Freeway incident detection models using automatic vehicle location data from buses

Material Information

Freeway incident detection models using automatic vehicle location data from buses
Hoeschen, Brian
Publication Date:
Physical Description:
vii, 70 leaves : illustrations ; 28 cm


Subjects / Keywords:
Traffic engineering -- Colorado -- Denver ( lcsh )
Motor vehicles -- Automatic location systems ( lcsh )
Traffic flow -- Measurement -- Colorado -- Denver ( lcsh )
Disabled vehicles on express highways -- Data processing ( lcsh )
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )


Includes bibliographical references (leaves 69-70).
General Note:
Department of Civil Engineering
Statement of Responsibility:
by Brian Hoeschen.

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:
43917582 ( OCLC )
LD1190.E53 1999m .H64 ( lcc )

Full Text
Brian Hoeschen
B.S., Colorado School of Mines, 1997
A thesis submitted to the
Ur versify of Colorado at Denver
in partial fulfillment
of the requirements for the degree of
Master of Science
Civil Engineering

This thesis for the Master of Science
degree by
Brian Hoeschen
has been approved
Bfuce Janson

Hoeschen, Brian (M.S., Civil Engineering)
Freeway Incident Detection Models Using Automatic Vehicle Location Data
Thesis directed by Assistant Professor Sarosh I. Khan
In recent years, transit agencies have installed automatic vehicle location systems (AVL) in
buses for fleet management. However, the data collected by an AVL system is not used for
traffic management. This s tudy investigates the feasibility of using data from a bus AVL
system for freeway incidei t detection. A Mahalanobis-distance based classifier, a
generalized linear classifier, and a neural network based classifier was applied to detect
freeway incidents. The performance of the classifiers was evaluated based on detector data
and bus AVL reports. The total number of AVL reporting per interval may vary based on
average bus headway and reporting interval. The impact of the number of reportings per
time interval on model performance was also investigated. This study suggests that AVL
data may be used by a traffic management center to increase the overall performance of an
incident detection model by mainly lowering the false alarm rate. It was also found that
decreasing the AVL reporting interval improves the performance of an incident detection
model only for low average bus headway. The performance of the generalized linear and
neural network classifiers was comparable. However, the neural network model outperforms
the generalized linear model with a one-interval persistence test.
This abstract accurately represents the content of the candidates thesis. I recommend its
Sarosh I. Khan

My thanks to the National Presto Foundation for the scholarship that allowed me to obtain
this degree.

1. Introduction ...................................................................1
1.1 Incident Detection..............................................................1
1.2 Lite ature Review...............................................................2
1.3 Obj-dives.......................................................................3
2. Background .....................................................................4
2.1 Der vers Bus AVL System........................................................4
2.2 CDOT Detector Data..............................................................5
2.3 Mic osimulation: CORSIM.........................................................5
2.3.1 S'nulating Buses................................................................6
2.3.2 CDRSIM Output..................................................................7
3. Field and Simulated Data........................................................9
3.1 Fielt Data......................................................................9
3.1.1 D tector Data..................................................................10
3.1.2 Vdeo Data......................................................................10
3.1.3 P obe Data....................................................................13
3.2 Simulation Data................................................................19
3.2.1 Simulation Input Data..........................................................19
3.2.2 Simulation Calibration........................................................21
3.2.3 Incident Detection Data.......................................................25
4. Methodology....................................................................32
4.1 Models Proposed................................................................32
4.2 Discriminant Analysis..........................................................32
4.3 Generalized Linear Model.......................................................34
4.4 Neural Network Model...........................................................35
5. Results .......................................................................37
5.1 Performance Measures...........................................................37
5.2 Variable Selection.............................................................38
5.3 Detector Model (PDet)..........................................................40
5.4 Bus Model (Pbus)...............................................................40
5.5 Detector and Bus Probe Model (Pdet&bus)........................................43
5.5.1 Combined Model.................................................................44
5.5.2 Comparing DA, GLZ, and NN Classifiers for the Combined Model..................45
6. Conclusions and Recommendations................................................49
A. Detector Reader................................................................51
B. TSD Reader.....................................................................54
C. Bus VPS Data Extraction........................................................57
D. CORSIM Input File ExcelA/BA Program............................................59
E. Access VBA Program and SQL Queries.............................................60
F. Classifier Parameters..........................................................66
References .........................................................................69

1.1: Lane Blocking Freeway Incident.................................................1
2.1: Denvers AVL System............................................................5
2.2: Typical Metered Freeway Ramp...................................................6
2.3: TSD Converter User Interface...................................................8
3.1: Picture of I-25 Network........................................................10
3.2: Five-min volume based on CDOT detectors and the video detection system (I-25 at
3.3: Bus Travel Time Extraction Using Video.........................................12
3.4: Autoscope Volume And Speed Detector Layout.....................................13
3.5: Sample VPS Data File...........................................................14
3.6: Sample Bus Assignment File.....................................................14
3.7: Steps in VPS and Bus Assignment Data Processing................................15
3.8: Schedule Adherence Data Processing.............................................16
3.9: All Buses Extracted From the VPS Data Based On Bus IDs in Test Network........17
3.10: GIS Buffer to Extract Bus-AVL Data by Location................................17
3.11: Replay of Buses on I-25 Between Evans and Logan at 8:06 AM, 12/9/98...........18
3.12: GIS Processing of the Bus-AVL Location Data....................................18
3.13: Link-Node Diagram for I-25 Network in CORSIM..................................20
3.14: Frequency of Percent Average Headway..........................................21
3.15: CORSIM Volume And Field Volume Comparison at Downing..........................23
3.16: CORSIM Volume vs. Field Volume................................................24
3.17: Comparing Simulation and Field Bus Space Mean Speec for Incident Case.........25
3.18: Section Diagram For I-25 Used for Incident Detection..........................27
3.19: Link and Network Distance for Buses...........................................29
3.20: Detector Variables.............................................................31
4.1: Euclidean Distance and Statistical Distance.....................................33
4.2: Neural Network Structure for the Poets.bus Model...............................36
5.1: Incident State Reported By Two Models..........................................38
5.2: DR vs. FAR for the Detector Models.............................................40
5.3: Performance of Bus Model for Different Average Bus Headway and Report Frequency
(H2= 2min headway, RI30=30-sec reporting interval).............................42
5.4: Incident Cases Only on a Squared Mahalanobis-Distance Plot for 2-min Average Bus
Headway, and 10-Second Reporting Interval (nBus= number of buses)..............43
5.5: Relative Performance of the PDet, Pbus and Pdetsbus DA Classifier for Intervals Bus
Data Available..................................................................44
5.6: DA, GLZ, & NN: Combined Model (Pcombined) Models for 10 Second Bus Report Interval
and Corresponding Detector Model (PDet).........................................45
5.7: Squared Mahalanobis distance of MDET and MCombined-2 Model for incident and non-
incident state vectors (3 bus reportings).......................................46
5.8: DA, GLZ, & NN: Combined Model (Pcombined) Models for 30 Second Bus Report Interval
and Corresponding Detector Model (Poet).........................................47
5.9: Time To Detect vs. Persistence Test for (NN=neural network, DA=discriminant analysis,
GLZ=generalized linear model)...................................................48

2.1: RTD Fleet Composition.............................................................4
2.2: Default Dwell Time Distributions for CORSIM Bus Stations..........................7
3.1: Data Collection Summary............................................................9
3.2: Custom Dwell Time Distribution Used to Match Bus Headways.........................21
3.3: Regression Statistics for CORSIM Calibration......................................23
3.4: Variance of Bus Travel Speed......................................................25
3.5: File Naming Convention Based on Variables.........................................26
3.6: Total Simulation Cases...........................................................26
5.1: Variables Selected for Each Model and Model Performance...........................39
5.2: Performance of the Bus Models.....................................................41

1. Introduction
From the early 1970's, research has been conducted to develop incident detection algorithms
for freeways to aid traffic engineers. Algorithms have been developed applying pattern
recognition techniques, pattern matching techniques or comparative analysis, statistically
based algorithms, and traffic flow modeling. These models typically use data collected from a
fixed location based sensor.
In recent years, transit agencies have installed automatic vehicle location systems (AVL) in
buses for fleet management. However, this data currently is not used for traffic management.
This study investigates the feasibility of using data from a bus AVL system for incident
detection. This study also presents the performance of a Mahalanobis-distance based
classifier, a generalized linear model based classifier, and a neural network based classifier
to detect freeway incidents using data from fixed and mobile sensors.
1.1 Incident Detection
An incident is any event that causes a reduction in the capacity of a roadway. A simple
freeway incident is shown in Figure 1.1. This inciden' occurs
Downstream Station ------- Upstream Sti tion
Lane 1 - Vdli d1i Sdi - V ii> Oui, Sui
Lam 2 Vd2i VU2> Ou2, su2
Lane 3 Vd3> d3f Sd3 VU3! u3> u3
Figure 1.1: Lane Blocking Freeway Incident
between upstream and downstream detector stations. Each detector reports three basic
traffic measures used to describe traffic flow:
Volume (V): number of vehicles crossing a section of roadway over a given time, usually
expressed as equivalent hourly flow rate, vehicles per hour
Occupancy (0): ratio of the time that vehicles are present at a detector station in a traffic
lane and the time of sampling.
Speed (S): Arithmetic mean of individual vehicle speed

When an incident occurs, typically the decn ase in volume at the upstream station and
increase in volume of a downstream station is accompanied by an increase in the upstream
occupancy and a decrease in the downstre m occupancy. Occupancy and volume difference
by lane may also be observed. The magniti de of the difference depends on the severity (e.g.
number of lanes blocked, duration) and the location of the incident and the traffic volume.
For any section of freeway, traffic measure: from an upstream and downstream station
provides the data to automate the process i.-f detecting incident and non-incident conditions
of freeway sections and therefore, pattern-r icognition techniques have been applied to
incident detection.
1.2 Literature Review
Payne and Tignor [7] developed pattern-matchinc algorithms, based on decision trees, for
freeway incident detection. These threshold-basi d algorithms, better known as the California
Algorithms, are based on the traffic pattern that o ;curs during an incident. Time-series based
algorithms were also developed to detect inciden s [1], They rely on forecasting short-term
traffic parameters. Significant deviations of estimated values from observed traffic
parameters lead to an incident alarm. Levin and classify incident and non-incident da'a. This algc ithm uses the ratio of the difference
between upstream and downstream one-minute ccupancies and historical incident data.
Based on the incident and incident-f ee condition A distributions of incidents and traffic
measures, the probability of the occi rrence of an incident at a pai icular location is
developed. This algorithm was shown to perforrr better than the C alifornia Algorithm, but
had a higher mean time to detect. The McMaste Algorithm [2] ap >lies catastrophe theory to
the two dimensional analysis of traffc flow and or cupancy data, by separating the areas
corresponding to different states of i affic conditic ns. When specific changes of traffic states
are observed over a period of time,; n incident £ arm is given. Neural network approaches
were also applied to freeway incidt nt detection [C; 4; 5], The multi-layered feed-forward
neural network and the probabilistic leural netwi rk architecture w< re applied.
More recently [6], data from two contiguous dete ;tors were utilized in an algorithm that tracks
changes in the difference between the current ci mulative sum of the log-likelihood ratio of
the probability density function of normal and ch; nged conditions (CUSUM) to the minimum
value up to the current time period. The algorith Vs performance was compared to the
California Algorithm [7] and a low-pass filtering £ gorithm [8] using data from the California
PATH database. Recently, Peta and Das [1998] modified the California Algorithm and the
McMaster Algorithm by adjusting the thresholds 'or the algorithms using the least square and
error back propagation (EBP). Their work show; that the neural net learning algorithm, EBP,
exhibited better learning capabilities for both algorithms.
Discriminant analysis (DA) has been applied to detecting incidents on urban arterials, in the
ADVANCE project [9; 10]. Fisher linear discriminant functions were developed based on
volume and occupancy deviation, occupancy ratio and volume to occupancy deviation for
upstream and downstream detectors; and link travel time ratio using probe data. Their work
included an evaluation of a fixed detector based algorithm, a probe vehicle based algorithm
and an integrated algorithm. However, 25% of the vehicles in the traffic stream were
considered probes and an incident detection interval of 7 minutes was used. One study

