs...................................................................33
3.1.7 Control Processing....................................................34
3.1.8 Road Stratification and Sampling......................................35
vi


3.1.9 Statistical Analysis....................................................36
3.1.10 Nominal Results.......................................................37
3.2 The Test Methods.........................................................39
3.2.1 Monoscopic Aerial Photography...........................................39
3.2.2 Aerial Film to Digital...................................................40
3.2.3 Block Adjustment and Orthorectification..................................41
3.2.4 Pixels and Resolution...................................................43
3.3 Digital Elevation Models..................................................45
3.3.1 Accuracy of the DEM.....................................................49
3.3.2 Shuttle Radar Topography Mission, SRTM...................................50
3.3.3 National Elevation Dataset..............................................54
3.3.4 NED-SRTM Comparison.....................................................56
3.3.5 DEM Derived from Stereo Satellite Imagery................................57
3.3.6 Intermaps NEXTMap Product..............................................59
3.3.7 DEM error Contribution to Horizontal Error..............................59
3.4 The Collection of Ground Control Points..................................64
3.4.1 Distribution of GCPs.....................................................64
3.4.2 Mobile Mapping Vehicles, MMV.............................................67
3.4.3 Mobile Mapping Van GCP Collection - An Example......................102
3.5 The Collection of Check Points...........................................107
vii


3.5.1 Check Points
107
3.5.2 Verification............................................................107
3.6 Terrestrial LiDAR.........................................................122
3.6.1 LiDAR Technology.........................................................123
3.6.2 LiDAR Application.......................................................125
3.6.3 Use of the Ground Truth 3D Model........................................126
4. Background on Accuracy.....................................................127
4.1 Horizontal Accuracy.......................................................127
4.1.1 Accuracy and Precision...................................................127
4.1.2 Relative and Absolute Accuracy in ADAS..................................130
4.2 Vertical Accuracy.......................................................134
4.2.1 Ellipsoidal and Orthometric Heights......................................136
4.2.2 ADAS and Slope..........................................................146
5. Test Data Collection and Processing........................................153
5.1 Area of Interest..........................................................154
5.1.1 General Character of the AOI.............................................154
5.1.2 Corner Coordinates of the AOI...........................................155
5.2 Data Collection ........................................................155
5.2.1 Imagery .................................................................155
5.2.2 Collection of GCPs for the Test..........................................158
viii
5.2.3 Collection of CPs for the Test
165


5.2.4 Collection of LiDAR for the Test.........................................168
5.3 Processing.................................................................175
5.3.1 Control...................................................................176
6. Bundle Block Adjustment and Orthorectification..............................181
6.1 First Bundle Block Adjustment and Orthorectification.......................183
6.1.1 The BEM derived GCPs and their Use.......................................184
6.1.2 Modification of the DEM..................................................192
6.1.3 Three Results............................................................193
6.1.4 Horizontal Absolute Accuracy of First Bundle Block Adjustment and
Orthorectification.........................................................200
6.1.5 Evaluation of the First Bundle Block Adjustment and Orthorectification..203
6.2 Second Bundle Block Adjustment and Orthorectification......................218
6.2.1 The New BEM derived GCPs and their Use....................................218
6.2.2 Two Results..............................................................225
6.3 Horizontal Absolute Accuracy from Second Bundle Block Adjustment and
Orthorectification.........................................................229
6.4 Vertical Accuracy .........................................................232
6.4.1 TwoDEMs...................................................................232
6.4.2 Collection of Heights by the MMV.........................................241
6.4.3 Vertical Absolute Accuracy of the Orthorectified Imagery..........243
7. Validation..................................................................244
IX


7.1 Two Centerlines
244
7.1.1 Four Accuracy Trials..................................................246
7.1.2 Comparison............................................................257
7.2 Statistics............................................................264
7.2.1 A Horizontal Absolute Accuracy Dataset................................265
7.2.2 A Horizontal Relative Accuracy Dataset................................283
7.2.3 A Vertical Absolute Accuracy Dataset..................................291
7.2.4 A Vertical Relative Accuracy Dataset..................................298
7.2.5 Compilation...........................................................302
8. Conclusion...............................................................366
9. Recommendations..........................................................373
10. Review of Literature....................................................377
Appendix
A. Accuracy and RMSE.......................................................386
B. Fifteen Road Segments in Orange County..................................389
C. Instrument Specifications...............................................421
Bibliography................................................................431
x


LIST OF FIGURES
Figure
3.1 Experimental Design.....................................................25
3.2 Bundle Block Adjustment.................................................28
3.3 Relief Displacement.....................................................30
3.4 Block Adjustment........................................................42
3.5 Potential GCPs and CPs, Image Resolution and Photo-
Identifiability........................................................44
3.6 Profile View of DEM Errors Contribution to Ortho Error.................47
3.7 Satellites Pitch and Roll to Image......................................48
3.8 Single Pass Stereo Collection...........................................57
3.9 Off-Nadir Angle in Aerial 9x9 Film Cell...............................63
3.10 GCP Distribution.......................................................65
3.11 Single Receiver -Autonomous GPS Positioning............................74
3.12 2nd Alternative Relative GPS Positioning.............................77
3.13 OmniSTAR HP Subscription DGPS..........................................80
3.14 The NGS CORS Base Stations as of November 2006.......................... 83
xi


3.15 Post-Processing CORS Data.............................................84
3.16 The Urban Canyon Problem..............................................95
3.17 BEM in Miami...........................................................105
3.18 Example GCP............................................................106
3.19 Checkpoints...........................................................114
3.20 An Image Chip of a Check Point........................................114
3.21 Potential CPs..........................................................119
3.22 Example CP Sketch Sheet................................................120
3.23 Example CP Photographs.................................................121
4.1 Accuracy and Precision.................................................128
4.2 EDMaps Absolute and Relative Accuracy for ADAS........................132
4.3 An Ellipse.............................................................137
4.4 Biaxial Ellipsoidal Model of the Earth.................................137
4.5 Ellipsoidal Height.....................................................138
4.6 The Geoid and Mean Sea Level...........................................140
4.7 Orthometric Height.....................................................142
4.8 Slope Gradient.........................................................146
4.9 Typical Vertical Curve.................................................148
4.10 Conic Section..........................................................149
5.1 Area of Interest.......................................................154
5.2 Left Image Right Image ..............................................157
xii


5.3 The MMV Route......................................................160
5.4 An Example of a Static GPS Collect of Redundant ORAGCP.............164
5.5 An Example of Static Collection of CP ORANGE003....................166
5.6 Twenty-One Check Points, CPs across the AOI........................167
5.7 1 mph to 15 mph class roads Segment in Quadrangle 6..............171
5.8 45 mph to 55 mph class roads Segment in Quadrangle 13............172
5.9 65 mph class roads Segment in Quadrangle 4.......................172
5.10 25-35 mph class roads Segments in Quads 1-3, 5, 7-12 and 14-
18................................................................173
5.11 25-35 mph class roads Segment 11 in Quadrangle 14................174
5.12 25-35 mph class roads 0+00 Station Segment 11 in Quadrangle
14................................................................175
6.1 30-Frames of Monoscopic Aerial Imagery.............................183
6.2 AOI with Ortho-Mosaic of 30 Frames and the 6 BEM derived
GCPs..............................................................185
6.3 BEM001 on Aerial Imagery...........................................187
6.4 BEM006 on BEM and on Aerial Imagery................................188
6.5 NED Unmodified (Left) and Modified NED (Right).....................192
6.6 GCP BEM003 on 0.50m Resolution Imagery.............................209
6.7 Red Points: Trimble Track & Novatel/IMU Track agree within
0.50 meter........................................................213
6.8 3 BEM derived GCPs where MMV tracks agree within 0.50
meter.............................................................214
xiii


6.9 3 BEM derived GCPs where MMV tracks do not agree within
0.50 meter............................................................218
6.10 AOI with Ortho-Mosaic of 30 Frames and 22 BEM derived
GCPs..................................................................219
7.1 Horizontal Absolute Trial (stations 0+00 to 1+20)......................247
7.2 Horizontal Absolute Trial (stations 1+20 to 2+40)......................248
7.3 Horizontal Absolute Trial (stations 2+40 to 3+00)......................249
7.4 Horizontal Relative Trial (stations 0+00 to 1+20)......................251
7.5 Horizontal Relative Trial (stations 1+20 to 2+40)......................252
7.6 Horizontal Relative Trial (stations 2+40 to 3+00)......................253
7.7 Vertical Curve (stations 0+00 to 3+00).................................255
7.8 Vertical Trials (stations 0+00 to 3+00)................................256
7.9 0+00 Station of Quadrangle 3 Segment 3.................................260
7.10 Quadrangle 3 Segment 3................................................262
7.11 Quadrangle 12 Segment 3...............................................263
7.12 Histogram of Horiz. Absolute Error Segment 10 Original................268
7.13 Normal Quantile Plot of Horiz. Absolute Error Segment 10
Original..............................................................269
7.14 Resampling............................................................274
7.15 Histogram of Horiz. Absolute Error Segment 10 Bootstrap...............275
7.16 Normal Quantile Plot of Horiz. Absolute Error Segment 10
Bootstrap.............................................................276
xiv


7.17 Calc, of Horiz. Absolute RMSE Segment 10 Original...................281
7.18 Histogram of Horiz. Relative Error Segment 10 Original..............285
7.19 Normal Quantile Plot of Horiz. Relative Error Segment 10
Original...........................................................286
7.20 Histogram of Horiz. Relative Error Segment 10 Bootstrap............288
7.21 Normal Quantile Plot of Horiz. Relative Error Segment 10
Bootstrap...........................................................289
7.22 Histogram of Vert. Absolute Error Segment 10 Original...............293
7.23 Normal Quantile Plot of Vert. Absolute Error Segment 10
Original............................................................294
7.24 Histogram of Vert. Absolute Error Segment 10 Bootstrap..............295
7.25 Normal Quantile Plot of Vert. Absolute Error Segment 10
Bootstrap...........................................................296
7.26 Histogram of Vert. Relative Error Segment 10 Original...............299
7.27 Normal Quantile Plot of Vert. Relative Error Segment 10
Original............................................................300
7.28 Separately Compiled Horizontal Absolute RMSE........................304
7.29 Profile View of Quadrangle 14, Segment 11...........................310
7.30 Plan View of Quadrangle 14, Segment 11.............................311
7.31 Combined Aggregation Horizontal Absolute RMSE.......................314
7.32 25mph-35mph Histogram of Horiz. Absolute Error Bootstrap...........318
7.33 25mph-35mph Histogram of Horiz. Relative Error Bootstrap............320
7.34 25mph-35mph Histogram of Vert. Absolute Error Bootstrap.............322
xv


7.35 25mph-35mph Histogram of Vert. Relative Error Bootstrap.......324
7.36 lmph-15mph Histogram of Horiz. Absolute Error Bootstrap.......330
7.37 lmph-15mph Histogram of Horiz. Relative Error Bootstrap........332
7.38 lmph-15mph Histogram of Vert. Absolute Error Bootstrap.........334
7.39 lmph-15mph Histogram of Vert. Relative Error Bootstrap.........335
7.40 Profile View of Quadrangle 13, Segment 1......................339
7.41 Plan View of Quadrangle 13, Segment 1..........................340
7.42 45mph-55mph Histogram of Horiz. Absolute Error Bootstrap......343
7.43 45mph-55mph Histogram of Horiz. Relative Error Bootstrap.......345
7.44 45mph-55mph Histogram of Vert. Absolute Error Bootstrap........347
7.45 45mph-55mph Histogram of Vert. Relative Error Bootstrap........348
7.46 Distortion and Deviation Between Imagery Derived Centerlines..351
7.47 Profile View of Quadrangle 4, Segment 1.......................353
7.48 Plan View of Quadrangle 4, Segment 1...........................354
7.49 65mph Histogram of Horiz. Absolute Error Bootstrap.............358
7.50 65mph Histogram of Horiz. Relative Error Bootstrap.............360
7.51 65mph Histogram of Vert. Absolute Error Bootstrap..............362
7.52 65mph Histogram of Vert. Relative Error Bootstrap..............364
B.l Quadrangle 6 Segment 1 0+00 Station Detail.....................389
B.2 Quadrangle 6 Segment 1 Novatel.................................390
B.3 Quadrangle 6 Segment 1 Novatel and Trimble.....................391
xvi