investigated a neural network based approach to combine the detector and probe vehicle
discriminant scores. The integrated neural network model performed the best.
1.3 Objectives
This study investigates the feasibility of using data from a bus AVL system for freeway
incident detection. This study presents the performance of three classifiers: Mahalanobis-
distance based classifier, generalized linear classifier, and a neural network based classifier.
These classifiers are used to detect freeway incidents based on detector data and data from
buses acting as probes, reporting location at pre-specified intervals. Sepa.ate models were
developed based on detector data only, bus data only and both detector at d bus data. The
performance of all three classifiers was evaluated and compared. Report in'ervals of 10, 20,
and 30 seconds were used for the bus probes. Depending on average bui headway and
reporting interval, the total number of probe reportings per time interval vai es. The effect of
probe reportings per time period on model performance was also investigated.
Highlights of this investigation include:
Pro )e data applied to incident detection for freeways
Ust of space mean speed of probe vehicle estimated based on reporting intervals of
10, 20, or 30 seconds
Shorter sampling interval or incident detection intervals
Effect of number of probe vehicles (average headway) and probe reporting frequency
on the performance of incident detection models
A comparison of the performance of three classifiers for freeway incident detection
Simulation of a bus location system in a traffic microsimulation

2. Background
This investigation was based on data collected from a local transit agency's AVL system and
the state transportation agency's freeway surveillance system. This chapter describes the
AVL system, surveillance system and a micro-simulation used for this study.
2.1 Denvers Bus AVL System
The Regional Transportation District (RTD) in Denver installed an automatic vehicle location
system in 1993. The system consists of 1,335 vehicles in its fleet. RTD installed this system
to: (1) develop more efficient schedules (2) improve ability of dispatchers to adjust on-street
operations (3) increase safety through better management [11], Their fleet composition is
provided in Table 2.1 below:
Vehicle Type Quantity
Fixed route service buses 936
16th Street Mall buses 27
Light Rail 17
ADA, Access-a-Ride vehicles 175
Maintenance and supervisor 180
Total 1,335
Table 2.1: RTD Fleet Composition
Denver's AVL system components are shown in Figure 2.1. Each vehicle in the RTD fleet is
equipped with an Intelligent Vehicle Login Unit (IVLU) and an antenna mounted on it to
receive GPS signals and the differential correction signals in real-time. The real-time
differentially corrected location is transmitted to the dispatch center. The communications of
the AVL system consists of nine microwave channels in the 450 MHz range. Seven channels
are used for fixed-end voice communications and the two remaining are used for data
In urban environments, because of signal degradation and obstruction, the Denver AVL
system integrates the GPS with inertial sensors or dead-reckoning (DR) sensors. The
location accuracy of the GPS is 1-2m. However, no specific data was available on the
accuracy of Denvers integrated GPS-DR system. Bus location data is available every two
minutes through this AVL system and was used for this study.

Differentially corrected location, vehicle, route and
message data
Voice Communications
Voice Communications
Voice Communications
Differentially corrected
location, vehicle, route and
message data
Field Supe visor
and Maintenance
Figure 2.1: Denvers AVL System
2.2 CDOT Detector Data
In 1981, the Colorado Department of Transportation (CDOT) implemented ramp metering on
five entrance ramps to northbound I-25. Four of the entrance ramps for this project, Evans,
Colorado NE, Colorado NW, and SB University, are equipped with ramp metering. The
University location is not operational, however detector data is still collected. A typical
metered freeway ramp is shown in Figure 2.2. A central computer controls the metering rates
based on detector data. The detector data includes volume, speed, and occupancy for each
lane for the mainline detectors and volume and occupancy for the ramp detectors. The
mainline detector data was used for this study.
2.3 Microsimulation: CORSIM
The study network was represented in a microscopic traffic simulation model CORSIM [12].
This simulation model has been extensively used and calibrated for US conditions. CORSIM
is a microscopic simulation model developed initially in the 1970s [13] and has been
enhanced subsequently [14]. It is an interval-based, stochastic traffic simulation model that
represents traffic flow and operations for both surface street and freeway networks. Vehicles
are moved through the network based on a vehicle-following model, in response to traffic
control and demands. Interactions between vehicles and interruptions are explicitly modeled.

Incidents or lane blockac es can be simulated, and vehicles respond by switching lanes where
sufficient gaps exist. It also simulates bus routes on a surface street network or an integrated
freeway and surface street network, but not on a freeway network only. Version 4.3 of
CORSIM was used for this research.
2.3.1 Simulating Buses
Bus operation in a CORSIM network is described in terms of bus routes, stations, and flow
rates. Each route is assigned a route number, which is then used to specify the flow rates,
release offsets, and stations that buses may stop at along the route. The bus route is defined
by specifying the path of nodes and stations that the bus traverses as it travels through the
network. The bus flow rate is the mean headway for buses that service a particular route.
This entry may be input for each time period to reflect changes in mean headway over time.
Also, each bus route may be given a start time offset so that a bus for each route will not
enter the network at the same time. (It is important to note that this offset begins at the start
of the initialization period and not at the start of the simulation.) The routes can traverse the
interface nodes between sub-networks so that a route may cover NETSIM and FRESIM

streets and highways. A route must have at least one station stop. However, stations cannot
be specified on a FRESIM link.
The bus stations are defined according to the link they are on, the distance from the
downstream node, and their capacity (in numbers of buses). The average dwell time for
buses at stations and the percentage of buses that do not stop can also be specified. The
dwell time distribution can be specified as well. A random number, k, between 1 and 10, is
used to enter a distribution of dwell time factors (in percent). There are six such distributions,
one for each of six possible bus station types. The type of station is just the statistical
distribution of dwell times applicable to the station. The default values are shown in (Table
Type/k 1 2 3 4 5 6 7 8 9 10
1 40 60 70 80 90 100 120 130 140 170
2 24 48 59 75 85 94 111 126 155 223
3 30 47 65 77 90 103 116 137 157 178
4 0 29 59 75 92 108 125 148 170 194
5 0 18 3 5 70 104 125 144 156 167 180
_S 0 0 5 48 96 120 144 171 198 223
Table 2.2: Default Dwell Tim Distributions for CORSIM Bus Stations
These stations may be either attached tc a bus route; buses on this route stop at the station,
or unattached, buses on this route do nc t stop at the station. While servicing passengers at
a station, if a bus blocks the rightmost la le of moving traffic, this lane is coded as
unprotected. If, on the other hand, a bus in dwell does not block a lane, the station is
"protected". An unprotected station must be in the rightmost lane [15].
2.3.2 CORSIM Output
CORSIM provides a text based output file that includes average measures of effectiveness
(MOE) for predefined simulation intervals, for the entire network, sub network, and links. This
output files also contains the detector data. In addition to this standard output, one-second
bus data was extracted. Detector Output
The detector data in CORSIM includes volume, mean speed, mean headway, and mean
occupancy rate for each detector over specified report intervals. A program was written to
extract the detector data from the text output files [Appendix A]. Bus Output
CORSIM reports two types of bus MOE. One type is based on route-specific, network wide
data. This output includes; bus trips, total travel time, mean travel time, person trips, and

person trc vel time. The other MOE group is based on data from each link and is available
only for N iTSIM links and not FRESIM links. This output includes; bus trips, person trips,
travel tim<, moving time, delay time, M/T, speed, and stops [12].
As part of this study, average bus travel speed for specific links in FRESIM was needed.
This is no available as standard CORSIM output. A program was developed to extract
second b; second bus location data from CORSIMs animation files. This program uses an
ActiveX o ntroller [16] to extract information from the CORSIM binary output files. A sample
VBA prog am is shown in Appendix B. Another similar program was written for this project
that extracts bus data from the animation files. The user interface for this program is shown
in Figure 2.3. When run in batch mode, this program allowed the extraction of the bus data
from an unlimited number of animation files at once. This program could also be used for
extracting any specific CORSIM vehicle type or all vehicle types at once. The program also
allows the extraction time period to be changed.
& TSD Converts* 1.
Paramdef$ Single C Batch
TSD File: [FAT hesis\13915232. tscj u
Convert From: 1 ^ te| CO o o m
Fleet Type: |3;Bus 3
Convert 1 Exit 1
Figure 2.3: TSD Converter User Interface

3. Field and Simulated Data
This chapter describes the data collected in the field and from a microscopic traffic simulation
model, CORSIM. The field data section describes 'he detector data, video data, and bus
probe data. The simulation section describes the c ilibration of the CORSIM network, the
simulation data extracted, and the database developed for incident and non-incident data.
3.1 Field Data
Field data collection included detector, video, vehicl probe, and bus probe data. Table 3.1
shows the field data collected for two days.
NB I-; 5: Evans to Logan Ir cident Non-Incident
f. 2/9/98 3/24/99
Video @ Logan 6: >0 9:00
Video @ Evans 6: -0 9:00
Video @ Downing 6; '0-9:00 6:00 9:00
CDOI Detector @ University 6:00-9:00
CDO' Detector @ Colorado NW 6:00 9:00
CDOI Detector @ Colorado NE 6:00 9:00 6:00-9:00
CDO' Detector @ Evans 6:00-9:00 6:00-9:00
RTD/.VL 6:00-9:00 6:00-9:00
Placer GPS 7 runs
ProXR GPS 6:00-9:00
Table 3.1: Data Collection Summary
The test network is a northbound section of Interstate Highway 25 in Denver, Colorado
between Evans Ave. and Logan St (Figure 3.1). It consists of the following entrance and exit
Entrance from Evans Ave.
Exit to Colorado Blvd.
Entrance from NB Colorado Blvd.
Entrance from SB Colorado Blvd.
Exit to NB University Blvd.
Entrance from NB University Blvd.
Exit to SB University Blvd.
Entrance from SB University Blvd.
Exit to Downing St.
Exit to Buchtel Blvd.
Entrance from Buchtel Blvd.

The posted speed limit is 55 mph. It has three through lanes of traffic with one auxiliary lane
between the entrance from NB University and the Buchtel exit. This section of 1-25 has
approximately one percent heavy vehicles. Data was also collected for this location when an
incident or lane blockage occurred.
3.1.1 Detector Data
Mainline-paired loop detectors are located at Evans, and Colorado; and ramp detectors are
located at several on-ramp locations: Evans, Colorado, and University. 5-min volume and
speed data was collected for the mainline and ramp detectors. This summary data was
obtained from CDOT for the days specified in Table 3.1. The detector data includes volume,
speed, and occupancy for each lane for the mainline detectors and volume and occupancy
for the ramp detectors. The average speed and the total volume across all lanes were also
used for this project.
3.1.2 Video Data
In addition to the detector data, the freeway was video taped from several overpass locations.
This video data was used for two purposes; to estimate bus travel time and for vehicle
Travel Time
The video recordings for the Evans and Logan locations were used to match buses and
vehicles to estimate travel time. Each video camera was synchronized with GPS time and
each videotape was time stamped. By comparing the upstream and downstream video
recording, bus and auto travel times was determined (Figure 3.3). The travel time data was
collected for all buses traveling through the network. Video Detection System
The video data was also used with a video detection system to extract traffic data. The
detection system uses machine-vision technology to measure and calculate:
Vehicle presence
Time and space occupancy
Vehicle length
For this study, volume and speed data was collected from locations where CDOT detectors
were not available: Downing and Logan. There were overpasses at these locations that
provided the best field of view [17]. Volume data was also extracted at Evans to compare
against the CDOT detector data, which was also available at this location. Five-minute
volume collected using the video detection system was compared to the detector data for one
lane of northbound I-25 at Evans (shown in Figure 3.2). A speed and count detector layout
for I-25 at the Logan overpass is shown in Figure 3.4.
NB I-26 Q Evans Lane 1
Figure 3.2: Five-min volume based on CDOT detectors and the video detection system
(i-25 at Evans)

Video @ Evans Video @ Logan
Figure 3.3: Bus Travel Time Extraction Using Video

Figure 3.4: Autoscope Volume And Speed Detector Layout
3.1.3 Probe Data
In addition to the detector and video data, vehicle location data or probe data was also
collected. The vehicle probe data was collected using test vehicles equipped with GPS
receivers and the bus probe data was extracted from RTDs AVL system. Bus Probe Data
Bus data was collected from the AVL system, for the test network, during a recurring
congestion period of 6:00 am to 9:00 am and for one incident condition or non-recurring
congestion. There are normally 12 northbound bus routes during the morning peak period for
the test network, with route headway varying from 6 to 42 minutes. The average observed
bus flow rate was 27 buses per hour. Denvers Bus AVL system does not report speed,
therefore the location data was processed in a geographic information system (GIS) to
calculate bus speed. Two different methods were used to extract the bus location data,
depending upon the type of data available.
13 VPS Data Extraction
Bus AVL data collected from RTD provided bus location every two minutes for all buses in
Denver. A bus identification is assigned to buses and is used in their Vehicle Positioning
System (VPS) database. As shown in Figure 3.5, the VPS file provides time, vehicle ID, and
bus location in NAD 27 state plane coordinates.
A Bus Assignment data file provides cross-reference on bus routes to Bus ID or Vehicle
Location Unit ID (VLUID). A few lines of the bus assignment file are also shown in Figure
<06:30:09>VPS msg: Location Report getting VLUID 8032: (errno num = 4)
<06:30:09>VPS msg: Pack location report x = 2159704 y = 715336:
(errno num = 4)
<06:30:09>VPS msg: Location Report getting VLUID 5011: (errno num = 4)
<06:30:09>VPS msg: Pack location report x = 2101992 y = 681555:
(errno num = 4)
<06:30:09>VPS msg: Logon Notice Handler for vehicle 1513
<06:30:09>VPS msg: logon: vluid = 1513 bid_signout_avail = 1
allow_mismatch_logon = 1
<06:30:09>IVLU 5025 raw status 3 raw dev 15 raw task seq 1 raw trip
seq 12: (errno num = 4)
<06:30:09>RSA Status Handler getting VLUID 5025: (errno num = 4)
<06:30:09>5025 0+1 tps 12
<06:30:09>P 5025 st 3 dev 15 task 1 trip 12 ha 1 g 2: (errno num = 4)
<06:30:09>VPS msg: Pack location report x = 2143747 y = 627856:
(errno num = 4)
<06:30:09>VPS msg: Location Report getting VLUID 8339: (errno num = 4)
Figure 3.5: Sample VPS Data File
Bus ID=VLUID Rniit*
\ /
1 2 363 1506 B 54 1132 992 140
1 2 364 1948 Z 74 1145 993 152
1 2 365 8081 7 6X 70 1070 993 77
1 2 366 7021 82X 70 1166 993 173
1 2 367 9034 100X 72 1094 993 101
1 2 368 8042 0L 76 1148 994 154
1 2 369 5068 B 46 1121 994 127
1 2 370 7019 8X 71 1118 995 123
1 2 371 5065 7 8X 75 1160 995 165
1 2 372 9042 120X 77 1108 995 113
1 2 373 9051 34 72 1122 997 125
1 2 374 8055 10 23 1106 998 108
Figure 3.6: Sample Bus Assignment File

From the VPS data for all buses, data for the test network and bus routes between 6:00 am
to 9:00 am were extracted using a Fortran90 routine [Appendix C]. The following flowchart
Figure 3.7: Steps in VPS and Bus Assignment Data Processing Schedule Adherence Database
RTD has also developed a real time database of the VPS data for run time analysis and to
perform statistical analysis of schedule adherence. This database along with route and
interline route information can be used to extract the bus position data.
As shown in Figure 3.8, the route table contains the route number, name, and a unique route
ID. The interline table contains the route number and beginning time for a sequence of
routes. To extract the bus position information, the first step was to extract the route ID for
each route of interest (Figure 3.8). This route could be part of an interline route, so the
interline route ID was also extracted. These route IDs were then used to extract the bus ID,
time, and location data from the bus table.
15 GIS Processing
For the test network, the Bus-AVL data was filtered by location, direction of travel and
number of reports. For every bus ID extracted from the VPS data, there is an associated
time and location every two minutes for the data collectic n period of 6:00-9:00 AM (Figure
3.9). From this data, only data for northbound buses alo ig the test network were retrieved.
The first step in this process was to select the reports based on location. A geographic
information system (GIS) software was used to display tl e buses and select those within a
buffer around the network (Figure 3.10) [18]. The next s ep was to extract the buses based
on direction by comparing consecutive reports of locatioi in terms of state plane coordinates
(NAD 27) x and y for

Figure 3.9: All Buses Extracted From the VPS D< a Based On Bus IDs in Test Network
each bus. Subsequently, only individual buses with 'wo or more r< portings were kept in the
data. GIS linear referencing tools were used to estir ate distance raveled along the roadway
between consecutive reports and a bus speed was < alculated for t ach pair of reportings.
ArcView was also used to replay the buses in order o check the ir tegrity of the data (Figure
3.11). Figure 3.12 highlights this process. Five-min ite average b s speeds were then
estimated based on the individual bus speeds.
Figure 3.10: GIS Buffer to Extract Bus-AVL Data by Location

Figure 3.11: Replay of Buses on I-25 Between Evans and Logan at 8:06 AM, 12/9/98
Figure 3.12: GIS Processing of the Bus-AVL Location Data

3.2 Simulation Data
The 1-25 network was simulated in CORSIM from 6:00-9:00 AM. The data collected on
March 24, 1999, was used as input data for the non-incident case and the data collected on
December 9, 1998, was used for the incident case.
3.2.1 Simulation Input Data
A network is represented in CORSIM as a series of links and nodes. A network is divided
into links to include entrance or exit ramps and represent changes in geometry.
Figure 3.13 shows the link-node diagram for the I-25 test network between Evans and Logan.
Dual loop detectors were placed in the CORSIM network at the same locations as CDOT and
video detectors for calibration purposes.
Time-varying entry volumes and ramp volumes were specified for different time periods in
CORSIM. CORSIM allows up to 19 time periods to be specified for an entire simulation run.
The three-hour tim* period used for this study was divided into ten-minute intervals or 18
CORSIM time peric ds.
The mainline and r; mp detector data were used to specify the flow for the 18 time periods.
The entry volume specified in CORSIM included one-percent heavy vehicles (determined
based on the video recording).
The bus frequencies were represented in CORSIM based on the bus headway data
collected. This was accomplished using two different methods. The first method was to
match bus for bus using the AVL data. Every bus that had a location reporting for a specific
day was coded as a separate route in CORSIM using an initial offset to the match the actual
appearance time and a large headway (greater than the simulation time) so that only one bus
appeared for each route. For this method, the mandatory bus station was placed at the
beginning of the network with a dwell time of zero.
The second method involved coding only one route with the same average headway and
using the bus station average dwell time and dwell time distribution to vary to bus headways.
The video data provided the time at which each bus entered the network. Using this data, an
average headway was calculated from all buses as well as headway and percent average
hear way for each bus. Figure 3.14 shows the distribution of percent average headway with
ten categories.

Figure 3.14: Frequency of Percent Average Headway
The relative frequencies from this histogram were then used to develop a custom dwell time
distribution in CORSIM, shown in Table 3.2. In order to match the frequencies shown in the
graph, fifty percent, or five, of the k-values are between 0 and 65, thirty percent are between
65 and 130 and so on. The only requirement in CORSIM for the dwell time distribution card
is that the numbers sum to 1000.
One advantage of this method is that the number of buses and the t' Tie at which buses
appear in the network changes when running multiple simulation rur s, unlike the first method
in which the buses are coded to appear at the same time. Another < isadvantage of the first
method is that the amount of coding required is proportional to the n imber of buses, which
for this project was more than 80.
Type / k 1 *) t 3 4 5 6 7 8 9 10
Custom 40 0 55 60 65 86 106 120 170 248
Table 3.2: Custom Dw< II Time Distribution Used to Match Bus Headways
3.2.2 Simulation Calibration
To calibrate the simulation, several parameters in CORSIM were modified and field and
simulated traffic measures were compared. Calibration Parameters
Car-following parameters and the rubbernecking factors were modified within CORSIM to
better represent the field conditions.
21 Car-following Parameters
CORSIMs car-following model is the Modified Pitts model.
headway(t + At) = L + m + kVf, + bk(V,, VfJ) (3.1)
L = vehicle length
m = buffer distance or car-following constant (default is 10ft)
k = driver sensitivity factor (default is 0.6 to 1.5 based on driver
b = 0.1 if Vf>V,
0 otherwise
V, = following vehicle speed
V| = lead vehicle speed
Field and simulated volume, speed, and headway comparisons were made for different
values of m and k values. The best results were obtained for m=5ft. This value was used
for both incident and non-incident cases. Rubbernecking Factor
This factor is used in CORSIM to simulate incident conditions. As an incident or lane
blockage occurs, capacity is reduced in adjacent lanes as drivers slow down. This is
modeled in CORSIM as a rubbernecking factor. Essentially, the headway of each vehicle
within the incident section is increased by the rubbernecking factor:
f -1 I'i 0\
J rubber-necking ~ , , 1 W
he£ dwayinCident = headway of vehicle within the incident section
hec dway = headway of vehicle outside the incident section
The length of the ini ident section in CORSIM is defined as the length of the vehicles involved
in the incident plus one vehicle. This capacity reduction parameter can be applied in any
unblocked lanes. It could also be used to simulate an incident where partial or no lane
blockage occurs.
To calibrate the incident case, field and simulated volume, speed and headway data was
compared. The incident and rubbernecking factors that provided the best fit to the field data
were as follows:
One lane blocked for 5 minutes with a rubbernecking factor of 33% on all other lanes.
A rubbernecking factor of 5% across all lanes for five minutes after the lane blockage
is cleared.
22 Volume and Speed Comparison
Five-minute field and simulated detector speed and volume data were compared at Evans,
Colorado NE, Colorado, NW, University, Downing and Logan for the incident case and non-
incident case. The CORSIM output file includes the detector data for specified time intervals.
The detector reader program (section ) was used to extract the detector data from the
output file. The simulated detector volume and speed were compared for every detector
location (Figure 3.13).
Figure 3.15 shows a comparison of field and simulated detector volumes for the non-incident
Corsim 5 min Volume Comparison
I-25 N6 12-9-96 Incident Case
Figure 3.15: CORSIM Volume And Field Volume Comparison at Downing
The final calibration results for all volume and speed data for both the incident and non-
incident cases are shown in Table 3.3. These calibration results are in agreement with
previous calibration effort using 30-second data [19;20] for incident and non-incident data.
Corsim vs. Actual Slope R2
Incident Volume 1.020 0.28
Incident Speed 1.070 0.61
Non-Incident Volume 0.922 0.66
Non-Incident Soeed 1.004 0.32
Table 3.3: Regression Statistics for CORSIM Calibration