B.4 Quadrangle 13 Segment 1 3+00 Station...............................392
B.5 Quadrangle 13 Segment 1...........................................393
B.6 Quadrangle 4 Segment 1 0+00 Station................................394
B.7 Quadrangle 4 Segment 1.............................................395
B.8 Quadrangle 1 Segment 1 0+00 Station................................396
B.9 Quadrangle 1 Segment 1 Novatel.....................................397
B.10 Quadrangle 1 Segment 1 Trimble.....................................398
B. 11 Quadrangle 2 Segment 2 Detail 0+00................................399
B.12 Quadrangle 2 Segment 2 Novatel....................................400
B.13 Quadrangle 3 Segment 3 0+00 Station...............................401
B.14 Quadrangle 3 Segment 3 Novatel....................................402
B.15 Quadrangle 5 Segment 5 0+00 Station...............................403
B.16 Quadrangle 5 Segment 5 Novatel....................................404
B.17 Quadrangle 7 Segment 5 0+00 Station...............................405
B.18 Quadrangle 7 Segment 5 Novatel....................................406
B.19 Quadrangle 8 Segment 7 0+00 Station...............................407
B.20 Quadrangle 8 Segment 7 Novatel....................................408
B.21 Quadrangle 9 Segment 7 0+00 Station...............................409
B.22 Quadrangle 9 Segment 7 Novatel....................................410
B.23 Quadrangle 10 Segment 8 0+00 Station..............................411
XVII


B.24 Quadrangle 10 Segment 8 Novatel...................................412
B.25 Quadrangle 11 Segment 9 0+00 Station..............................413
B.26 Quadrangle 11 Segment 9 Novatel...................................414
B.27 Quadrangle 12 Segment 10 0+00 Station.............................415
B.28 Quadrangle 12 Segment 10 Novatel..................................416
B.29 Quadrangle 12 Segment 10 Trimble..................................416
B.30 Quadrangle 14 Segment 11 0+00 Station.............................417
B.31 Quadrangle 14 Segment 11 Novatel..................................418
B.32 Quadrangle 15 Segment 12 0+00 Station.............................419
B.33 Quadrangle 15 Segment 12 Novatel..................................420
xviii


LIST OF TABLES
Table
1.1 The evolution of ADAS map accuracy.......................................4
3.1 Comparison of NED and SRTM DEMs..........................................56
3.2 Off-nadir angles and horizontal error...................................60
3.3 Example GGA NMEA string.................................................87
3.4 Denver survey-grade test January 8, 2007..............................97
4.1 10 %o = 1 percent........................................................146
5.1 Comers of the area of interest...........................................155
5.2 Proportional calc, of 300m segments based on MPH road classes...........170
5.3 Continuously operating reference station control........................177
5.4 21 Checkpoints..........................................................179
5.5 10 Redundant GCPs.......................................................180
6.1 Novatel BEM first bundle block adjustment residual report..............190
6.2 Trimble BEM first bundle block adjustment residual report..............191
6.3 Control orthorectification results April 2008........................... 196
6.4 Trimble BEM orthorectification results April 2008.......................197
xix


6.5 Novatel BEM orthorectification results April 2008................... 199
6.6 Trimble BEM CPs that exceed 95% limit April 2008..................................................202
6.7 Novatel BEM CPs that exceed 95% limit April 2008..................................................212
6.8 BEM positions compared with GPS/GNSS at all redundant GCPs........................................207
6.9 Redundant GCPs Trimble and Novatel tracks agree within 0.50m......................................215
6.10 Redundant GCPs Trimble and Novatel tracks disagree 0.50m or more...216
6.11 Novatel BEM second bundle block adjustment residual report.......................................221
6.12 Trimble BEM second bundle block adjustment residual report.......................................223
6.13 Trimble BEM orthorectification results October 2008..............................................226
6.14 Novatel BEM orthorectification results April 2008................................................228
6.15 Trimble BEM CPs that exceed 95% limit October 2008...............................................231
6.16 Novatel BEM CPs that exceed 95% limit October 2008...............................................232
6.17 Comparison of GPS/GNSS heights & NEXTMap DEM heights.............................................234
6.18 Comparison of GPS/GNSS heights and NED DEM heights...............................................235
6.19 GPS/GNSS heights compared w/NED DEM heights on roadways..........................................236
6.20 GPS/GNSS heights compared w/NEXTMap heights on Roadways...237
6.21 GPS/GNSS heights compared w/NED DEM heights off roadways..238
6.22 GPS/GNSS heights compared w/ NEXTMap heights off roadways.239
6.23 Heights collected by MMV...........................................241
6.24 Range of height on multiple MMV passes.............................242
7.1 Horizontal absolute accuracy dataset.................................267
xx


7.2 Segment 10 horiz. absolute results..............................276
7.3 Horizontal relative accuracy dataset............................283
7.4 Segment 10 horiz. relative results..............................289
7.5 Vertical absolute accuracy dataset..............................291
7.6 Segment 10 vert, absolute results...............................297
7.7 Vertical relative accuracy dataset ellipsoidal heights........298
7.8 Vertical relative accuracy dataset orthometric heights........301
7.9 25mph-35mph road class compiled results.........................305
7.10 25mph-35mph road class horizontal relative accuracy............307
7.11 25mph-35mph road class vertical absolute accuracy..............308
7.12 25mph-35mph road class vertical relative accuracy..............312
7.13 25mph-35mph road class aggregated results......................315
7.14 25mph-35mph horiz. absolute results............................319
7.15 25mph-35mph horiz. relative results............................321
7.16 25mph-35mph vert, absolute results.............................322
7.17 25mph-35mph vert, relative results.............................324
7.18 lmph-15mph horiz. absolute results.............................325
7.19 1 mph-15mph road class horizontal relative accuracy............326
7.20 lmph-15mph road class vertical absolute accuracy...............326
7.21 1 mph-15mph road class vertical relative accuracy..............327
7.22 lmph-15mph road class aggregated results.......................328
xxi


7.23 1 mph-15mph horiz. absolute results............................331
7.24 lmph-15mph horiz. relative results.............................333
7.25 lmph-15mph vert, absolute results..............................334
7.26 lmph-15mph vert, relative results..............................336
7.27 45mph-55mph horiz. absolute results............................337
7.28 45mph-55mph road class horizontal relative accuracy............337
7.29 45mph-55mph road class vertical absolute accuracy..............338
7.30 45mph-55mph road class vertical relative accuracy..............340
7.31 45mph-55mph road class aggregated results......................341
7.32 45mph-55mph horiz. absolute results............................344
7.33 45mph-55mph horiz. relative results............................346
7.34 45mph-55mph vert, absolute results.............................347
7.35 45mph-55mph vert, relative results.............................349
7.36 65mph horiz. absolute results..................................350
7.37 65mph road class horizontal relative accuracy..................352
7.38 65mph road class vertical absolute accuracy....................352
7.39 65mph road class vertical relative accuracy....................354
7.40 65mph road class aggregated results............................356
7.41 65mph horiz. absolute results..................................359
7.42 65mph horiz. relative results..................................361
7.43 65mph vert, absolute results...................................363
xxii


7.44 65mph vert, relative results..........................................365
7.45 Road class results....................................................367
7.46 Area of interest results..............................................368
C.l Crossbow IMU700CB- inertial measurement unit...........................422
C.2 Novatel Propak-LBplus-GPS receiver......................................423
C.3 Trimble SPS851-GPS receiver.............................................424
C.4 Trimble R8 GPS receiver..............................................426
C.5 Leica ScanStation 2 LIDAR scanner....................................427
xxiii


1. Introduction
Running late, a rushed driver in an unfamiliar city is about to make a mistake, but just
as he starts to swerve into an impossible lane change- the maneuver is aborted. And
instead of a collision, his car, his time and his health are preserved by an onboard
Advanced Driver Assistance System, ADAS. It notes the surrounding traffic, slows
the vehicle and returns him to his proper lane, and then provides him with an alternate
route so he will make his appointment on time and in one piece.
There are many components of ADAS, but its foundation is digital base mapping
built on three-dimensional road geometry. This information directly supports in-
vehicle subsystems by presenting a model of the vehicles environment. It is also the
model on which the predictive capabilities anticipated by ADAS rely. There are a
variety of them. They include: adaptive cruise control, curve anticipation, predictive
lighting, transmission assistance, collision avoidance, intelligent speed adaptation,
stability control, lane departure warning and more. All depend on the positions of
road centerlines, curve radii, slope and etc. These features must be correct within
close limits on the digital base mapping for ADAS to be successful.
1


The fundamental importance of accurate base mapping to ADAS was corroborated by
the authors of the Enhanced Digital Mapping (EDMap) Project Final Report which
included representatives from DaimlerChrysler, Ford, GM, NAVTEQ, Toyota
Technical Center and Crash Avoidance Metrics Partnership. In the report they
submitted to the United States Department of Transportation, the Federal Highway
Administration and National Highway Traffic and Safety Administration on
November 19, 2004 they wrote:
A digital map database and the associated navigation system are important to
the development of Advanced Driver Assistance Systems (ADAS). Digital
map navigation provides a connection between the vehicle and the roadway
infrastructure that is not possible with other ADAS sensors such as radar or
computer vision. While digital map navigation does not obviate the need for
other ADAS sensors, it serves as a necessary component in the development
of future driver assistance systems. (EDMap Team, 2004)
The achievement of just such digital road mapping and validation of its accuracy will
be the subject of this study.
1.1 Purpose of the Dissertation
While ADAS is the impetus for the mapping of road geometry, it is important to
mention here that the subject of this study is not ADAS itself nor its many interfaces.
The subject of this study is the mapping that supports ADAS, the scope of the
mapping and the verification of its positional accuracy.
2