Corsim Volume vs Actual Volume
1-25 NB 3/24/99 Non-Incident Case
Actual Volume (vph)
Figure 3.16: CORSIM Volume vs. Field Volume Bu Travel Speed Comparison
The TSD Read* r program discussed in section 2.3.2 2 was used to extract the bus position
information for t ie first and last links in the network and the travel time was calculated. Using
this travel time i iformation and the distance from the beginning to the end of the network, the
space mean sp ed was calculated. Figure 3.17 shows that bus space mean speed
calculated from '.he video and from the AVL system is highly correlated to the bus speed
estimated from he simulation. It may also be noted the correlation of the GPS speed and the
video speed wt re also found to be high (0.92), an indication that the positioning system
accuracy is high, given both real-time correction and the integration of GPS and inertial
The individual bus travel time calculated from the video was also used to calculate average
travel speed and variance of travel speed. The individual bus travel time obtained from the
location output of the TSD Reader was used to calculate average travel speed and variance
of travel speed for the simulation. The travel speed variance for both the field and simulation
is shown in Table 3.4 and was not found to be significantly different at the 95% level of

Bus Travel Time Comparison
CORSIM Bus-travel Time
Figure 3.17: Comparing Simulation and Field Bus Space Mean Speed for Incident Case
Bus Speed Variance Field 169.40
Bus Speed Variance CORSIM 171.06
Table 3.4: Variance of Bus Travel Speed
3.2.3 Incident Detection Data
Final parameters based on the calibrated CORSIM test network for both incident and non-
incident cases were used to generate simulation runs to develop and test incident detection
models. A program was written to develop a variety of CORSIM simulation cases. The TSD
Reader and Detector reader programs (section 2.3.2 ) were used to read the CORSIM output
files and a database was created to store this information. Another program was developed
to calculate the bus location and speed based on the reporting interval. Bus data and
detector data was aggregated into 30-second intervals in the database. The following
sections describe this in detail. Incident and Non-Incident Cases
Incident and non-incident cases were developed based on several variables. These
variables include flow level, incident location, incident duration, bus headway and bus report

The test network was divided into five sections for incident detection based on current
detector locations (Figure 3.18).
Table 3.5 shows the incident and non-incident cases developed and the file naming scheme.
_______________________________________# in eight digit filename (12345678 trf)_______________________
1 2 3 4 5 6 78
Run Flow Bus Headway I Incident Inc Duration | Inc Lane Inc Location
1 Random #1 1 Low (\4c=0.3) 2 2 minute 0 Non-Incident 0 Non-Incident 0 Non-Incident 00 Non-Incident
2 Random #2 2 Medium (vfc=0 6) 6 6 minute Incident 5 5 minutes 1 Right 31 Section 3. 1353 ft from Beginning of Section
n Random #n 3 High (vfc=0 9) 9 10 minute 9 10 minutes 2 Center 32 Section 3. 2706 ft from Beginning of Section
33 Section 3, 4059 ft from Beginning of Section
41 Section 4, 396 ft from Beginning of Section
42 Section 4, 791 ft from Beginning of Section
43 Section 4, 1187 ft from Beginning of Section
Table 3.5: File Naming Convention Ba ed on Variables
The first digit represents the run number used to identify th j different non-incident runs. The
second digit represents the flow level. (The v/c ratio was u ^ed to calculate the volumes for
each flow level based on the capacity of the base, non-inci lent simulation. The v/c ratios
used were 0.3, 0.6, and 0.9 for low, medium and high flow evels respectively.) The third digit
reprt sents the average bus headway used in the simulation. (The custom dwell time
distri >ution described in Section was used along vr th 2, 6, and 10-minute average
heac vay) The fourth digit indicates whether or not the sinr jlation included an incident. The
fifth (igit indicates the duration of the incident. (All incidents began five minutes into the
simu ation and lasted either five or ten minutes.) The sixth digit designates the lane in which
the ii cident occurs, right or center. The seventh and eight i digits indicate the incident
locat on; the section number and distance from the beginn' lg of the section to the incident.
Incid ,-nts were only simulated on sections three and four e id were located at specified
dista ices from the upstream end (30, 60, and 90 percent c f the section length.) The total
nurr er of simulation cases is shown in Table 3.6
| Incident I Non-Incident Variables Cases T i Cases
Run Flow Headway Incident Inc Duration Inc Lane Inc Location 1 9
3 I 3 3 i ; 3 J 2 2 6 i
Total 216 81 297
Table 3.6: Total Simulation Cases

/ Buchtel /l O) c c o Section 5 £ z CO i c /^^^x^Section 4 ID Z 'e 5 c D Section 3 £ z o T3 (0 o o ^/^ection 2 LD Z o TJ ns L_ o o O Section 1 \ Evans

m m m m S Ei a m a m a m I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I a|a|a aiaia a a a m a a m m m a s m a a a a m a
< 3514 * t 1319 r 4510 . t 778 > ^ 1424 > <4553 (Ne
16183 ^ 12669 < ^ (11350 , 6840 f 6062
Detector Corsim Bus Station ^
Figure 3.18: Section Diagram For I-25 Used for Incident Detection.
27 CORSIM File Generator
In order to create the 297 simulation files, an Excel/VBA program was written. A
representation of the base '.trf file was imported into excel as a fixed width text file. Since
most of the cards in the CORSIM input file have column widths in multiples of four units, the
column width for the excel sheet was set to 4 units. The cards that did not have four unit
columns were formatted so that when the excel sheet was saved as a space delimited text
file, it matched the original '.trf file. This excel sheet, [trf master] (where [ ] represents a
sheet name), was then saved as an excel workbook. This excel workbook contained all of
the variables discussed in the previous section in a different worksheet [variables]. An
additional sheet contained a list of the file names to be created with each of the digits in a
separate column [filenames]. Seven additional reference cells were linked by if statements
to the cards in [trf master] and the variables to which they pertained. As the number in this
reference cell changed, so would the corresponding value in the [trf master] sheet. For
example, one sheet contained all of the volumes and turn percentages associated with the
three flow levels. If the reference cell for flow had a value of 2, then the card types in the [trf
master] sheet would show the volumes and turn percentages pertaining to a medium flow
level. Therefore, all possible combinations of variable values could be reflected in the '.trf file
by changing seven reference cells. A VBA program was written that changed the reference
cells and saved the resulting .trf file for the 297 variable combinations (Appendix D).
CORSIM was run in batch mode in groups of 50 simulations because of disk space
limitations (30-60 Mb for each simulation). The TSD Converter (Section ) was also
run in batch mode and the second by second bus date was extracted for the group of 50
simulations. Simulation Database
The next step in processing the simulation date was to setup the bus and detector data in a
database. Microsoft Access was used to deve'op the simulation database. Detector Data
Six detector stations were simulated and located at the beginning and end of each of the five
sections. The detector reader progiam described in section was used to extract the
30-second detector data for each file. These values were then imported into the detector
table in the database. The fields int luded in the database were as follows:
File Name
1s1 through 7th digit in File Name (as separate fields)
30 Second Time Step
Station n Lane m Volume (each n,m combination as separate fields)
Station n Lane m Speed (each n,m combination as separate fields)
Station n Lane m Occupancy (each n,m combination as separate fields)
Station n Total Volume
Station n Average Speed
Station n Average Occupancy
28 Bus Probe Data
The bus data extracted from the CORSIM binary files contained the following fields:
File Name
Time Step (every second for each bus in the network)
Bus ID
Bus Position (Distance from the upstream node)
The second by second bus data from the simulation was modified to replicate RTDs fixed-
time location reports. This was accomplished by adding reporting intervals to the time at
which each bus entered the network. The bus location was then extracted for each report
time. The bus location, distance from upstream node, was converted into a network distance
from the beginning of the network (Figure 3.19). This was calculated from a relational
database table that contained the distance from each node number to the beginning of the
Mean speed was calculated for each bus based on the report interval and the distance from
the previous location. A section number was then assigned to this report location and
rounded to the nearest 30-second interval. The speed was assigned to the section the bus
last reported. Three separate tables were developed for report intervals of 10, 20, and 30
seconds that contained the following fields.
File Name
Actual Time Step
30 Second Interval
Bus ID
Network Location
Space Mean Speed

Section Number
The next step was to aggregate the bus data into 30-second vectors similar to the detector
data. This was done by averaging the bus speeds for eac h 30-second interval. The number
of buses that reported speeds during the same 30-second interval was placed in an
additional field: nBus. This aggregation of number of buses and bus speeds was repeated
for all five sections in each of the three report interval tables, 10, 20, and 30 seconds. The
Access VBA program and SQL queries are shown in Appendix E. 30-Second Vector Data
The bus data and detector data tables were joined on the common fields, Filename and 30
Second Interval. Several additional variables were calculated from the detector variables,
including differences between lanes and differences between stations. All variables were
also time-lagged by one and two intervals and treated as additional variables. For example,
vector t includes the total volume for station 3 and the total volume for station 3 at t-1 and t-2.
The final variables includec in the database were as follows:
Each Digit of the F lename
Vector Incident or Jon-Incident
Each of the Follow ng Variables at time t, t-1, and t-2 (T me Lagged by 30 Second
Station n Lane m Volume, Occupancy, Speed
Station n Lane m t id Lane m+1 Difference in Volume, Occupancy, Speed
Station n Total Vol ime, Average Occupancy, Average Speed
Station n and Stat' >n n+1 Difference in Total Volume, Average Occupancy, Average
Section n Average Bus Speed for 10, 20, 30 second Report Intervals
Section n Number of Buses 10, 20, 30 second Report Intervals
Each of the Previous Variables (Excluding the first 3 bullets) Time Lagged by One 30
Second Interval
Each of the Previous Variables (Excluding the first 3 bullets) Time Lagged by Two
one 30 Second Intervals
The final database table included sixty, 30-second intervals for each simulation and 297
simulations for a total of 17,820 vectors. This consists of 14,580 non-incident vectors and
1620 incident vectors in both sections three and four.


4. Methodology
Three different classifiers were developed for freeway incident detection: discriminant
analysis, generalized linear model and neural networks. This chapter presents the models
proposed, an overview of the classifiers, and the measurement vector for all classifiers.
4.1 Models Proposed
The followir g three models were developed based on detector date only, bus AVL data only
and both de tector and bus data. Detector data may not be available if the infrastructure is not
available. T us data may not be available if the frequency of servicr is not available. The
following m dels were developed to classify a measurement vector as an incident-state or
non-incident-state, every 30-seconds.
Bus Model (PBus) model developed based on bus data only, for time intervals bus data is
available; bus data may not be available every 30-second detection interval.
Detector Model (PDet )- model developed based on detector data only; detector data
available every 30-second interval
Detector and Bus Model (Pdet&bus) model developed based on both bus anc detector data,
for time intervals bus data is available; both bus and detector data may not be ; vailable every
30-second interval
Models were evaluated with an average headway varying from 2 to 10 minutes, reporting
location every 10, 20 or 30 seconds. The calibration set for all three methods was developed
for incidents located at the beginning and end of section 3. The incidents located in the
middle of section 3 were reserved for the test set.
4.2 Discriminant Analysis
A minimum-distance based classifier is used in discriminant analysis. Instead of using
Euclidean distance metric, which measures the distance between the measurement vector v
to be classified and a reference vector u, the following method is used to measure statistical
distance. This method accounts for the correlation between the measurement vectors:
dk(v) = (v~Pk)T K~liv~Pk) (4-1)

dk (v) = discriminant function
v = measurement vector
k = number of classes or groups
k - mean vector
K = covariance matrix
Figure 4.1 shows that the Euclidean distance between P, and the center of the cluster Q, is
greater than the Euclidean distance between O and Q. However, P appears to belong to the
cluster more than 0 does. Therefore, based on statistical distance, P is closer to Q than to
Figure 4.1: Euclidean Distance and Statistical Distance
This non-Euclidean metric and the corresponding minimum-distance classifier is known as
the Mahalanobis metric and the Mahalanobis classifier.
The forward stepwise method was used for model development. When stepping forward, the
variable that made the most significant contribution to the discrimination between the incident
and non-incident groups was selected based on the largest F value.