1.1.1 The Scope of the Mapping
The creation and maintenance of worldwide digital road base maps are the purview of
two companies. They are Tele Atlas and Navteq. This is a condition the European
Commission calls a "duopoly market for navigable digital maps."(New Europe 2008.)
Tele Atlas's database covers approximately 14.6 million miles in 64 countries
(Sharma 2008) and Navteqs includes, data at various levels of detail for 69
countries on six continents, covering approximately 11 million miles of roadway
worldwide. (Reuters 2008). The scope is enormous.
1.1.2 Current Specifications
In light of the size of these databases the accuracy required is rigorous. Despite that,
the current requirements for ADAS compliant road centerline accuracy are not solidly
defined. This study will rely on the specifications presented by Mr. V. Blervaque in
his presentation at the ITS World Congress in Nagoya Japan. At that meeting Mr.
Blervaque also went further and theorized that the ADAS map accuracy specification
will be increased in the future, Road Accuracy digitized to ... lm relative. Absolute
accuracy (now 5m) to be 2m by 2012. (Blervaque 2004). The interpretation of these
standards here will be; absolute horizontal accuracy 5 meters, relative horizontal
3


accuracy 1 meter and slope accuracy 1%, all root-mean-square-error, RMSE. It will
also address vertical accuracy, both its absolute and relative components.
1.1.3 Future Specifications
It appears that Mr. V. Blervaques prediction was right. More recent information, as
shown in the slide (Table 1.1) taken from a presentation by Mr. Stephen TSiobbel of
Tele Atlas at the ITS World conference in London in 2006, indicates that ADAS
standards will increase significantly in upcoming years.
Table 1.1 The evolution of ADAS map accuracy (TSiobbel 2006)
Roadmap Requirements 2006 2008 2010 2012 2014 2016
Geometry Related Accuracy
Vehicle-position-accuracy (GPS, DGPS INS ) 7- 3 m +/- 2 m *>- 1.5 m *i-1 m +/-1 m +/- 1m
Absolute accuracy ** 2 4 m 2 4 m 1 5-3m 1 5-3m 1 -2m 0 5-15 m
Relafcve accuracy 1 -2m 05- 2m 06 1.5 m 0.5 1 rr 03-05m 0 3m
The application of such increasing accuracy standards to growing road databases that
already contain millions of miles is a daunting challenge. For example, Table 1.1
predicts that both the absolute and relative positional accuracies will be submeter.
However, some such as Meng Lu, et al. think that may be unrealistic as they wrote in
4


their paper Technical Feasibility of Safety Related Driving Assistance Systems
presented at the ICTCT conference in Helsinki in 2005:
As a conclusion it can therefore be said that sub-meter positioning
using satellite technology in moving vehicles seems difficult to
achieve. Use of a position with sub-meter accuracy would require a
map database of similar or better accuracy, of which the economical
feasibility vet has to be demonstrated [NextMAP Consortium 2000].
A proposed solution to cover satellite outages is the use of
pseudolites (local augmentation) [El-Rabbany 2002], but it is
questionable if this is cost-effective and useful if sub-meter level
positioning is not possible, (emphasis added)
On a multi-lane road such (a) system should provide lane
discrimination: in which lane the vehicle is, and where in that lane.
This would require a horizontal accuracy of about 0.3 m, which, as
stated before is difficult to achieve. If this would be achievable, also
a highly precise digital map would be required, of which the practical
and economical feasibility vet needs to be demonstrated, (emphasis
added) (Meng Lu et al. 2005)
The purpose of this study is the demonstration of just such feasibility.
1.2 Rationale
ADAS is at an intersection of commerce and societal good. The first to most
economically build a substantial body of ADAS compliant road geometry will
certainly realize an advantage in the marketplace. At the same time such mapping
will also mark a step toward reducing the leading cause of death for people ages 2 to
34 motor vehicle crashes. (Insurance Information Institute).
5


1.2.1 Safety
Some studies assert that human error is the cause of approximately 65% of motor
vehicle accidents (Dhillon 2007, 110) others suggest that human error is the sole
cause of only about 57% but a contributing factor in as many as 90% of the cases
(Treat et al. 1977). While the precise contribution of human error is in question,
drivers are certainly facing more distractions and crashes are too common. In the
United States in 2006, 42,642 people died in motor vehicle crashes and an additional
2,575,000 people were injured (NHTSA). The World Health Organization estimates
that 1.2 million people are killed and approximately 50 million injured worldwide in
road accidents every year. Estimates are that these figures will increase 65% in the
next 20 years unless prevention measures improve substantially. (Peden et al. 2004).
To do that there are several elements that need to be ameliorated.
Research has shown that accidents occur for one of three principle
reasons. The first is perceptual error. Sometimes critical information
was below the threshold for seeing the light was too dim, the driver
was blinded by glare, or the pedestrian's clothes had low contrast. In
other cases, the driver made a perceptual misjudgment (a curve's
radius or another car's speed or distance).
Drivers often misjudge road curvature, the speed of their own or
another vehicle, distance, etc.
6


Lastly, accidents sometimes occur because drivers accurately
perceive and interpret information but fail to respond appropriately
because they make the wrong decision or because they make the
right decision but perform the wrong response. (Green et al. 2004)
ADAS enabled navigation devices can help improve these situations in several ways.
For example, they can help minimize distractions due to complex intersections,
unfamiliar landmarks, road signs, lane choices and etc. In a recent article entitled
Driving Toward Crashless Cars, Steven Ashley writing in the Scientific American
listed these related key concepts:
Smart safety systems on todays high-end autos are taking ever
greater control from drivers to avoid collisions or, at least, reduce
injuries and fatalities. Within a few years, cars will steer clear of
accidents without any driver input at all. So called crashless cars
will emerge because of customer expectations of safety, government
pressure, crowded roads, an older, less capable population, and the
adoption of light-weight vehicles with less crashworthy structures.
Engineers have meanwhile demonstrated robotic vehicles. Together
with the crashless car, this development means that the driverless car
cannot be far off. (Ashley 2008, 86-92)
However, for these tools to be effective they must have a foundation of mapping that
contains sufficiently detailed and accurate information about the road itself.
7


1.2.2 Accuracy
It is important to note that the development and testing of ADAS applications
involves several stakeholders from design engineers to automotive manufacturers and
suppliers. Each must deal with the specification, realization and assessment of ADAS
technology in their own way. Starting in the earliest development stages overall
hazard analysis, risk assessment and controllability of safety relevant systems have to
be quantified. This is not possible without consideration of the underlying map data
accuracy.
The stakes for accuracy are especially high when it comes to preventative and active
safety applications. For example, a hazardous situation could result if an incorrect
map-based curve warning led to an erroneous system message and subsequent driver
over-reaction. Avoiding such outcomes starts with a clear understanding of the
accuracy statements of the mapping foundation, and that understanding should be
one that neither over- nor under-estimates the map datas integrity.
Further when error margins are as slim as they are in this realm, claims of accuracy
must be proven not assumed. It follows that clear, practical and reliable validation
procedures are required. Procedures by which accuracy statements are supported and
proven by measurements made in the field not extrapolated from ancillary datasets.
8


And it is necessary to consider methods that accommodate updates. Mapping data is
most often static but road alignments change.
1.3 Originality and Scope of the Dissertation
The originality of the work lies in the process described here and the problem to
which that process is addressed. In other words, the uniqueness is the arrangement of
its steps rather than the steps themselves. Also distinctive is its application to the
construction of mapping that complies with the accuracy specifications of Advanced
Driver Assistance, ADAS.
The scope of the study on which this dissertation is based includes the presentation of
procedures, processes and methods intended to achieve particular road centerline
accuracies. Specifically those accuracies mentioned in § 1.1.2 Current Specifications
are absolute horizontal accuracy 5 meters, relative horizontal accuracy 1 meter and
slope accuracy 1 %, all root-mean-square-error, RMSE.
A real-world test of these procedures, processes and methods was conducted over a
~ 430 sq. km area of interest, AOI beginning in August 2007 to December 2008. This
work will hereinafter be referred to as the Test. However, before the details of the
9


Test itself are discussed some background information on the methods used in the test
will be offered.
The methods used in the Test relied on imagery among other things. The imagery
involved was collected from an airborne platform in 2007 and also from vehicles
traveling on the AOI roadways themselves, also in 2007. Some topics pertinent to
airborne as well as spacebome platforms, bundle block adjustment and
orthorectification will be addressed in §3.1 Imagery.
The accuracy of the orthorectification of imagery is dependent on several elements;
Terrain relief and sensor orientation relative to that terrain cause satellite and aerial
imagery to not show features in their correct locations. Orthorectification transforms
the central projection of such imagery into an orthogonal view of the ground. In
other words orthorectification is the production of orthoimages, on which scale is
constant and accurate measurements of distance and direction on ground features,
such as road centerlines can be made. This process requires two data sets: the
parameters of image acquisition, the position and orientation of the camera at the
instant of collection, and the relief of the terrain represented by a digital elevation
model, DEM. The pertinence of this digital representation of the topographic
diversity, its effect not only the vertical aspect of the orthorectified imagerys
accuracy but also the horizontal will be presented in §3.2 Digital Elevation Models.
10


The quality of the orthorectification necessary to support the objective mapping
accuracy also requires ground control points, GCPs. That topic and how the required
GCPs were collected from the imagery captured by vehicles traveling on the
roadways themselves will be considered in §3.3 The Collection of Ground Control
Points.
As mentioned earlier when error margins are as slim as they are in this realm, claims
of accuracy must be not be assumed. Therefore, a necessary step toward validation of
road centerlines derived from imagery must be the verification of the accuracy of the
imagery itself. After orthorectification controlled by GCPs that verification relies on
check points, CPs. The presentation of the methods and procedures used to provide
those CPs for this study is found in §3.4 The Collection of Check Points.
Before consideration of the actual data from this study itself it is valuable to address
the differences around the concept of accuracy that are important to the meaning of
this study. For example, the difference between accuracy and precision is crucial.
The difference between the two components of accuracy itself absolute and relative
- is also of considerable consequence. This distinction is of interest because the
meanings of absolute and relative accuracy tend to vary with practice from discipline
to discipline and therefore the particular sense of these terms used here must be
defined. On close examination variation is also apparent in the definitions of vertical
11


and horizontal accuracies. Therefore, the two are each explained in detail in their
own sections. Horizontal accuracy is discussed in §4.1 Horizontal Accuracy and
vertical accuracy in §4.2 Vertical Accuracy.
The road centerlines derived from the orthorectified imagery which have been
databased are then compared with the same road centerlines as collected by direct
field survey measurement. The Test data and statistical analysis regarding the
outcome of this study are presented in §7. Testing and Validation.
The conclusions reached can be found in §8. Conclusion.
Finally an account of selected work that has been previously published on topics
pertinent to this study by accredited scholars and researchers is given in §9. Review of
Literature.
12


2. Hypotheses
2.1 First Hypothesis: Horizontal Absolute Accuracy
2.1.1 Hypothesis
Road centerlines with an absolute geometric accuracy of 5m horizontal RMSE,
HRMS can be achieved across ~ 430 sq. km on a foundation of off-the-shelf imagery
and elevation modeling, specifically 0.50m pixel orthorectified monoscopic aerial
imagery and the appropriate National Elevation Dataset, NED, derived digital
elevation model, DEM.
2.1.2 Methodology
The orthorectification that will provide the basis for the extraction of the road
centerlines to be databased will rely upon photo-identifiable ground control points,
GCPs. The GCPs used to register the aerial imagery will be harvested from data
captured via a mobile mapping vehicle, MMV. The vehicles position and orientation
while collecting that imagery will be determined by an Inertial Navigation System,
13


INS, containing an Inertial Measurement Unit, IMU, augmented with a real-time
corrected global positioning system, GPS, measurements. The INS and the real-time
corrected GPS will be integrated such that the GPS receivers will be the main position
sensors, while the IMU will be the main orientation sensor.
Verification of the accuracy of the orthorectification of the aerial imagery will be
done by coordinate comparisons at photo-identifiable check points, CPs. To quantify
the absolute horizontal error of the result the position of each CP as it appears on the
orthorectified aerial imagery will be compared with its post-processed position
collected by survey-grade static global positioning system/global navigation satellite
system, GPS/GNSS, observations.
Validation of the absolute accuracy of the road centerline subsequently extracted
from the orthorectified imagery will be based on direct field survey measurement.
Road segments will be sampled from the area of interest, AOI, quadrangle and
surveyed in the field. Each surveyed segment will be divided into equally spaced
stations augmented by stations at each deflection. Then the coordinate of the point at
each station defined in the databased road centerline will be compared the coordinate
of the same point as determined by the field survey.
14