The Wilks lambda was used to test the discriminatory power of the model [21], Wilks lambda
statistic for the overall discrimination is computed as the ratio of the determinant of the within
group groups variance-covariance matrix (W) over the determinant of the total variance-
covariance matrix (T):
_ det(W)
Partial lambda represents the unique contribution of the respective variable to the
discrimination between groups. In other words, the partial lambda is the ratio of Wilks lambda
after adding the respective variable over the Wilks lambda before adding the variable.
The partial lambda is computed as the multiplicative increment on lambda that results from
adding the respective variable:
Partial A=-^s- (4.3)
The corresponding F-statistic is computed as:
(n-q-p) 1 partial A
(tf"1) partial A
where, n is the number of cases, q is < e number of groups, anc p is the variables.
4.3 Generalized Linear Model
The second statistical method applied to the incident detection data was a generalized linear
model (GLZ). GLZ is a generalization of the linear model which specifies a linear relationship
between a response variable Y, and a set of predictor variables, X, where the dependent
variable is normally distributed.
A linear model with response variable Y and predictor variables X would be the following
Y = (bQ + + ^2^2 ^ (4.5)
b0 = regression coefficient for the intercept
ty = regression coefficient for the predictors
e = error variability that cannot be accounted for by the predictors (assumed to
be zero)
The generalized linear model can be used to predict responses both for dependent variables
with discrete distributions and for dependent variables which are nonlinearly related to the

predictors [21]. This is accomplished by defining a distribution of the dependant variable,
binomial for this case, and a link function, which was assumed to be logit for the incident
data. A generalized linear model is represented by the following equations.
Y ~ g(bo + +b2X2 +bkXk +e) (4 6)
g() = a function
The inverse of g(), say f(), is called the link function.
f(U) = b0+blX1+b2X2+bkXk (4.7)
U = expected value of Y
The link function used for this data is:
/(w) = log(w/(l-w)) (4.8)
also known as the logit link function.
4.4 Neural Netv ork Model
In recent years neur< I networks (NN) have been applied to a variety of pattern recognition
problems and outper'ormed some statistically based techniques. The performance of the
statistical classifiers : uch as the DA, depends on how well the assumptions about the density
functions fit the data hey describe.
An artificial NN consists of many processing elements (PEs) that are connected to each
other. The processing elements can be arranged in layers, with an external layer rece ding
input from data sources, which is passed onto other processing elements through
interconnections, and on to an output layer of processing elements. Since the structur s are
distributed in nature, they are known to capture highly nonlinear mappings between inf Jts
and outputs. One NN architecture, the multi-layer, feed forward NN was used in this s jdy
Three separate structures were developed for the three models, PDet, PBus, and Pdet&bus-
The NN consisted of inputs on the first layer corresponding to the number of variables for
each model, twelve units on the hidden layer for each model, and one output unit. The neural
network structure for the Pdet&bus model is shown in Figure 4.2.
Each structure was then trained and tested; one structure for the Pdet model and three
structures for the PBUS and Pdet&bus, one for each bus report interval.


5. Results
This chapter presents the measures used to evaluate the performance of the incident
detection models, the variables selected for the bus model (Pbus), detector model (Pdetector)
and bus and detector (PBus & detector) model. The parameters for each classifier are shown
in Appendix F.
5.1 Performance Measures
Incident-State Detection Rate (ISDR): An incident detection model is presented with an
incident or non-incident vector, every time interval. The ratio of the total number of incident
vectors classified correctly by an incident detection model and the total number of incident
vectors presented is the Incident-State Detection Rate. This is also referred to as the
classification rate for incident state vector.
r , ^ . Incident Vectors Classified Correctly _
Incident State Detection Rate = ---------------------------------------xl 00 (5.1)
Total Incident Vectors
Incident Detector Rate (IDR): The number of incidents correctly detected based on the total
number of incidents.
Incident Detection Rate (IDR)
Number of Incidents Detected ,
Total Number of Incidents
False Alarm Rate (FAR): The ratio of the number of non-incident vectors classified
incorrectly and the total non-incident vectors
. _ Non Incident Vectors classified Incorrectly ^
False Alarm Rate =----------------------------------------------xlOO (5.3)
Total Non incident Vectors
Time to Detection (TTD): Time elapsed between the beginning of an incident and when an
incident detection model first detects an incident.
A persistence test was also used in this study. A persistence test checks consecutive vectors
for incident classification. An n level persistence test would not recognize an incident until n
successive vectors were classified as incidents. A persistence test is used to reduce false
Typically, the IDR, FAR and TTD is used to report the performance of an incident detection
model. This study proposes ISDR as an additional performance measure. The need for using
ISDR may be explained by Figure 5.1. For example, the figure shows two incident detection
models. Model-1 detects almost the entire duration of Incident #2, i.e. high ISDR, however

model-2 detects the incident for shorter duration or intermittently both models detect all
three incidents. It is likely that a traffic operator may ignore an intermittent alarm. This
demonstrates he importance of using both performance measures ISDR and IDR to report
the performam e of detection models.
co 0 8
- 0.6
s 0.4
Incident #1
Incident #2
20 25
Time Step
Incident #3

1 "
A i .. (.' i1 i i \ ....
J f.
: i ; 1 i' 1' i1 i i
1 1 i : i i : i i i ; i i1 i1 i1 i1 i' i' i11' 1111 i111 1 1! i11 1 1 i ! j! i :i | i i :! 1 i i
: i : : i i : t : t 11 11 | | i 11 1 i h 1 1 1 1 1 h' ' f | 1 | 1 | i | 1 |
! ) i 1 1
- Model-1
Figure 5.1: Incident State Reported By Two Models
5.2 Variable Selection
The variables used in the model development are as follows:
Vu (t),Vd (/) = upstream and downstream station total volume
Ou (t),Od (/) = upstream and downstream station average volume
Su (t),Sd (/) = upstream and downstream station average speed
Su_d (/) = upstream and downstream speed difference

Q-d (0 = uPstream anc* downstream occupancy difference
O (t),Od^di (/) = speed and occupancy differences by lane for both upstream
and downstream locations
BS(t) = average bus speed
nBus(t) = number of bus reporting
These variables lagged one and two time intervals were also considered, e.g.
Table 5.1 shows the significant variables selected for all three models (Pbus, Pdetector, Pbuss.
detector), based on discriminant analysis, The performance on the PBus and PDet&bus
model is reported for vectors with bus data available only. These variables were used for
subsequent analysis of all three classifiers: discriminant analysis, generalized linear model
and the neural network model.
Variable Detector Model (PdbA Bus Model (Pbus) Bus and Detector (PofcT 4 bus)
Upstream Station Volume (Vu t) t-1, t-2* t-2
Upstream Station Speed (Su t) t
Downstream Station Volume (Vd t) t, t-1, t-2 t, t-1, t-2
Downstream Station Speed (Sd,t) t
Upstream & Downstream Speed t t
Difference (S^t)
Upstream Occupancy Difference
by Lane ((9^)
Downstream Occupancy t t
Difference by Lane (Od ^)
Average bus speed t t
Number of bus reportings t
Wilks Lamda 0.36 0.681,0.68",0.72-3 0.341,0.3420.35s
I F 1150 519.91,483.72,353.43 609.51,555.6*,479.23
*(t-n) data lagged n time interval
1, 2,3 for 10, 20, 30-second reporting intervals
Table 5.1: Variables Selected for Each Model and Model Performance
On Calibration Data

5.3 Detector Model (Pdet)
Results for the PDET model based on the discriminant analysis, generalized linear model, and
neural network model is shown in Figure 5.2. This plot shows that without a persistence test
the ISDR is between 80% to 85% and the FAR is between 0.3% to 0.6% for all classifiers.
Introducing a persistence test lowers the FAR to 0.10% with the ISDR between 60-65%.
Figure 5.2: DR vs. FAR for the Detector Models.
(NN= neural network, DA=Discriminant analysis, GLZ= generalized linear mode)
However, for a zero persistence test, the generalized linear model outperforms the DA and
NN models (higher ISDR and lower FAR). With one persistence test, the GLZ and DA
models are about the same while the NN model FAR drops to zero. The FAR is zero for all
classifiers for the higher persistence test. The IDR for all classifiers (DA, GLAZ, and NN) is
5.4 Bus Model (Pbus)
Bus headway varies from 2 to 10 minutes and they report their location every 10, 20 or 30
seconds. Therefore, bus data may not be available every 30-seconds. The Bus Model (Pbus)
was developed only for time intervals when bus data was available, however the ISDR and
FAR calculated was based on the total available vectors, not bus vectors only. This implies
that vectors without bus data available were classified as non-incident and vectors with bus
data were classified as either incident or non-incident. This allows the models performance

to be compared directly to the Detector Model (Pdet)- The performance of the bus model is
summarized in Table 5.2
Model Report Interval IDR ISDR FAR
DA 10 sec 75.00% 44.17% 0.766%
DA 30 sec 75.00% 40.28% 0.862%
GLZ 10 sec 70.83% 37.78% 0.000%
GLZ 30 sec 70.83% 33.33% 0.000%
NN 10 sec 70.83% 38.06% 0.000%
NN 30 sec 70.83% 35.83% 0.000%
Table 5.2: Performance of the Bus Models
The f N and GLZ models have a zero FAR and very similar ISDR and IDR. The detector
mod( I (Pdet) performs significantly better than the Bus Model (PBus) in terms of ISDR as
expe> ted, since bus data is not available every time interval however the IDR for Peus is
70%. The higher ISDR reflects the duration of the incident the incident alarm remains on.
Additional testing with one-interv; I persistence test has shown that the reporting frequency
does not affect the relative perfoimance of the bus models.
Figure 5.3 shows the performani e of a DA bus model based on average bus headway and
reporting interval tested on allectors (not only when bus data available). The figure shows
that the performance of the mod* I does not improve significantly by increasing the reporting
frequency, whet, the average bu. headway is high. For average bus headway of 2 minutes,
increasing the n porting interval 1 om 30 seconds to 20 seconds lowers the FAR
considerably; hi wever, increasing from 20 to 10 seconds does not improve the performance
as much. After applying one-per istence test (1 PT), there was no significant difference in
the performanct of the model for a given headway. Average bus headway has a significant
impact on the models perform; nee lower average headway provides better overall
performance on ISDR, IDR, and FAR.

Discriminant Analysis Bus Model
Performance on Test Set
0.00% 0 50% 1.00% 1.50% 2 00% 2 50%
FAR: False Alarm Rate
Figure 5.3: Performance of Tus Model for Different Average Bus Headway and Report
Frequency (H2= rimin headway, R 30=30-sec reporting interval)
The effect of the number of bu: reportings per int .-rval on the performance of the model may
be examined by observing the tatistical distance between state vectors and the incident and
non-incident groups. Figure 5.4 shows a plot of t ie squared Mahalanobis distance for the
Bus Model, for incident cases ( nly. Cases or measurement vectors based on the same
number of bus reportings per ir terval form arcs, "he arcs formed by fewer bus reportings per
time interval appear on the lower right corner of tl e plot, and a higher proportion of the points
on the same arcs fall on the non-incident side (Mt>M0). This shows that a higher number of
incident vectors are missed for fewer reportings per time interval. However as the number of
bus reporting increases, a higher proportion of the points on the same arc appear on the
incident side (M0>M1). For example, for 9 bus reportings, all points on the curve are on the
incident side, i.e. all incident points are correctly classified. The same effect is observed for
the non-incident vector plots. This, as expected, suggests that both FAR and the ISDR are a
function of the number of buses reporting. Therefore, the performance of the model may be
improved by introducing a persistence test when the number of bus reportings is low; and a
persistence test may not be necessary when the number of bus reportings is high (providing
lower TTD).

Figure 5.4:' icident Cases Only on a Squared Mahalanobis-Distance Plot for 2-min
Average But Headway, and 10-Second Reporting Interval (nBus= number of buses)
5.5 Detect r and Bus Probe Model (Pdet&bus)
For the time interval; both bus and detector data was available, the detector and bus model
(Pdet&bus) was dev< loped.
The Bus (PBUs) or Dt tector Model (Pdet) can be applied when either bus or detector data is
available. However, -'or certain highway corridors, both bus and detector data may be
available. For this case, both bus and detector data was used to develop the PDet& bus model.
As shown in Table 5.1, the same variables selected for the PDET model were selected for the
detector and bus me Je (Pdet&bus) model, except the upstream detector speed is replaced by
the average bus spe ;d reported in a given time interval.
Figure 5.5 shows th< performance of the PDet, Pdet&bus and the PBus model based on the
discriminant analysis, for intervals only bus reportings were available. It shows that using
both detector and bus data lowers the false alarm rate for all headway by 67%, with a less
than 5% decrease in ISDR.

a. 20%
80% -
60% -
Pdet. Pbus. Pdet&bus: Relative Performance
Pr| IS & DPT __________
0.00% 0.20% 0.40% 0.60% 0.80% 1.00%
FAR: False Alarm Rate
Figure 5.5: Relative Performance of the PDet, Pbus and Pdetsbus DA Classifier for
Intervals Bus Data Available
5.5.1 Combined Model
The Pdet & bus model was developed for 30-second periods, when bus data was available. A
combined model was developed, for every 30-second interval, as follows:
Combined = \PDET + a2PDET&BUS (54)
a, = 1 when only detector data available, else 0
a2= 1 when bus data available, else 0
This model applies the Pdetector s. bus model for time intervals bus data is available and the
Pdet model for time intervals bus data is not available.
Figure 5.5 shows that the Poets bus model performs better than the PBUS model in terms of
ISDR and FAR, and the Poets bus model may not perform better than the PDet model, but it
reduces the FAR considerably, with very little difference in ISDR. Figure 5.7 shows this more
clearly. For example, for the non-incident cases, fewer points were classified incorrectly by
the Pcombined model compared to the PDet\ however the PCombmed model missed a few incident
state vectors. This may lower the ISDR, however it does not lower the IDR. Therefore a

simple combined model based on Pdbt&bus and PDEt Model improves the overall
performance of the model.
5.5.2 Comparing DA, GLZ, and NN Classifiers for the Combined Model
Figure 5.6 and Figure 5.8 show the performance of the DA, GLZ, and NN classifiers for the
combined model, compared to each corresponding detector model, for 10 and 30 second bus
reporting intervals respectively. For the 10-second report interval the combined NN model
outperforms all other models at both the zero and one persistence level. With a bus report
interval of 30 seconds, the GLZ combined model outperforms the other models with zero
persistence. With one persistence level, the NN and GLZ models are the best. Two
important points can be made when comparing these two figures.
For higher reporting intervals, the neural network classifier performs better than the
generalized linear classifier; for lower reporting interval, the performance of both
classifiers is comparable.
Comparing tf e Pcombinea to the corresponding PDET for both 10 and 30 second report
intervals, the Pcombinea model always outperforms the PDEt model (based on I DR,
ISDR, and F/ R) for each type of classifier (DA, GLZ, and NN).
The report in erval has very little effect on the performance of the GLZ based
combined me del.
DR vs. FAR (Detector and Combined Model; 10 Sec Rl)
Figure 5.6: DA, GLZ, & NN: Combined Model (Pcombinea) Models for 10 Second Bus
Report Interval and Corresponding Detector Model (PDet)

Incident State Vectors Non-Incident State Vectors
Figure 5.7: Squared Mahalanobis Distance of PDet and PCombined2 Model for Incident and Non-Incident State Vectors (3 Bus Reportings)

DR vs. FAR (Detector and Combined Model; 30 Sec Rl)
Figure 5.8: DA, Gl 7., & NN: Combined Model (PCombinad, Models for 30 Second Bus
Repor Interval and Corresponding Detector Model (PDet)
The models were also ested based on average bus headway. The only significant finding
was that as the averag : headway increased, the change in report interval had no significant
effect. This is expectet based on equation 5.1; as the number of buses decrease, the
combined model perfor ns more like the detector model.
Figure 5.9 shows the time to detect (TTD) versus persistence level for the detector and
combined models for each classifier.
The TTD is lowest for the NN detector model with zero persistence and the NN combined
model is lowest with one persistence level. However, the difference in average TTD for all
the models is a maximum of 15 seconds with zero persistence. The TTD for the combined
models is less than or equal to the TTD for the detector models in all cases except the NN
classifier with zero persistence. Even with one and two persistence tests, the TTD for all
classifiers for the detector and combined models is within one time interval (30-seconds).

Time to Detect vs. P rsistence Level
Figure 5.9: Time To Detect vs. Persistence Test for (NN=neural network,
DA=discriminant analysis, GLZ=generalized linear model)

6. Co iclusions and Recommendations
This study investigated the feasibility of using data from an automatic bus location system for
freeway incident detection. A Mahalanobis-distance based classifier, a generalized linear
classifier, and a neur. I network based classifier was applied to detect freeway incidents. In
this study, CORSIM, a microscopic traffic model was used to simulate bus AVL system. The
performance of the classifiers was evaluated b; sed on simulated detector and bus AVL
The study has shown that location data available as part of a bus AVL system may be used
to estimate bus speeds. Though typically an AVL system is not designed to provide speed
data, this di ta may be retrieved using a geographic information system (GIS) and therefore it
may providt valuable information about traffic f ow.
For networl s without fixed sensors such as loo) detectors, bus location data may be used to
detect up ti 70% of incidents, depending on th bus headway and reporting interval.
However, d le to unavailability of bus reporting, incidents may not be detected continuously.
Even thoug i the detection rate, using bus data only, may not be as high compared to using
detector da a only, the false alarm rate is signif ;antly lower. And therefore, combining bus
and detectc r data improves the overall perfornr mce of all the classifiers. For high average
bus headw; y, increasing reporting interval has no significant impact on ISDR and FAR. For
average 2-i linute bus headway, increasing reporting frequency from 30 seconds to 20
seconds improves the performance.
In addition, a comparison of the three classifiers reveals the following:
Detector Model: The generalized linear classifier outperforms the other classifiers
based on ISDR and FAR
Bus Model: The neural network and the generalized linear classifier perform about
the same with respect to IDR, ISDR, and FAR.
Pbus & detector: For intervals when only bus reportings were available, for all
classifiers, the Pbus & detector model shows an improvement over the Pdetector
model by reducing the FAR for all classifiers
Pcombmed: For all three classifiers, a combined detector and bus model improves a
detector model by reducing the false alarm rate. Without a persistence test, the
performance of both generalized linear model and the neural network model are
comparable, except when buses report very frequently
The TTD for all classifiers for the detector and combined models is within one time
interval (30-seconds).
This study also presents a procedure to simulate buses in CORSIM and identifies limitations.
The specific effort includes:
Calibrating CORSIM to simulate buses in a freeway network

Developing routines to create multiple simulation input files with different parameters
Simulating a bus AVL system: extraction of probe vehicle location and speed at pre-
specified intervals.
Limitations identified are as follows:
Buses cannot be simulated on a FRESIM network without an integrated NETSIM
The bus flow rate remains constant for each route during each CORSIM time period.
To simulate a route with time-varying bus headway, each bus on that route must be
coded in the simulation as a separate route.
The offset for a route begins at the start of the initialization period instead of at the
start of the simulation and the initialization period varies depending on the size of the
network and the network volume
CORSIM bus stations can only hold a maximum of 6 buses.
The bus measures are not reported for FRESIM links. They are only reported for
NETSIM links.
This study suggests that data from an AVL system for buses may be used by a traffic
management center for incident detection. The pei'ormance of a detection system based on
either bus data only or a combination of bus and d tector data depends on the bus flow rate
for the network. The focus of this study was to con pare the performance of three classifiers
for incident detection. It also considers the impact f average bus headway and frequency of
bus reporting on model performance. The performance of the generalized linear classifier
and the neural network model was comparable, he vever, with a one-interval persistence test,
the neural network classifier outperforms the generalized linear classifier. This work may be
extended to find the best method of combining detector and bus data for incident detection.

o o
Appendix A. Detector Reader
Written by Kittichai Thanasupsin
C Last change: T 2 Jun 1999 9:27 am
PROGRAM Reading_CORSIM_outputdata
CHARACTER infile*20, oufile*20,TEXT(90000)*130,EXTN*4
INTEGER nd,j, sta(90000), lane(90000), initime(90000)
1,endtime(90000),E(bbl,b,q ,ns,K(90000)
REAL vol(90000),speed(90000),headwy(90000),occ(90000)
*, TVOL(90000),volsum(90000)
5 READ(20,'(A8)1, END = 100)INFILE
infile = INFILE(1:8)//EXTN(1:4)
OPEN(l, FILE= infile, STATUS= 'old')
OPEN(2, FILE= oufile)
ND = 19
ND1 = 3
OPEN(10, FILE =l.txt')
c Parti
c read and search data from 1 and write in temp, file l.txt which
c will be deleted later on.
DO 1=1, 90000
READ(1,'(A)', END= 20) TEXT(I)
N = I
20 DO J = 1, N
IF (K(j) .NE. 0) THEN
jj = 0
DO 1 = 1 nd
JJ = J+13
WRITE(10,40) text(JJ+1)(53:54),text(JJ+1)(21:21)
1, text(J+4) (77:81) text(J+5) (74 :78),text(JJ+1) (94:97)
1,text(JJ+1) (123:128)
40 FORMAT(T3,A3,T8,A2,T14,A5,T23,A5,T30,A8,T40,A8,T50,A8,T60,A8)
c end of part 1
OPEN(11, FILE = 'l.txt')
DO I = 1, 90000

read (11, 50, end =70)sta(i),lane(i),initime(i),endtime(i),vol(i)
1,speed(i),headwy(i), occ(i)
50 FORMAT(T3,13,T8,12,T14,I5,T23,I5,T30,f8.0 ,T40,F8.3,T50,F8.3
1,T60,F8.3 )
E = I
70 B = 1
N = 0
DO J = 1, E
BB1 = J
volsum(j) = 0
speedsum(j) = 0
C begin to calculate average speed, avg. headway
DO i = B, BB1
volsum(BBl) = volsum(BBl)+vol(i)
speedsum(bbl) = speedsum(bbl)+speed(i)
headwysum(bbl) = headwysum(bbl)+headwy(i)
occsum(bbl) = occsumfbbl)+occ(i)
q = bbl-b+1
tvol(bbl) = volsum(bbl)
avgspd(bbl) = speedsum(bbl)/(q)
avghy(bbl) = headwysum(bbl)/(q)
avgocc(bbl) = occsum(bbl)/(q)
B = BB1+1
DO I = 1, E ,ND
Nil = N1-ND1+1
N22 = N2-ND2+1
N33 = N3-ND3+1
N44 = N4-ND4+1
N55 = N5-ND5+1
N66 = N6-ND6+1
*,INFILE(4:4),INFILE(5:5),INFILE(6:6) INFILE(7:8),endtime(i)
*, (VOL (II), II=N6,N66,-1) TVOL (N6)