2.2 Second Hypothesis: Horizontal Relative Accuracy
2.2.1 Hypothesis
A road centerline with a relative geometric accuracy of lm horizontal RMSE, HRMS,
can be achieved across the same ~ 430 sq. km utilizing off-the-shelf imagery and
elevation modeling, specifically 0.50m pixel orthorectified monoscopic aerial
imagery and the appropriate NED derived DEM.
2.2.2 Methodology
The orthorectification and harvesting of photo-identifiable GCPs will be done with
the process described in §2.1 First Hypothesis: Horizontal Absolute Accuracy.
However, data from an additional sensor will be added to the suite onboard the
MMV. That additional sensor will be a survey-grade GPS/GNSS receiver whose
measurements will be post-processed. The two independent GPS tracks will allow a
comparison between the position of the MMV as determined by the real-time
corrected GPS/IMU integration with the position of the MMV as determined by the
post-processed survey-grade GPS/GNSS. This comparison will allow the selection
of GCPs where the two positions are in agreement within the aerial imagery
resolution, 0.50 meter, and thereby improving the integrity of the orthorectification
process.
15


Verification of the accuracy of the orthorectification of the aerial imagery will also be
done by photo-identifiable check points, CPs, in the same manner described in §2.1
First Hypothesis: Horizontal Absolute Accuracy. The validation of the relative
accuracy of the road centerline subsequently extracted from the orthorectified
imagery will also be based on a field survey sampled segments as described in § 2.1
First Hypothesis: Horizontal Absolute Accuracy.
2.3 Third Hypothesis: Vertical Absolute Accuracy
2.3.1 Hypothesis
A road centerline with an absolute geometric accuracy of 4m vertical RMSE can be
achieved across the same ~ 430 sq. km utilizing off-the-shelf imagery and elevation
modeling. Specifically 0.50m pixel orthorectified monoscopic aerial imagery and
the appropriate NED derived DEM.
16


2.3.2 Methodology
The orthorectification, the harvesting of photo-identifiable GCPs will be done with
the process described in § 2.2 Second Hypothesis: Horizontal Relative Accuracy.
Verification of the accuracy of the orthorectification of the aerial imagery will also be
done by photo-identifiable CPs in the same manner described in §2.1 First
Hypothesis: Horizontal Absolute Accuracy. The validation of the relative accuracy of
the road centerline subsequently extracted from the orthorectified imagery will also
be based on a field survey sampled segments as described in §2.1 First Hypothesis:
Horizontal Absolute Accuracy.
2.4 Fourth Hypothesis: Vertical Relative Accuracy
2.4.1 Hypothesis
A road centerline with an relative geometric accuracy of 0.30m vertical RMSE over
the length of a full station, 30m, that is 1 %, can be achieved along road centerlines
across the same ~ 430 sq. km utilizing off-the-shelf imagery and elevation modeling.
Specifically 0.50m pixel orthorectified monoscopic aerial imagery and the
appropriate NED derived DEM.
17


2.4.2 Methodology
The orthorectification, the harvesting of photo-identifiable GCPs will be done with
the process described in § 2.2 Second Hypothesis: Horizontal Relative Accuracy.
Verification of the accuracy of the orthorectification of the aerial imagery will be
done by photo-identifiable CPs in the same manner described in §2.1 First
Hypothesis: Horizontal Absolute Accuracy. The validation of the relative accuracy of
the road centerline subsequently extracted from the orthorectified imagery will be
based on a field survey sampled segments as described in §2.1 First Hypothesis:
Horizontal Absolute Accuracy.
2.5 Fifth Hypothesis: Statistical Analysis of Road Geometry Accuracy
2.5.1 Hypothesis
Where road segments have been chosen from an AOI in which the road centerlines
are proportionally stratified and a re-sampling plan employed data from the
comparison of the position of stations as databased and as observed can yield an
estimate of mean, standard deviation and standard error of the mean that is reasonably
representative of the whole population.
18


2.5.2 Methodology
The goal of the sampling of the population of road centerlines in the AOI and the
methods for selecting and observing a part of that population is to make inferences
about the whole population. Toward that end, road segments chosen from the AOI in
which the road centerlines have been stratified will provide a foundation for such
inferences. Each stratum will be chosen in the same proportion as it occurs in the
whole population. In other words, the sampling fraction in each stratum is made
equal to the sampling fraction for the population. (Kish 1995, 83).
In each stratum the comparison described earlier will be conducted, specifically the
differences in position of stations as databased and as observed in the field will be
measured. Then a bootstrap resampling plan will be used to calculate an estimate of
the mean, standard deviation and standard error of the mean within each stratum. The
values for each stratum will be reported and then weighted to provide a combined
estimate for the whole population.
19


2.5.2.1 Stratified Sampling
In broad terms, stratified sampling consists of the following steps:
(a) The entire population of sampling units is divided into distinct
subpopulations, called strata.
(b) Within each stratum a separate sample is selected from all the
sampling units composing that stratum.
(c) From the sample obtained in each stratum, a separate mean (or
other statistic) is computed. These stratum means are properly
weighted to form a combined estimate for the entire population
(d) The variances are also computed separately within each stratum
and then properly weighted and added into a combined estimate for the
population.
(Kish 1995,75)
This plan of stratified sampling will be followed with the caveat that the calculation
of the mean, standard deviation and standard error of the mean of each of the stratum
will incorporate a bootstrap resampling plan.
2.5.2.2 Computation of Statistics for Each Stratum Sample
When the segments are chosen a direct field survey of the samples will be done and a
3D truth model of each segment constructed. Each of these stratum sample segment
models will be divided into 10 equally spaced stations augmented by stations at each
deflection. Then the horizontal coordinate and height at each of the stations defined in
the databased road centerline will be compared to the horizontal coordinate and
height of the same point as determined by the direct field survey. These comparisons
20


will provide the components from which the absolute and relative, vertical and
horizontal, RMSE, mean, standard deviation and standard error of the mean will be
calculated for each road segment sampled.
2.5.2.3 Calculation of Statistics via Resampling
Inferences will be drawn from the sampled road segments statistics. These statistical
inferences will be based on the sampling distributions of sampled road segments
statistics. However, they will be limited by the smallness of the sample relative to the
size of the stratum from which it comes. Therefore, the bootstrap resampling strategy
will be employed. The bootstrap provides a way to derive the sampling distribution
for a stratum from which a road segment originates, at least approximately, from that
sampled road segment itself. In other words, resamples from the sampled road
segment will be used to represent the distribution that would result if there were many
more samples taken from the stratum.
More specifically, to bootstrap a statistic of a road segment sample hundreds of
resamples will be drawn from it, with replacement. The statistic will be calculated for
each resample and the bootstrap distribution produced. This bootstrap distribution
approximates the sampling distribution of the statistic across the whole stratum from
21


which the road segment sample was selected. In this way the bootstrap distribution
estimates the variation in the statistic based on the original data.
22


3. Background on Methods
3.1 Experimental Design
A flow chart of the experimental design of the Test described in this dissertation is
offered here in Figure 3.1 to inform follow-on research and serve as a template for
future work. The steps represented are more fully described in what follows and
improvements in the methodology are suggested at appropriate intervals in italics.
3.1.1 AOI
An Area of Interest, AOI that contains -400 sq. km is chosen that has variable terrain
with significant elevation change, urban canyons, vegetative cover and other error
sources in order to tax the processes and methods.
23


3.1.2 Imagery
Off-the shelf imagery provides the basis for the drawing of the road centerlines by
heads-up digitizing. The road centerlines so developed are subsequently databased.
The process by which the imagery is prepared to support that work is dependent on
several elements.
These elements include field collected photo-identifiable ground control points, GCP
These are collected via post-processed as well as real-time GPS/GNSS on a Mobile
24


to


Mapping Vehicle, MMV. These points define the reference coordinate system or
datum on which the photogrammetry relies. ITRFOO is the recommended datum.
Ground control points are surveyed at higher accuracy than is expected of the
photogrammetric product.
Also collected in the field are photo-identifiable check points, CPs, via static post-
processed GPS/GNSS. Check points are surveyed at higher accuracy than is expected
of the bundle block adjustment.
These data will support the bundle block adjustment. And finally the
orthorectification which is supported by a digital elevation model, DEM
3.1.3 Bundle Block Adjustment
As shown in Figure 3.1, after the imagery is acquired the aerial negative film is
scanned with a high precision photogrammetric scanner. The scanned images are
imported in to a project with support files. Those files include initial estimates of
image positions, camera calibration, location of GCPs and etc. Next, image pyramid
layers are generated. The interior orientation is done. The aerotriangulation for the
bundle block adjustment is carried out by software.
26


Bundle block adjustment is a process that determines the exact location and
orientation of each photo in a block at the instant of exposure. To accomplish this it
relies on image rays. Those rays are shown in Figure 3.2 as the lines from the
photographs to the ground surface.
Each of these rays connects an object point, the perspective center of the image, and
the projection of the object point on the image. Each image is a bundle of these rays.
Using these bundles along with ground control information and tie points a bundle
block adjustment is a least squares error minimization technique to establish the
position and orientation, that is the exterior orientation, of all of the photographs in a
block simultaneously. In other words the bundle block adjustment determines the
three translations: Xo, Yo, Zoand the three rotations: to,(j),K required to bring all the
corresponding rays to intersection as nearly as possible at the tie points and the
control points.
27


Photographs
Figure 3.2 BUNDLE BLOCK ADJUSTMENT
Residuals in the bundle block adjustment are the difference between
where the GCP has been placed, and where the bundle adjustment
computes the position of that point to be. The bundle adjustment solves
for the best possible solution for the location of each image, using all
GCPs, TPs, and EO (exterior orientation) data available. The criterion
for this solution is that the sum of the square errors is minimized (least
squares), which means that no single GCP or TP will fit perfectly. The
residuals are the remaining shift in the computed position. Residuals are
not errors to be corrected, they simply help you to see which GCPs fit
well (low residuals) and which don't fit quite as well (higher residuals).
High residuals can tell you if a single GCP is severely out of place, or
more often, where parts of the block of imagery do not fit the ground
well. (PCI Geomatics.2009)
28


An improvement to the process as described in the Test would be to validate the
accuracy of the bundle block adjustment with the check points because they were not
used as control in the solution. This step was skipped in the Test work. In the Test the
final ortho imagery was compared with the check points. While this in some sense
validates the final photogrammetric product of the process it does not separate the
errors attributable to the bundle block adjustment from those attributable to the
orthorectification and the digital elevation model.
3.1.4 Orthorectification
As shown in Figure 3.3 terrain relief and sensor orientation relative to that terrain
cause satellite and aerial imagery to not show features in their correct locations.
Orthorectification transforms the central projection of such imagery into an
orthogonal view of the ground. In other words orthorectification is the production of
orthoimages, on which scale is constant and accurate measurements of distance and
direction on ground features, such as road centerlines, can be made. This process
requires two data sets: the parameters of image acquisition, the position and
orientation of the camera at the instant of collection, and the relief of the terrain
29


represented by a digital elevation model, DEM. Both the horizontal and the vertical
accuracy of the orthorectification are dependent on the DEM.
Perspective
Center
.Relief
Displacement
Figure 3.3 RELIEF DISPLACEMENT
(Adapted from Mikhail et al. 2001, 235)
Improvement: It is unusual that the National Elevation Dataset, NED, digital
elevation model used in the Test met the 2.2m criteria. The expected vertical
30


accuracy from the 10m posting NED DEM, such as is used is -7m to -15m. This
typical vertical accuracy would certainly not support work that would satisfy the
ADAS horizontal or vertical accuracy specifications. However, a superior digital
elevation model would improve both the horizontal and vertical accuracy results. A
DEM developed from stereo satellite scenes, stereo aerial photography of the AOI or
other sources in future work may make it possible to meet both the ADAS relative and
absolute- horizontal and vertical accuracy standards.
Improvement: in the Test the first bundle block adjustment and orthorectification was
done with too few GCPs and the results were disappointing. The processing was
repeated with 22 GCPs and the results were improved by approximately a third. It
would be better to perform one bundle block adjustment and orthorectification with
adequate GCPs the first time. Also, tie points could be used to validate the DEM for
the project area.
3.1.5 GCPs
One of the unusual aspects of this work is that the GCPs are harvested from a mosaic
of imagery captured by a mobile mapping vehicle, MMV. This mosaic is known as
the Birds Eye Mosaic, BEM in the Tele Atlas lexicon. Because the MMVs position
and orientation is determined in real-time through a suite of on-board sensors the
31