*, (SPEED (II),II=N6,N66,-1),AVGSPD(N6)
1, (OCC(II),II=N3,N33,-1) AVGOCC(N3)
1, (OCC(II),II=N4,N44,-1),AVGOCC(N4)
1, (OCC(II) II=N5, N55,-1),AVGOCC(N5)
1, (OCC(II),II=N6,N66,-1),AVGOCC(N6)
99 FORMAT (A8,6(', ',A1),', ',A2, ', 1,15,76(', '(FI 0.3))
210 FORMAT(' Finished!!(',F8.3, CPU Seconds))

Appendix B. TSD Reader
Written by Brian Hoeschen (ActiveX by [14])
Sub ReadTSDFileO
' Macro to read vehicle information from a TSD file
' using the TSDReader ActiveX object

Dim TSD_Obj As Object
Set TSD_Obj = CreateObject('TSDProj.TSDReader)

'Place a single quote in front of the appropriate
'FileNan e option -
'the Inpi Box will let y< j specify a file location
'otherwii e the second jption lets you specify the path and file name
'FileNan e = InputBoxflnput TSD file name.", "Open TSD")
FileNam. = "C:\My Dc uments\UCD\Research\TSIS Projects\RTD l-25\l253241m.tsd"
'FileNan e = "C:\My D( cuments\TSIS Projects\RTD\tsdout.tsd"
TSD_Ot,.FileName = ileName
TSD_Ot(.IsOpen = True
If Not TSD_Obj.lsOpe i Then
MsgBo; ("Invalid file .ame entered!")
Exit Sub
End If
Open "Bus51 csv" For Output Shared As #260
'Select the field description and values
'you want to include in the output file. Make sure
'you choose the same fields in the loop below.
' Descriptions
' "Time_Step, _
' "VehJD", _
1 "Fleet", _
"V_Type", _
' "V_Length", _
' "LaneJD", _
' "Veh_Pos", _
' "Prv_USN", _
' "Turn_Code", _
' "Queue_Status", _
' "Acceleration", _
' "Velocity", _
' "Lane_Ch_Stat", _
' "Destination_Nd", _
' "Leader_Veh_ID", _
' "Follow_Veh_ID", _

' "Target_Lane"
' Values
' TSD_Obj.Time, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).ld, _
' TSDjDbj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Fleet, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).V_Type, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).V_Length, _
' TSDJDbj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).LaneJD, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Veh_Pos, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Prv_USN, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Turn_Code, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Queue_Status, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Acceleration, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Velocity, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Lane_Ch_Stat, _
' TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Destination_Nd, _
' TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Leader_Veh_ID, _
' TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Follow_Veh_ID, _
TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Target_Lane
Write #260, "Time_Step", _
"VehJD", _
"Fleet", _
"LaneJD", _
"Veh_Pos", _
"PrvJJSN, _
"Queue_Status", _
"Acceleration", _
"Velocity", _
"Lane_Ch_StfT, _
"Destination J.!d", _
Do While TSD_Obj.Re dNext
For i = 0 To (TSD_Ot,.VehLinks.Count -1)
i = 10
For x = 0 To (TSD_Obj.VehLinks.GetLink(i).Vehicles.Count -1)
'x = 0
If TSD_Obj.Time > 900 Then
If TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Fleet = 3 Then
If TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Prv_USN = 51 Then
Write #260, TSD_Obj.Time, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).ld, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Fleet, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).LaneJD, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Veh_Pos, _
TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Prv_USN, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Queue_Status, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Acceleration, _
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Velocity, _
TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Lane_Ch_Stat, _
TSD_Obj.VehLinks.GetLink(i). Vehicles.GetVehicle(x).Destination_Nd,
TSD_Obj.VehLinks.GetLink(i).Vehicles.GetVehicle(x).Target_Lane, Cr

Eid If
Enc If
End If
Write #260, "End of File"
Write #260, "File Name:" + FileName
Set TSD_Obj = Nothing
Close #260
End Sub

Appendix C. Bus VPS Data Extraction
Written by Kittichai Thanasupsin
C Last change: T 15 Feb 1999 9:25 am
program BUSIDXY
CHARACTER infile*20, oufile*20, text(90000)*80, busid(1000)*34
character inpt*20, bl(1000)*4
INTEGER k(90000), 1(90000),p(90000),q(90000)
WRITE(*,10) 1.Please enter BusID filename: '
READ(*,20) inpt
OPEN(3, FILE=inpt, STATUS='old')
write(*,10) 2.Please enter input filename: '
read!*,20) infile
open(l, FILE=infile, STATUS= 'old')
WRITE)*,10) 3.Please enter output filename: '
READ(*,20) oufile
OPEN(2, FILE= oufile)
WRITE(*, 10) 4.Please wait...............'
10 format(A\)
20 format(A)
WRITE(2, 100)
n =0
DO i = 1, 1000
READ(3, 25, END = 24) busid(i)
bl(i) = busid(i)
busid(i) = 'Location Report getting VLUID '//busid(i)
n = i
24 DO i = 1, 90000
read (1,25, end = 27) text(i)
m = i
END do
27 DO j = 1, n
DO i = 1, m
k(i) = INDEX(text(i), busid(j))
1(i+1) = INDEX(text(i+1), x = )
p (i) = INDEX(text (i), ' RSA Status Handler getting VLUID')
q (i+3) = INDEX (text (i+3) x = )
if (k(i) .NE. 0 .and. 1(i+1) .NE. 0) THEN
write (2,30) text (i) (2:9), text (i) (50:53), text(i + 1) (47 : 53),
text (i + 1) (60:65)
30 format (T2, A T13,A, T24, A, T35, A)
if (p(i) .ne. 0 .and. q(i+3) .ne. 0) then
if (text(i) (44 : 47) .eq. b1(j)) then
write (2,35) text(i) (2:9) text(i) (53:56) text(i + 3) (47:53) ,
35 FORMAT(T2, A T13,A, T24, A, T35, A)
100 FORMAT (T3,'TIME',T12 'BUS ID',T26, X', T37, 'Y')
110 FORMAT( ', 40('*'))

WRITE(*/*)' * *'
WRITE(* ) '*************Congratuation, It is finished***********'
WRITE(*,*)' * * *'
end program

Appendix D. CORSIM Input File Excel/VBA Program
Written by Brian Hoeschen
Sub IncidenttrfO
' Incidenttrf Macro
' Created 5/27/99 by Brian Hoeschen
Dim i As Integer
Dim FileName As Range
Dim OldFileName As String
OldFileName = "Master trf
For i = 81 To 296 Step 1
Sheets("File Names").Select
RangefH"). Select
ActiveCell.FormulaR1C1 = "=Rr' & i & "]C[-8]"
Rangef'JI"). Select
ActiveCell.FormulaR1C1 = "=R[" & i & "]C[-8]n
ActiveCell.FormulaR1C1 = =Rf & i & "]C[-8]''
ActiveCell.FormulaR1C1 = "=Rr & i & ]C[-8]"
Range("M1). Select
ActiveCell.FormulaR1C1 = "=Rf & i & "]C[-8]"
ActiveCell.FormulaR1C1 = H=R[ & i & n]C[-8]"
Range("01"). Select
ActiveCell.FormulaR1C1 = "=R[ & i & "]C[-8]"
'SheetsC'Master trf).Select
Set FileName = Worksheetsf'File Names").Range("Q13")
Sheets(OldFileName). Select
ActiveWorkbook.SaveAs FileName:= _
"C:\My Documents\Brian\Projects\Thesis\trf\" & FileName & trf, FileFormat:
xlTextPrinter, CreateBackup:=False
OldFileName = FileName
Next i
End Sub

Appendix E. Access VBA Program and SQL Queries
E.1 Bus Report Database VBA Program
Written by Brian Hoeschen and Robert Largent
Option Compare Database
Option Explicit
Private m_dbData As Database
Private Const FEET_PER_SEC_TO_MILES_PER_HOUR As Single = 0.681818181818
Sub Main()
On Error GoTo Main_Error
Dim recVeh As Recordset
Dim recFile As Recordset
Dim lgFileName As Long
Dim lgVehID As Long
Dim 'igStartTime As Long
Dim lgLastTime As Long
Dim igLoop As Long
Dim itSecLoop As Integer
Dim itlntervalLoop As Integer
Dim trSql As String
Dim trValues As String
Dim trFileName As String
Dim itStep As Integer
Dim itlncident As Integer
Dim ngSecAvgSpeed(4, 6) As Single
Dim igSecBusCount(4, 6) As Long
Screen.MousePointer = 11 'Hourglass
Set m_dbData = CurrentDb
m_dbData.Execute Query:=DELETE FROM [BusProbelO]"
m_dbData.Execute Query:="DELETE FROM [BusProbe20]
m_dbData.Execute Query:="DELETE FROM [BusProbe30]"
Set recVeh = m_dbData.OpenRecordset("StartLast Time)
If Not recVeh Is Nothing Then
Do While Not recVeh.EOF
IngFileName = CLng(recVeh.Fields("FileName"))
IngVehID = CLng(recVeh.Fields("VehlD"))
IngStartTime = CLng(recVeh.Fields("StartTime"))
IngLastTime = CLng(recVeh.Fields("LastTime"))
Call msDolt(lngFileName, IngVehID, IngStartTime, IngLastTime, 10)
Call msDolt(lngFileName, IngVehID, IngStartTime, IngLastTime, 20)
Call msDolt(lngFileName, IngVehID, IngStartTime, IngLastTime, 30)