BEMs are used to extract from the images the x, y and z coordinates of photo-
identifiable GCPs with great flexibility. They are chosen in the pavement of the
road, in areas of high contrast, flush with the ground, stable, unobstructed by
vegetation or shadows and be more than lm in size.
The vehicles position and orientation while collecting that imagery is determined by
an Inertial Navigation System, INS, containing an Inertial Measurement Unit, IMU,
augmented with a real-time corrected global positioning system, GPS, measurements.
The INS and the real-time corrected GPS are integrated such that the GPS receiver is
the main position sensors, while the IMU is the main orientation sensor. It is
important to note that the data from the GPS receiver used in the MMV data stream is
not post-processed. However, an additional sensor is added to the suite onboard the
MMV. That additional sensor is a survey-grade GPS/GNSS receiver whose
measurements are post-processed with data from Continuously Operating Reference
Stations, CORS.
Because these are two satellite navigation signal receivers, two different Birds-Eye
Mosaic, BEM datasets can be produced and from these two sets of GCPs can be
derived, one based on GPS/IMU corrected by a real-time subscription service and
another base on post-processed data from the GPS/GNSS receiver. However, the
additional survey-grade receiver is not intended to provide a separate, continuous
32


source of GCPs. Its purpose is to provide redundancy for the MMV measurements
and thereby increase the reliability MMV track. In other words, because the two
sensors sets are operating simultaneously on the circuit when it is driven, the main
objective is to augment and support the other MMV sensors at specific points on the
route. This facility is most useful at points chosen as GCPs.
One way of improving the integrity of the chosen GCPs is to combine the survey-
grade post-processed MMV track and the real-time corrected/IMU MMV track on the
circuit, compare the positions along the route driven and determine where the two
tracks- one with the survey-grade receiver and one without agree within close
limits, say Vi m. Then choose GCPs in the area where the two tracks are in such close
agreement.
Improvement: A better approach is to generate not two different BEM datasets, but
rather one. And, instead of choosing GCPs where two separate MMV tracks closely
agree, limit the selection of GCPs to those areas where the precision dilution of
precision PDOP is low and the signal to noise ratio SNR is high and enhance the
integrity of the selection of GCPs.
Improvement: GCPs can be used to validate the DEM in the project area.
33


3.1.6 CPs
Verification of the accuracy of the results of the processing of the imagery is done by
coordinate comparisons at photo-identifiable check points, CPs. To quantify the
errors of the result the position of each CP as it appears on the imagery is compared
with its post-processed position collected by survey-grade static GPS/GNSS.
Twenty check points are required. Regarding the distribution of the collected CPs
across the AOI in the Test, the FGDC report: Geospatial Positioning Accuracy
Standards, Part 3: National Standard for Spatial Data Accuracy, Subcommittee for
Base Cartographic Data, Federal Geographic Data Committee, FGDC-STD-007.3-
1998 specifies: A minimum of 20 check points shall be tested, distributed to reflect
the geographic area of interest and the distribution of error in the dataset.
(FGDC 1998, 7)
3.1.7 Control Processing
Both the GPS\GNSS data collected by static observations on the GCPs and CPs as
well as the survey-grade Mobile Mapping Vehicle, MMVdata are post-processed.
It is important to point out that both sets of data are post-processed based on exactly
the same set of Continuously Operating Reference Stations. It is also important to
explain that the coordinates of the control are in all cases expressed in the datum
34


WGS84 (G1150) which is virtually coincident with ITRF00 (2002.0). This is an
international datum that obviates the need to transform work from datum to datum
around the world and provides consistency throughout the database.
3.1.8 Road Stratification and Sampling
Road segments are chosen from the AOI road classes into which the road centerlines
have been stratified. The recommended road classes are based on miles-per-hour,
mph as a convenient representation of differences in design characteristics. Each
stratum is chosen in the same proportion as it occurs in the whole population.
After stratification into classes, road segments are sampled from the area of interest,
AOI, quadrangle from sub-quadrangles into which it has been subdivided. The road
samples are chosen where the Mobile Mapping Vehicle has not been driven and are
chosen a minimum of 1 km distant from the roads which have been collected by the
MMV. The samples are selected to included curves, overhead obstructions, urban
canyons, changes in height and other error sources.
Each sampled road segment is surveyed in the field by terrestrial LiDAR scanning.
Each sampled road segment is divided into equally spaced stations augmented by
35


stations at each deflection in the alignment drawn from the roads representation on
the ortho imagery. Then the coordinate of the point at each station is also defined in
the databased road centerline that is created from heads-up digitizing on the ortho
imagery and subsequently compared to the coordinate of the same point as
determined by the LiDAR field survey. Validation of the accuracies of the road
centerline subsequently extracted from the processed imagery is based on acceptance
of direct LiDAR field survey measurement as ground truth.
Improvement: In the Test the 15 road segments samples were measured over 4 road
classes only 1% of the roads in the AOI were represented in this sample. A larger
sample would provide more robust results.
3.1.9 Statistical Analysis
In four accuracy trials, one for each of the following categories- horizontal absolute,
horizontal relative, vertical absolute and vertical relative each of the 15 road
segment samples across the 4 road classes are tested. The comparisons are between
the LiDAR derived centerlines and the ortho imagery center lines. The mean,
standard deviation, standard deviation of the mean and root mean squared error is
calculated for each 15 road segment samples in each of the four accuracy categories
36


individually. The mean, standard deviation and standard deviation of the mean of
these calculations are checked by bootstrap resampling.
Improvement: In the Test two datasets were statistically evaluated separately. The
dataset derived from the MMV GPS\IMU real-time sensors was considered
separately from the dataset derived MMV GPS\GNSS post-processed receiver. This
will be rendered unnecessary by the earlier mentioned improvement to limit the
selection of GCPs to those areas where the precision dilution of precision PDOP is
low and the signal to noise ratio SNR is high. In that way there will be only one
dataset to consider not two.
3.1.10 Nominal Results
Using the methodology above and the following resources:
Imagery -
0.50m pixel monoscopic RGB aerial photography captured with a calibrated
metric mapping camera with a wide-angle (90), a 6-inch focal length lens that
uses a 9 x 9 inch film format.
Digital Elevation Model -
37


10m posting and a vertical RMSE of 2.2m
Ground Control Points -
derived from mosaiced imagery captured by 6 calibrated 1280x980 color
digital cameras with 90 lenses synchronized by an electrical trigger signal
mounted on a Mobile MappingVehicle, MMV equipped with a Crossbow
IMU700CB, six degree-of freedom, 6DOF, IMU that uses third generation
Fiber Optic Rate Gyroscope technology with index positioning is provided by
an on board Novatel Propak LBplus GPS receiver and augmented by a
Trimble SPS851 survey-grade receiver whose data is post-processed.
Specifications for all instruments are given in Appendix C.
Check Points -
collected by static post-processed GPS\GNSS observations with a Trimble R8
receiver. Specifications for this instrument are given in Appendix C.
road centerline accuracy results within these ranges can be obtained:
Horizontal absolute accuracies RMSE from 1.50m to 3.00m
Horizontal relative accuracies RMSE from 0.40m to 0.50m
Vertical absolute accuracies RMSE from 1.20m to 5.50
38


Vertical relative accuracies RMSE from 2.00% to 4.70%.
3.2 The Test Methods
More detail on the methods and processes used the in the Test follow. Imagery is a
fundamental component of the data path street map vendors use to develop the
centerlines that populate their databases. Therefore it is important to minimize the
effect of both the systematic and random errors that are inevitable in its production.
3.2.1 Monoscopic Aerial Photography
The imagery available for the Test was off-the-shelf monoscopic 8-bit RGB aerial
photography. The positive aspects of this approach include the fact that film-based
aerial imagery is widely understood, it has been available for nearly a century and
original film resolution can be high though it might not be completely preserved
when scanned However, the outcome depends somewhat on the scanner's bit depth.
There is another advantage available to the capture of imagery with aerial cameras.
As described by Jacek Grodecki and Gene Dial in their paper, Block Adjustment of
High-Resolution Satellite Images Described by Rational Polynomials:
39


Owing to the dynamic nature of satellite image collection,
photogrammetric processing of satellite imagery is more complicated
than is aerial frame camera processing. Aerial cameras acquire the
entire image at an instant of time with a unique exposure station and
orientation. High-resolution pushbroom satellite cameras, including
Ikonos, use linear sensor arrays that acquire a single image line at an
instant of time. Consequently, each line of a pushbroom satellite
image has a difference exposure station and orientation. (Grodecki
and Dial 2003, 59)
Nevertheless, 11-bit panchromatic satellite imagery would be appropriate as well, and
may be used in future work. Some advantages it might offer are: its coverage is
worldwide, its contrast is more pronounced and, compared to aerial imagery, the
swaths are large and the strips long, from 10km to 200km. These qualities contribute
to consistency in content and lighting conditions which can facilitate feature
extraction, a characteristic that could be helpful in developing road centerlines. The
use of satellite imagery can also reduce the number of ground control, GCP and
check points, CP required for rectification. This is due in part to the fact that the
satellite Rational Polynomial Coefficient, RPC, sensor model has fewer adjustable
terms than an aerial sensor model.
3.2.2 Aerial Film to Digital
A digital version of the aerial film photography was generated by scanning the
original film negatives or film diapositives which produced ~0.50m pixel s. Next, the
orthorectification and mosaicking was done with a process known as block
40


adjustment, bundle block adjustment soft copy geopositioning or
aero333333triangulation a least squares error minimization process typically used to
solve both aerial and satellite image blocks.
3.2.3 Block Adjustment and Orthorectification
The tie points, TPs, are image points. They are photo-identifiable points located in
overlapping images and are derived from two or more contiguous strips of
photographs as illustrated by the red dots in the Figure 3.4. Figure 3.4 was adapted
from figures 3, 4 and 5 in Dr. Jacek Grodeckis article Satellite Photogrammetry:
Processing Innovations Maximize High Resolution Imagery Options. Regarding
satellite imagery Figure 3.4 illustrates the situation for a pushbroom satellite raster
model with a narrow field of view.
41


001 003 003_______-3/ _ _ 009 Oil OH
* i T"1 r r r
in 1 : i a
i < ! % 4' $ * % i G> <4
. ; ! j i
L. i ; i
002 004 006 006 010 012 014 000 001 002 003 004
Stereo
Mono
A single GCP improves accuracy of all stereo images but in the case
of a mono block it improves accuracy of only the image onto which
it falls (GCP is triangle/square; tiepoints with dots).
Legend:
0
Confidence Region
Mono
In a mono block with a single GCR a more uniform accuracy
improvement can be realized by adding a cross-strip.
Figure 3.4 BLOCK ADJUSTMENT
(Adapted from Grodecki. 2006)
Block adjustment is also mentioned, with some guidelines on GCPs, in the initial
report of the European Commission Directorate General JRC, Joint Research Centre,
Italy titled, Guidelines for Best Practice and Quality Checking of Ortho Imagery,
p. 17, November 20, 2003.
As an alternative to single frame processing, and if appropriate
software is available, multiple image frames or a block of images
- for the same zone can be processed together. The block processing
uses ground control points (GCPs) and tie points (points observed on
images but not on the ground), combined with sensor geometry to
42


calculate the best fit for all images together. It is not recommended to
use less than one GCP per image frame in the block.
(European Commission Directorate General JRC 2003, 17)
In stereo imagery each area is captured from two points of view, the improvement in
accuracy from GCPs propagates through the block, however, in the monoscopic
block, each area is captured from only one point of view and the GCP only improves
the accuracy of the image on which it falls. To achieve a consistently accurate
solution in a monoscopic block it is recommended to have a GCP in each one of the
frames.
Another alternative is the use of a crossing image as illustrated in Figure 3.4. When
monoscopic imagery is used a crossing image enhances the effectiveness of few
GCPs. This alternative was not available for the Test.
3.2.4 Pixels and Resolution
As mentioned earlier the digital imagery used in the Test had ~0.50m pixel s. In this
regard it is important to distinguish between accuracy and resolution; they are
certainly not the same. The size of the pixel as it relates to the resolution of an image,
while important, is entirely different from the positional accuracy of that image. The
positional accuracy is concerned, of course, with the relationship of the imagery to
43


the Earth. It can be confusing. For example, a product that is quoted as 0.50m
orthophoto may refer to either the pixel resolution of the image or the absolute spatial
position relative to the ground coordinate system. In this instance it is the pixel
resolution.
At this resolution it is problematic whether or not a 0.50m sized feature would be
distinguishable on the imagery. Photogrammetrists generally assert that the ground
resolution should typically be considered to be half the size of the smallest object to
be resolved. In other words, at best the smallest object that can be certainly identified
in imagery with a 0.50m pixel will be ~lm in size.
Figure 3.5 POTENTIAL GCPS AND CPS, IMAGE RESOLUTION AND PHOTO-
IDENTIFIABILITY
44


To illustrate the practical implication of this idea to this study please see the images
in Figure 3.5. The image on the left shows several CP locations. The image on the
right is an enlargement of the double walkways in the first image. In this close-up
the twin walkways contrast nicely with the grass but determining exactly which pixel
represents the edge of a comer between them is clearly difficult.
In short, higher image resolution increases certainty in GCP and CP photo-
identification, lower resolution decreases it. Hence resolution affects the level of
certainty in photo-identification which affects the level of horizontal accuracy
achievable through orthorectification using photo-identifiable check points and
ground control points.
3.3 Digital Elevation Models
A few words about the DEMs are in order not only to clarify their use in the Test but
also to point the way for future work around the world. A DEM is a digital
representation of the terrain absent the vegetation and man-made structures. This
absence is the primary distinction between a DEM and a DSM, Digital Surface
Model, which includes the canopy.
45


The DEM is the raster representation of the map created from those digital vectors
and their elevations. Alternatively, they can be semi-automatically extracted from
stereo satellite scenes. DEMs can also be produced from stereo aerial photography.
Some excellent work in DEM creation has been done with Interferometric Synthetic
Aperture, IFSAR, such as the Shuttle Radar Topography Mission, SRTM and the
NEXT Map product from Intermap. DEMs are also created over limited areas with
Light Detection and Ranging, LiDAR.
The alternative sources of DEMs described here are pertinent to this study because
the overall accuracy of orthorectified imagery is very dependent on the accuracy of
the DEM. The more topographically diverse the landscape, for example, the terrain
in the eastern portion of the AOI for this study, the more distortion that variation
contributes to the imagery, in such circumstances the importance of the DEM
increases.
The accuracy of the orthorectified imagery is also very dependent on the size of the
elevation angle of the line of sight from the Earth to the sensor. The best elevation
angle is 90, along the nadir of the sensor. At that angle inaccuracies the DEM
contributes are minimized. But as that elevation angle flattens, as it decreases, the
error that the DEMs inaccuracies contribute to the final ortho-photograph increases.
As is shown in Figure 3.6 which was adapted from figure 7 in Dr. Jacek Grodeckis
46


article Satellite Photogrammetry: Processing Innovations Maximize High Resolution
Imagery Options.
Ortho Error
Figure 3.6 PROFILE VIEW OF DEM ERRORS CONTRIBUTION TO ORTHO
ERROR
(Adapted from Grodecki 2006)
In other words, to maintain a particular accuracy standard the accuracy of the DEM
and the size of the elevation angle must be considered. Aerial camera work is most
affected by this phenomenon. Elevation angles on an aerial camera are a function of
the distance from the principal point. The elevation angles for particular camera
models are fixed, and given that aerial cameras, such as the Wild RC-30
47


used in this study, also have wide fields of view, i.e. 90 many of the resulting pixels
have a low elevation angle. This unfortunately increases somewhat the contribution
of the DEMs inaccuracies to the bias in the imagery.
Satellite Position II
Satellite Position I
High-resolution satellites are agile.
Elevation angles can be varied to
y meet the desired ortho accuracy.
Elevation 4I: More accuracy
Elevation a2: Less accuracy
Elevation .4 2
i_______________________________________________
Figure 3.7 SATELLITES PITCH AND ROLL TO IMAGE
(Adapted from Grodecki 2006)
Satellites, on the other hand, are agile and can pitch and roll to image an area. The
elevation angles can be controlled, as is shown in Figure 3.7 which was adapted from
figure 10 in Dr. Jacek Grodeckis article Satellite Photogrammetry: Processing
Innovations Maximize High Resolution Imagery Options.
48


It is therefore possible to specify the elevation angles to meet the desired ortho
accuracy. By considering the error inherent in the DEM in each case it is possible
then to extrapolate the limit of the elevation angles that will produce horizontal
accuracy within the desired range.
In any case it follows that improving the accuracy of the DEM improves the accuracy
of the resulting orthorectified imagery both vertically and horizontally.
3.3.1 Accuracy of the DEM
Commonly used DEM products include the less accurate off-the-shelf DEMs, Shuttle
Radar Topography Mission SRTM, derived from space-based Interferometric
Synthetic Aperture Radar, IFSAR and the USGS National Elevation Dataset, NED
compiled from a variety of sources. And the more accurate measured DEMs derived
from stereo imagery and IFSAR from aircraft. All are digital raster products. Among
the several factors that influence the accuracy of these or any DEM are the quality,
density and distribution of the source data. In practice these boil down to three
components. One is the post spacing or posting, also known as resolution. Another is
the vertical accuracy as it is inferred from the sources or as it is measured and finally
the horizontal accuracy of the DEM.
49


Posting is the ground distance between the elevations or z-values which are at
regularly spaced intervals on a grid. The post spacing is usually specified in units of
whole feet or meters. For example in the Shuttle Radar Topography Mission, SRTM,
finished data is available in 1 in the United States and 3 arc second postings
elsewhere. The 1 arc second data is also known as 30m data, and the 3 arc second is
known as 90m data.
3.3.2 Shuttle Radar Topography Mission, SRTM
Interferometric Synthetic Aperture Radar, IFSAR, aka InSAR, instruments operating
simultaneously in X and C band were flown for 11 days on the Space Shuttle in
February of 2000. The project produced the first near-global high resolution DEM
available to the public and the scientific community. In the summer of 2005 finished
DEMs became available. Finishing means the definition and flattening of water
bodies, delineation of coastlines, and removal of spikes and wells in the data along
with the filling of small voids. This data is now available online at Global Land
Cover Facility (GCLF) and United States Geological Survey (USGS) for most of
the world between 60 N Latitude and 56 S Latitude. In data poor regions of
the world SRTM is a boon, before SRTM the best available global elevation
dataset provided roughly one kilometer spatial resolution. The SRTM geographic
coordinate system is based on the WGS84 datum. The vertical datum for SRTM is the
50


currently available worldwide Earth Gravity Model 1996, EGM96. The fact that
SRTM covers 80% of the Earth is synoptic and available off-the-shelf are its main
strengths today.
However, there are drawbacks to the dataset. Finished SRTM of 1 arc
second, which is the 30m posting, covers the United States, its possessions
and territories but is unavailable elsewhere, the 3 arc second, which is the
90m posting, covers the rest of the world. There remain large no-data lacunae in
the SRTM data, particularly in areas of high-relief and deserts. The voids have been
filled in various ways, some are quite inaccurate
NASA has specified that SRTM has an absolute 90% (LE90) vertical accuracy of less
than or equal to 16 meters, and a relative 90% (LE90) vertical accuracy of less than
or equal to 10 meters. The relative accuracy describes the error in a local 200-km
scale (Yastikli et al. 2006). The absolute horizontal accuracy 90% (CE90) circular
error is specified at 20 m. The mission specified RMSE is 9.7m.
It is worthwhile to note here that linear error 90%, LE90, is used to quantify the
vertical error in the DEM versus the real world. At every point within the DEM in
question, there is a vertical error. LE90 is the error range which would include 90%
of the pixels within the DEM. Absolute LE90 is defined as the LE90 calculation for
51


the DEM with no corrections applied. As such the error includes the effects of
position and elevation inaccuracies. Relative LE90 is a measure of the error in the
surface shape of the DEM. Circular error 90%, CE90, is a measure of the combined
errors in latitude and longitude of the DEM. CE90 is a circular radius in meters,
which would include 90% of the positional errors of the DEM versus the real world.
This estimation was borne out by a comparison against GCPs conducted by N.
Yastikh, G. Ko^ak and G. Buyiiksalih in Istanbul, Turkey in 2006.
Accuracy of SRTM X and C-band DEMs were also checked against
the GCPs measured by GPS campaigns and RMSE-Z values were
found to be about 5.6 m and 9.6 m for X- and C-band respectively
under the estimation of 10 % mean slope value. (Rabus et al. 2003)
There have been several other studies that have also evaluated the vertical accuracy of
SRTM. For example, in January of 2006 Ashton Shortridge published his report,
Shuttle Radar Topography Mission elevation data error and its relationship to land
cover, in Cartography and Geographic Information Science. In his work Mr.
Shortridge compared SRTM in Michigan just northwest of the city of Detroit to an
Oakland County spot height dataset. He reported that the county dataset was held as
ground truth and, The County elevation data set was derived via a photogrammetric
survey on high-resolution aerial orthophotographs collected in April, 2000
(Shortridge 2006). He mentioned that The fidelity of the County data set to bare-
52


earth is quite high; accuracy is reported as within one international foot of actual
surface height. (Shortridge 2006). Concerning his results he wrote:
Mean error for the 7,044 points in the SRTM dataset was 3.2
meters, indicating that the point-based estimate is just over 3.2
meters too high on average. The error distribution was unimodal
and positively skewed, indicating a greater number of large
positive magnitude errors. Root mean squared error (RMSE) was
6.53 meters. This corresponds to a 90 percent absolute accuracy of
10.73 meters, about two-thirds as large as the maximum objective
in the SRTM specifications. (Shortridge 2006).
Elsewhere in his report Mr. Shortridge points out that SRTM errors are not random
especially as associated with forested areas, even in leaf off conditions, where it
significantly overestimates elevations. And as reported by Kon Joon Bhang and his
colleagues from Ohio State University in their report of SRTM data, Topographic
change on the surface is a factor of the errors of SRTM DEM high relief areas have
larger errors. (Bhang et al. 2005)
It is worth noting that ellipsoidal heights can be derived from SRTM data by proper
application of the EGM96 reference (Amitabh et al. 2006). A new EGM08 geoidal
model has recently become available. (Dr. Daniel Roman, Research Geodesist,
Geosciences Research Division, NGS, email message to author April 14, 2009). The
official Earth Gravitational Model EGM2008 has been publicly released by the
53


National Geospatial-Intelligence Agency (NGA) EGM Development Team.
(National Geospatial-Intelligence Agency)
3.3.3 National Elevation Dataset
The National Elevation Dataset, NED is the standard publicly available DEM
produced by the United States Geological Survey, USGS. Elevations are stored in
raster using geographic coordinates which more closely represent a bare-earth model
than does SRTM. NED has been derived from a variety of data sources of vintages
from 1925 to the present processed in 1 by 1 tiles. The source data for NED have a
range of elevation units, horizontal datums, and map projections. Older DEMs
produced by methods that are now obsolete have been filtered during the NED
assembly process to minimize artifacts that are found in data produced by these
methods. However, NED is likely to be less current than SRTM in many instances
and it certainly does not have global coverage. It covers the coterminous U.S.,
Alaska, Hawaii, Puerto Rico, and Virgin Islands.
The horizontal datum of NED is NAD83, except for Alaska, which is NAD27. The
vertical datum is the North American Vertical Datum 1988, NAVD88, except for
Alaska, which is the National Geodetic Vertical Datum 1929, NGVD29 as was most
54


of the original source data. Since it was referenced to NGVD 29 a vertical datum
conversion was required and this transformation has been accomplished by the
application of the National Geodetic Survey NGS tool VERTCON rather than the use
of any geoid model. The results are variable.
An estimation of ellipsoidal heights from NED based on the reference ellipsoid for
NAD83, Geodetic Reference System 1980, GRS80, is possible and could be
accomplished by application of a NGS geoid model such as GEOID03 or GEOID06.
GEOID06 currently only covers the state of Alaska. A new geoidal model will be
released by NGS shortly. (Conversation with Dr. Daniel Roman, Research
Geodesist, Geosciences Research Division, NGS, November 2006)
NED has an actual resolution of one arc-second for the United States, Hawaii, Puerto
Rico and the island territories with the exception of a resolution of two arc-seconds
for Alaska. However, the data product is available in 1 arc second-30m, 1/3 arc
second-10m, 1/9 arc second-3 m. The 1/3 arc second and the 1/9 arc second are not
available throughout the US. The 1/3 arc second product closely represents the
contour available on USGS quad sheets and edge matching issues between those
quads have largely been resolved. NED is updated bimonthly to incorporate the best
available DEM data and as more 1/3 arc second data covers the US it will become
55


seamless at that posting. Its horizontal accuracy is variable. Its vertical accuracy is
generally within the limits shown in Table 3.1.
3.3.4 NED-SRTM Comparison
Here is a comparison between two off-the-shelf DEMs.
Table 3.1 Comparison of NED and SRTM DEMs
NED (US) SRTM (US) SRTM (World)
Resolution 1 arcsec (30m) 1 arcsec (30m) 3 arcsec (90m)
Source Data Maps/aerial photos Radar (IFSAR) Radar(IFSAR)
Source Resolution 10m & 30m posting 30m posting 90m posting
Source Dates 1925 present Feb 2000 Feb 2000
Source Type Bare Earth First Return First Return
Horizontal Datum NAD83* WGS84 (orig) WGS84 (orig)
Vertical Datum NAVD88* EGM96 EGM96
Height Accuracy 7m to 15 m** 10m to 20m*** 10m to 20m***
* Except Alaska-NAD27 and NGVD29
**Depends on the original source DEM and if it was level 1, level 2, or 10m resolution.
***<=16 m absolute vertical height accuracy, <= 10 m relative vertical height accuracy and <=20 m
absolute horizontal circular accuracy. All accuracies are quoted at the 90% level, consistent with
National Map Accuracy Standards.
56


3.3.5 DEM Derived from Stereo Satellite Imagery
As reported in the May 2006 ASPRS journal PE&RS, Recent research has shown
that detailed digital elevation models (DEMS) can be generated with high-resolution
satellite stereo pairs. (Tao 2006).
Figure 3.8 SINGLE PASS STEREO COLLECTION
(Grodecki 2006)
57


Same pass satellite stereo collection is shown in Figure 3.8 which was adapted from
figure 9 in Dr. Jacek Grodeckis article Satellite Photogrammetry: Processing
Innovations Maximize High Resolution Imagery Options. To collect the first of the
two images that constitute the stereo pair the satellite yaws, rolls and pitches to the
target as required while pointing in a forward direction. A hundred or so seconds
later, after the first image is collected the satellite is maneuvered to again image the
same area, this time pointing in a backwards direction.
Information concerning this same pass approach to stereo pairs is pertinent to this
study in that it facilitates automated feature extraction because the scene content and
lighting conditions are virtually the same for the two images. Another one of the
benefits of stereo imagery is that it can be used as a source for DEM generation. The
stereo images can be block adjusted together in the ground station to improve relative
orientation and remove the y-parallax. There is some independent testing in support
of the resulting DEM. Describing their work in the Journal of Photogrammetry and
Remote Sensing September 2005 Rau and Chen reported, The experimental results
demonstrate that when IKONOS stereo images are utilized for DSM generation a
potential positioning accuracy less than two meters can be achieved. (Rau and Chen
2005). A DSM differs from a DEM in that it is a digital surface model including
trees, buildings, houses, etc. A DEM has these features removed.
58


3.3.6 Intermaps NEXTMap Product
NEXTMap is Intermap Technologies United States DEM project. They produce
DSMs and DEMs from Orthorectified Radar Imagery, ORI with a 1,25m pixel using
IFSAR.
The vertical datum is NAVD88 as defined by GEOID99. The geographic coordinates
are expressed in the horizontal datum NAD83 at an advertised RMSE of 2m. The
DEM resolution, that is the posting, is 5m and is available in an advertised vertical
RMSE of 0.70m and lm. (Intermap Technologies, NEXTMap)
3.3.7 DEM error Contribution to Horizontal Error
The relationship between the DEM error and the resulting horizontal error in the
orthorectified imagery can be expressed:
£^=tan6,(SD£)
In which:
e onho is the horizontal error in the orthorectified image
e dem is the vertical error in the digital elevation model, DEM
0 is the off-nadir angle
59


Table 3.2 Off-nadir angles and horizontal error
£ DEM Elevation 4(Q)* Off-Nadir 4 (0)** £ ortho Source as published
16m 45 45 16m SRTM (LE90 absolute)
16m 63.5 26.5 8m SRTM (LE90 absolute)
16m 76 14 4m SRTM (LE90 absolute)
10m 63.5 26.5 5m SRTM (LE90 relative) and NED
7m 16.6 13.4 1.66m NED
3m 61 29 1.66m IKONOS -GeoEye
lm 45 45 lm NEXTMap -InterMap (RMSE)
*The elevation angle is expressed from target to sensor otherwise the curvature of the
earth plays a role.
**The off-nadir expression of the angle is true only for aerial images due to low
flying height.
In the circumstance of the ADAS specification requiring 5m absolute horizontal
positional accuracy, it is desirable to use the most accurate DEM available, which
may be that built from stereo satellite capture or IFSAR from aerial capture, i.e.
60


NEXTMap. It is also desirable to specify the maximum allowable off-nadir angle
and/or minimum elevation angle value of the sensor during the data collection. If
possible it may be best to set the limits to achieve a resulting horizontal error in the
DEM that does not exceed ~1.66m 1/3rd of the 5m specification.
The best line of sight is down along the nadir of the sensor, which is the same as an
elevation angle of 90. At that angle the vertical inaccuracy in the DEM to horizontal
errors is zero. But as that elevation angle flattens, as it decreases, the error that the
DEMs inaccuracies contribute to the final ortho-photograph increases. In other
words, to maintain a particular horizontal accuracy standard both the accuracy of the
DEM and the size of this angle taken together must be considered. And the more
variation in the terrain the more distortion that variation contributes to the imagery, in
such circumstances the importance of the DEM increases.
The topographical variations in the surface of the earth and the tilt of
the satellite or aerial sensor affect the distance with which features
on the satellite or aerial image are displayed. The more
topographically diverse the landscape, the more distortion inherent in
the photograph. Image data acquired by airborne and satellite image
sensors are affected by systematic sensor and platform-induced
geometry errors, which introduce terrain distortions when the image
sensor is not pointing directly at the Nadir location of the sensor.
Terrain displacement can be hundreds of meters. For example, if the
IKONOS satellite sensor acquires image data over an area with a
kilometer of vertical relief, with the sensor having an elevation angle
of 60 (30 from Nadir), the image product will have nearly 600
61


meters of terrain displacement. Additional terrain displacement can
result from errors in setting the reference elevation. Low elevation
angles of images, imperfect terrain models, and variability of sensor
azimuth and elevation angles within an image limit accuracy
potential if image orthorectification is attempted. For this reason,
when new high resolution satellite image data is acquired over rough
terrain, high elevation angles of the sensor is required. (Satellite
Imaging Services)
Aerial imagery is more affected by this phenomenon than is satellite imagery.
Elevation angles on an aerial camera are a function of the distance from the principal
point (nadir). Unfortunately the elevation angles for particular camera models are
fixed. Aerial cameras have wide fields of view. They typically have a 6-inch focal
length lens with a 90 field of view. In that case many of the pixels have low
elevation angles which can exacerbate the contribution of the DEMs inaccuracies to
the bias in the imagery.
62


5m 5m
Right Image Left Image"
I Om Horizontal Ortho Error
Figure 3.9 OFF-NADIR ANGLE IN AERIAL 9X 9 FILM CELL
As illustrated in Figure 3.9 a standard 9-inch x 9-inch photo taken with a 6-inch focal
length camera implies a ~37 off-nadir angle at the edge of the photo. If a DEM has
an error of 7m the resulting horizontal error can be as large as 5m.
In aerial work the distance of a particular point of interest from the center, the
principal point, of an image affects the magnitude of horizontal error attributable to
the DEM. A point near the edge of a component, i.e. a 9x9 film cell, of an image
63


mosaic will have more of such error and will reflect the horizontal error of all the
features in that region of the photograph. A point near to the center of the cell will
have less horizontal error attributable to the DEM and will reflect the horizontal error
of all the features in that region of the photograph.
3.4 The Collection of Ground Control Points
As described in the previous section of this study, §3.1.3 Block Adjustment and
Orthorectification, orthorectification of high-resolution imagery into high accuracy
imagery requires Ground Control Points, GCPs. It is important to integrate the
collection of these points into street database vendors existing processes as
efficiently and effectively as possible. As mentioned in §2.1.2 Methodology, in this
study the GCPs used to register the aerial imagery will be harvested from data
captured via a mobile mapping vehicle (MMV). Both street database vendors
currently operate fleets of such vehicles.
3.4.1 Distribution of GCPs
In their Tampa Bay, Florida report in September 2003 entitled Geometric Modelling
and Processing of QuickBird Stereo Imagery, Mapping and GIS Xutong Niu et. al,
of Ohio State University tested various geometric modeling and photogrammetric
processing methods for mapping using stereo QuickBird images. Each of these
64


methods were supported by different configurations of the seven (7) photo-
identifiable GPS field collected GCPs illustrated in Figure 3.10.
Figure 3.10 GCP DISTRIBUTION
(Niu et al. 2003)
Across the Quickbird scene in Figure 3.10 which is ~16/2 km E-W and ~20km N-S,
the 7 points symbolized by green triangles were used as GCPs or CPs for different
configurations. The positional accuracy of these points was reported to have
accuracies of 0.06m horizontal and 0.09m vertical. Some of the conclusions of the
report are pertinent here. They wrote:
It is recommended that a scale and translation model or an affine
model with four to six well-distributed GCPs be used to achieve a
high level of accuracy. These methods seem to be most practical for
use in mapping applications. (Niu et al. 2003)
65


Concerning the scale and translation model test using four (4) GCPs distributed at the
comers of the scene they wrote:
Two GCPs are necessary for this model. In the object space, if these
two GCPs used are distributed in the cross-track direction, the
computed RMSEs of the ground points are relatively smaller than if
they are distributed in the along-track direction. This trend is
consistent with the results achieved using simulated IKONOS images
(Zhou and Li, 2000). The RMSEs calculated by using the method in
image space with two GCPs shows similar results associated with the
GCP distribution. In order to increase redundancy, more GCPs
should be used. With four evenly distributed GCPs, the result is
improved (less than 71 cm in the horizontal and 65 cm in the
vertical) in both object and image spaces. (Niu et al. 2003)
This conclusion is in concert with other studies and experience. Therefore, should
satellite imagery be available the GCPs to be used in the orthorectification of a scene
should include four (4) to six (6) well distributed GCPs, (i.e. at the comers of the
scene and across track). However, if the work relies on monoscopic aerial imagery,
as it does here, to achieve a consistently accurate solution in a monoscopic block it is
recommended to have a GCP in each one of the frames. That and the smaller
footprint of the aerial images require more GCPs when aerial photography is used.
Therefore, the efficiency of their collection via MMV will be beneficial if the method
can be shown to offer the appropriate accuracy.
66


3.4.2 Mobile Mapping Vehicles, MMV
Mobile mapping began in the same era as GPS. It resulted from the work of
researchers at the Center for Mapping at Ohio State University and the Department of
Geomatics Engineering, University of Calgary in the 1980s (Tao and Li, xii). The
first symposium on Mobile Mapping Systems, MMS, in May of 1995, was at Ohio
State University and the technology continues to develop rapidly today.
The major impetus behind mobile mapping has been from its beginnings the need for
highway information. Most mobile mapping vehicles have been driven along
highway and railway corridors and asset information recorded on video image
sequences with positioning and orientation provided by the integration of GPS and
Inertial Measurement Systems, IMUs.
3.4.2.1 GPS/INS
An IMU is a closed system of gyroscopes and accelerometers used to detect changes
in the roll, pitch and yaw of a moving vehicle. In other words, an IMU measures the
3-axis angular rate and the 3-axis linear acceleration of the vehicle. An Inertial
Navigation System, INS, contains an IMU. It includes a computer that can dead
67


reckon using the IMUs measurements. An INS computes the navigation solutions
and attitude references from the IMU data.
Inertial systems accumulate error, that is, the slight INS measurement errors that
occur as the MMV moves are additive. This accumulation of errors eventually
constitutes a drift of the measured position of the vehicle away from its actual
position. This is one reason an INS is usually part of a suite of sensors including
GPS. Since GPS determines position from external sources, satellites, it can help
correct the drift in the internal INS processes. (Titterton 2004, 410)
However, in practice the GPS signal from the satellites may be obstructed by bridges,
trees, tunnels and buildings in an urban area. During these outages the INS measures
the velocity and direction changes of the vehicle with accuracy. Therefore, the
GPS/INS combination is symbiotic. Although each technology can in principle
determine both position and orientation, they are usually integrated in such a way that
the GPS receiver is the main position sensor, while the IMU is the main orientation
sensor. (Tao and Li, 5). In other words, velocity and direction changes can still be
determined by the INS and fill in the gaps during GPS signal outages. In this way the
GPS measurements are used to index, update and correct the INS systems drift.
68


3.4.2.2 Relative Accuracy from MMV
In the realm of positioning the integration of an INS with GPS receivers is the
foundation of the sensor combination. Nevertheless, a typical MMV vehicle also
utilizes other positioning sensors as well such as a distance measuring instrument,
DMI. A DMI may be an odometer or a pick-up on the vehicles anti-lock brake
system, ABS, which is not only useful for operational reasons, i.e. keeping a fixed
distance between camera exposures, but also measuring the distance traveled.
The INS and DMI then can be thought of as being primarily concerned with track
measurements from point-to-point as the vehicle drives down the road. The
evaluation of such measurements would reveal the relative accuracy of the along-
track results and along with an ABS signal to determine wheel direction of travel on
the vehicle, the cross-track as well. Using these devices experiments indicate that the
relative accuracy delivered by an MMV may be within 0.50m within a range of 100m
from the vehicle ( K. Miksa, pers.comm.)
As defined by the American Society for Photogrammetry and Remote Sensing,
ASPRS, relative accuracy is a "measure that accounts for random errors in a data set.
Relative accuracy may also be referred to as point-to-point accuracy. The general
measure of relative accuracy is an evaluation of the random errors ... in determining
69


the positional orientation (e.g., distance, azimuth) of one point or feature with respect
to another. (Maune 2007,471). For example, a stake set to control a comer in
construction of a building must relate to the stakes at the other comers properly, but
their relationship to the center of mass of the Earth is immaterial. In other words,
their accuracy relative to one another is critical, but their absolute accuracy is of little
or no concern in such a context. The term relative accuracy is also used elsewhere
with different connotations, i.e. topology
3.4.2.3 Absolute Accuracy from MMV
On the other side of the configuration of sensors is the GPS component of the mobile
mapping vehicle. It can be thought of as being primarily concerned with
measurements in a wider context, specifically the NAVSTAR, the GPS Satellite,
constellation and by implication the earth-centered-earth-fixed datum on which the
system is based, currently, World Geodetic System of 1984 GPS Week 1150,
WGS84 (Gil50). The evaluation of such GPS derived positions then is in the realm
of absolute accuracy as it is generally understood in a mapping context.
As defined by ASPRS, absolute accuracy is, "a measure that accounts for all
systematic and random errors in a data set. Absolute accuracy is stated with respect to
a defined datum or reference system. (Maune 2007, 471). This definition is
70


corroborated by the Ordnance Survey, the national mapping agency of Great Britain.
A measure which indicates how closely the coordinates of a point in Ordnance
Survey map data agree with the true National Grid coordinates of the same point on
the ground. As the true position can never be known exactly, the statistic is quoted
relative to the best known position determined by precise survey methods.
(Ordnance Survey Glossary 2009)
However, as implied by these quotes, the meaning of absolute accuracy is the same in
both but the reference system involved is not. That will be different from country
to country. For example, the most current datums in the United States are National
Spatial Reference System (NSRS), including the North American Vertical Datum of
1988, NAVD 88 and the North American Datum of 1983,
'NAD83(C()RS96)(2003.00). The reference ellipsoid for these is Geodetic Reference
System of 1980, GRS 80. Clearly this differs, for example, from the reference
ellipsoid for the Ordnance Surveys National Grid which is the Airy Spheroid of
1830, OSGB36.
71


3.4.2.4 WGS84 (Gil50) ITRF00(2002.0)
Fortunately there is a solution. It happens that the datum for GPS is currently
WGS84 (G1150). This is the third and latest realization of the WGS 84 reference
frame. It was implemented on January 20, 2002 and redefined WGS84 with respect
to the International Terrestrial Reference Frame of 2000, ITRF00(2002.0). These two
reference systems agree with one another within +/- 0.02m in each component of x, y
and z in an Earth-Centered-Earth-Fixed, ECEF context (Malys et al. 1997). In other
words, it is possible to work in this system around the world.
GPS geographic coordinate positions will be expressed in ITRF00(2002.0)
coordinates in this study. These are virtually identical with WGS84 (G1150). There
are clear advantages to this approach to the wider application of the procedures
discussed here. In such cases it will not only obviate the need to transform work from
datum to datum, it will also provide worldwide datum consistency. On the other
hand WGS84 (G1150) and ITRF00(2002.0) differ from NAD83 (1986) up to ~1.5m
horizontally and ~1.0m vertically across North America (Craymer 2006). The
difference between the two datums is ~lm in x, ~2m, in y and ~0.50m in z in an
Earth-Centered-Earth-Fixed, ECEF context. (NIMA).
72


3.4.2.5 Alternative GPS Configurations
Since WGS84(G1150) is the datum on which the GPS Navigation Message is
currently based, if one is doing what is known as autonomous positioning with GPS
the results are in that system (Petrovskyy and Vyacheslav 2007). As a practical
matter autonomous positioning, also known as absolute or point positioning, is done
using a lone GPS receiver. However, such positioning without any sort of relative
correction to ameliorate the unavoidable GPS errors, i.e. ionospheric error, clock bias,
ephemeris error, etc. is inaccurate. Autonomous positioning has a positional accuracy
from ~ describes it this way:
The NAVSTAR GPS was originally conceived and designed to
provide point positioning and velocity of a user with a single usually
low-cost hand-held GPS receiver. This is termed absolute point
positioning, as distinguished from relative positioning when a
second receiver is employed. GPS absolute positioning is the most
widely used in military and commercial GPS positioning method for
real-time navigation and location. It is usually not sufficiently
accurate for precise surveying, mapping, or hydrographic positioning
uses horizontal accuracies are typically only in the 10 to 30m
range. (USACE)
73


if *
Figure 3.11 SINGLE RECEIVER -AUTONOMOUS GPS POSITIONING
Even though this sort of GPS positioning produces coordinates in an absolute context
and in the appropriate datum, WGS84 (G1150), unaided, it is not capable of
delivering the sort of accuracy required in the work that is the subject of this
dissertation. For that it is necessary to adopt a different scheme. The arrangement
that will provide the necessary absolute accuracy in the sense of the ASPRS
definition is what is known as relative GPS positioning. This requires a minimum of
74


two GPS receivers. The Federal Geodetic Control Commission described the
situation as follows:
There are two methods by which station positions can be derived;
point positioning and relative positioning. In the point positioning
method, data from a single station are processed to determine three-
dimensional coordinates (X, Y, and Z) referenced to WGS-84 earth-
centered reference frame (datum).
To perform geodetic surveys at the decimeter-level or better, one
must deploy relative positioning techniques. In relative positioning,
two or more GPS geodetic receivers receive signals simultaneously
from the same set of satellites. These observations are processed to
obtain the components of the base line vectors between observing
stations [station coordinate differences (dx, dy, dz)] (Hull 1989, 3-4)
Relative GPS, also known as Differential GPS, DGPS, depends on one of the two
GPS receivers involved being located at a known position. Please recall that the
ASPRS definition of absolute accuracy requires that known position be on a defined
datum or reference system. In the Test that is the subject of this dissertation that will
be WGS84 (G1150) ITRF00(2002.0). That known position could be a monumented
station in the nations government maintained network, i.e. the NSRS in the US.
Utilizing this sort of station necessitates its occupation with a GPS receiver, known as
a base station that simultaneously tracks the same constellation of GPS satellites that
are being tracked by the GPS receiver on the MMV. Under such an arrangement
there are three alternatives. One alternative is that post-mission the data from the
receiver on the known station, the base station, is combined with the data from the
75


MMV receiver, also known as the rover, and they are post-processed together to
calculate the positions collected by the van, please see Figure 3.15. In the second
alternative the GPS receiver on the base station is provided with an RF transmitter
through which it beams real-time corrections to the rover on the MMV, please see
Figure 3.12. The third alternative is similar to the other two, but with an important
difference. The base stations and their data are managed by a commercial concern or
a governmental agency. This organization then provides either the real-time RF
correction signal or the recorded data that was archived during the mission, or both,
please see Figure 3.13. These services are available through various means,
sometimes free on the internet and sometimes through a subscription. The base
station data from these organizations is available in submeter, code phase, and
centimeter carrier phase, level services.
76


DGPS
Characteristic DGPS
Satellites required for initializing 3 minimum: 4 required for sub-meter accuracy
Time required for initializing Instantly
On-the-fly initialization (obtaining centimeter accuracy while moving} No
Receiver Single frequency sufficient
Accuracy Sub-meter (Horizontal Axis only)
Base Station requirement Operator-owned, fee based correction service provider, or free radio beacon broadcasts (e g.. Coast Guard or WAAS)
Figure 3.12 2nd ALTERNATIVE RELATIVE GPS POSITIONING
77