m_dbData Execute Query:="DELETE FROM BusAggregateSpeed"
strSql = "SELECT DISTINCT FileName FROM Veh1;
Set recFile = m_dbData.OpenRecordset(Name:=strSql)
If Not recFile Is Nothing Then
Do While Not recFile.EOF
strFileName = CStr(mfVariantValue(recFile.Fields("FileName''),""))
If Len(strFileNsme) > 0 Then
For IngLoop = 30 To 1800 Step 30
Erase sngSecAvgSpeed
Erase IngSecBusCount
For intlntervalLoop = 10 To 30 Step 10
For intSecLoop = 1 To 5
strSql = "SELECT AVG(MeanSpeed) AS AvgMeanSpeed, COUNT(VehlD) AS SecBusCount" &
"FROM [BusProbe" & CStr(intlntervalLoop) & "]" & _
WHERE FileName = & StrFileName & "" & _
"AND Section =" & CStr(intSecLoop) & "" & _
"AND [30Seclnterval] = & CStr(lngLoop) &
Set recVeh = m_dbDca.OpenRecordset(Name:=strSql)
f Not recVeh Is Nothii g Then
sngSecAvgSpeed(( ntlntervalLoop /10), intSecLoop) = _
CSng(mfVariant' alue(recVeh,Fields("AvgMeanSpeed"), 0))
lngSecBusCount((i tlntervalLoop /10), intSecLoop) = _
CLngfmfVarian.' alue(recVeh.Fields("SecBusCount"), 0))
End If
Next intSecLoop
Next intlntervalLoop
intlncident = 0
If Mid$(strFileName, 5, 1) = "5" And (IngLoop >= 330 And IngLoop <= 600) Then intlncident = 1
If Mid$(strFileName, 5, 1) = "9" And (IngLoop >= 330 And IngLoop <= 900) Then intlncident = 1
strSql = "INSERT INTO & _
"BusAggregateSpeed (" & _
"FileName, [30Seclnterval], Incident," & _
Sec1AvgSpeed10, Sec1nBus10, Sec2AvgSpeed10, Sec2nBus10, Sec3AvgSpeed10,
"Sec3nBus10, Sec4AvgSpeed10, Sec4nBus10, Sec5AvgSpeed10, Sec5nBus10," & _
"Sec1AvgSpeed20, Sec1nBus20, Sec2AvgSpeed20, Sec2nBus20, Sec3AvgSpeed20,
"Sec3nBus20, Sec4AvgSpeed20, Sec4nBus20, Sec5AvgSpeed20, Sec5nBus20," & _
"Sec1AvgSpeed30, Sec1nBus30, Sec2AvgSpeed30, Sec2nBus30, Sec3AvgSpeed30,
"Sec3nBus30, Sec4AvgSpeed30, Sec4nBus30, Sec5AvgSpeed30, Sec5nBus30" & _
strValues = VALUES ( & _
strFileName & "," & CStr(lngLoop) & & CStr(intlncident) & "," & _
CStr(sngSecAvgSpeed(1, 1)) & "," & CStr(lngSecBusCount(1, 1)) & "," & _
CStr(sngSecAvgSpeed(1,2)) & & CStr(lngSecBusCount(1, 2)) & & _

CStr(sngSeoAvgSpeed(1, 3)) & & CStr(lngSecBusCount(1, 3)) & &
CStr(sngSecAvgSpeed(1,4)) & & CStr(lngSecBusCount(1,4)) & &
CStr(sngSecAvgSpeed(1, 5)) & & CStr(lngSecBusCount(1, 5)) & &
CStr(sngSecAvgSpeed(2, 1)) & & CStr(lngSecBusCount(2, 1)) & &
CStr(sngSecAvgSpeed(2, 2)) & & CStr(lngSecBusCount(2, 2)) & &
CStr(sngSecAvgSpeed(2, 3)) & & CStr(lngSecBusCount(2, 3)) & &
CStr(sngSecAvgSpeed(2, 4)) & & CStr(lngSecBusCount(2, 4)) & &
CStr(sngSecAvgSpeed(2, 5)) & & CStr(lngSecBusCount(2, 5)) & &
CStr(sngSecAvgSpeed(3, 1)) & & CStr(lngSecBusCount(3, 1)) & &
CStr(sngSecAvgSpeed(3, 2)) & & CStr(lngSecBusCount(3, 2)) & &
CStr(sngSecAvgSpeed(3, 3)) & & CStr(lngSecBusCount(3, 3)) & &
CStr(sngSecAvgSpeed(3, 4)) & & CStr(lngSecBusCount(3, 4)) & &
CStr(sngSecAvgSpeed(3, 5)) & & CStr(lngSecBusCount(3, 5)) &
m_dbData.Execute Query:=strSql & strValues
Next IngLoop
End If
recFile. JoveNext
End If
End If
If Not recFile Is Nothing Then
Set recFile = Nothing
End If
If Not recVeh Is Nothing Then
recVeh. Close
Set recVeh = Nothing
End If
If Not m_dbData Is Nothing Then
Set m_dbData = Nothing
End If
Screen.MousePointer = 0 'Default
Exit Sub
MsgBox "There was an error:" & CStr(Err.Number) & -" & Err.Description
Resume Main_Finish
End Sub
Private Sub msDolt(ByVal FileName As Long, ByVal Vehld As Long, ByVal StartTime As Long,
ByVal LastTime As Long, ByVal IntervalStep As Integer)
On Error GoTo msDolt_Error
Dim recAII As Recordset
Dim strSql As String
Dim IngDis As Long
Dim IngPrevDis As Long

Dim IngLoop As Long
Dim intSection As Integer
Dim sngMeanSpeed As Single
Dim lng30Seclnt As Long
IngDis = 0
IngPrevDis = 0
For IngLoop = StartTime To LastTime Step IntervalStep
sngMeanSpeed = 0
intSection = 0
lng30Seclnt = 0
strSql = "SELECT V.*, BN BusNetwork FROM Veh1 AS V, BusNetwork BN & _
'WHERE V.FileName = & CStr(FileName) & "" & AND V.VehID = & CStr(Vehld) & "" & _
"AND V.USN IN (1,2,3,4,7001,33,35,51,5,6,7,8,9,13,14,15,16,17,"&_
"18,19,20,21,22,23,24,25,259,26,27,28,29,30,31,34,36)" & _
"AND V.TimeStep-' & CStr(lngLoop) & "" &
Set recAII = m_dbData.OpenRecordset(s Sql)
If Not recAII Is Nothing And Not recAII.EC Then
With recAII
IngDis = CLng(.Fields("BusNetwork") Fields("VehPos"))
If IngLoop > 1 Then
sngMeanSpeed = CSng(((lngDis -1 gPrevDis) / (IntervalStep)) _
lng30Seclnt = Abs(lnt((-1 IngLoop, / 30) 30)
If IngDis >= 4568 And IngDis < 606! Then
intSection = 1
Elself IngDis >= 6062 And IngDis < 3840 Then
intSection = 2
Elself IngDis >= 6840 And IngDis < 11350 Then
intSection = 3
E Iself IngDis >= 11350 And IngDis < 12669 Then
intSection = 4
Elself IngDis >= 12669 And IngDis <16183 Then
intSection = 5
End If
strSql = "INSERT INTO & _
"[BusProbe" & CStr(lntervalStep) & "] (" & _
"ID, FileName, TimeStep, Link, USN, DSN, VehID, Fleet, VehType, VehLength, DriverType," &
"Lane, VehPos, PUSN, TurnCode, QueueStat, Accel, Vel, ChgStat, TargetLane, & _
"DestNode, LeadID, FollowlD, BusNetwork, NetDistance, MeanSpeed, Section,
"30Seclnterval" & _
'VALUES (" & _
CStr(.Fields("ID")) & & CStr(.Fields(FileName")) & & _
CStr(.Fields("TimeStep")) & "," & _
CStr(.Fields("Link")) & & CStr(.Fields("USN")) & & _
CStr(.Fields("DSN")) & & CStr(.Fields('VehlD")) & & _
CStr(.Fields("Fleet")) & "," & CStr(.Fields('VehType")) & "," & _
CStr(.Fields('VehLength")) & & CStr(.Fields("DriverType")) & & _
CStr(.Fields("Lane")) & "," & CStr(.Fields('VehPos")) & "," & _
CStr(.Fields("PUSN")) & ", & CStr(.Fields('TurnCode")) & & _
CStr(.Fields("QueueStat")) & & CStr(.Fields("Accel")) & & _

CStr(.Fields(Vel")) & & CStr(.Fields("ChgStat")) & & _
CStr(.Fields("TargetLane")) & & CStr(.Fields("DestNode")) & & _
CStr(.Fields("LeadlD")) & & CStr(.Fields("FollowlD")) & & _
CStr(.Fields( "BusNetwork")) & & CStr(lngDis) & & _
CStr(sngMeanSpeed) & & CStr(intSection) & & CStr(lng30Seclnt) &
m_dbData Execute Query:=strSql
End If
IngPrevDis = IngDis
End With
End If
If Not recAII Is Nothing Then
Set r> cAII = Nothing
End If
Next IngLoop
If Not recAII Is Nothing Then
Set recAII = Nothing
End If
Exit Sub
MsgBox There was an error:" & CStr(Err.Number) & -" & Err.Description
Resume msDolt_Finish
End Sub
Private Function mfVariantValue(ByVal Value As Variant, ByVal Default As Variant) As Variant
mfVariantValue = Default
If Not lsNull(Value) Then mfVariantValue = Value
End Function
E.2 Database SQL Queries
Distinct FileName
FROM [Master BusProbe20];
Distinct Fleet3 Query

ORDER BY FileName, VehID, TimeStep;
Distinct List of VehiclelDs for Fleet3
SELECT DISTINCT Veh1 .FileName, Veh1 .VehID
FROM Veh1;
StartLast Time
SELECT DISTINCT FileName, VehID, Min(TimeStep) AS StartTime, Max(TimeStep) AS
FROM [Distinct Fleet3 Query]
GROUP BY FileName, VehID;

Appendix F. Classifier Parameters
Discriminant Analyis
Modd Report Interval
Drscnminant Functor Analyse Summery
Classification Function*, grouping: INCIDENT
Step 9. N of vars > modal 9. Grouping INCIDENT (2 grps) W4kt' Lambda: .36327 approx F (9.5906)=1150 2 p Lambda Lambda 1,5906 p-leval Toler. (R-Sqr.) S3AS 0.725676715 0 6526874
S4TV 0.373793513 0 971648547 171 C78537 0 0127572224 0.872427762 S3TVT2 0 00285676 0 0065474
S3AS 0.366695739 0.984751642 91 45117168 1 63625E-21 0 383186238 0 606811762 S4TVT2 -0 000990108 0.0014963
S3TVT2 0 401732981 0 904259086 626 3140259 0 0 311876996 0.686021004 S4AS 7 00652504 6 4592228
S4TVT2 0.36548666 0.993936479 36 02976227 2 05876E-O9 0.478785574 0.521214426 S3S4SD 0.75044173 0.9646963
S4A8 0 39971903 0.908615145 5925714722 0 0.156176666 0.643823314 S4AO 15.6429081 14.76272
S3S4SO 0 387497246 0 937479436 393 8716431 0 0 34479332 0.65520668 S4TVT1 0.004090168 0.0030033
S4AO 0.374997705 0 968727767 190 6560516 0 0 059473565 0 940526426 S3TVT1 0.005535389 0 0057434
S4TVT1 S3TVT1 0.371029913 0.36351499 0.979067353 0.899326017 126 1481705 3.971546412 5 60269E-29 0.046321191 0 438166242 0.377313405 0.560611758 0.622666625 Constant -267.8299866 -230 52045
Discriminant Function Analysis Summary
Step 2. N of vars r model. 2; Grouping INCIDENT (2 grp*)
W*s' Lambda 66133 approx F (2,22231=519 87 p<0 0000
Wiles Partial F-remova
Lambda Lambda 1.2223 p-ievel
8 BS10 0.944645643 0.721253612 859.1328125
1- Tolar.
Tolar. (R-Sqr.)
0 0 944131651 0.055668146
Classification Functions, grouping. INCIDENT
G_1:0 G_2.1
p=.87152 p=.12848
S3BS10 0.315941632 0.1646641
S3NB10 2 470160246 2.6773329
Constant -10.32317638 -6.2255316
8 1B10 C.664622566 0.995189546 10.74531174 0.001061546 0.944131651 0.055666149
C enminant Function Analysis Summary
f p 2, N of vars ri model: 2; Grouping INCIDENT (2 grps)
h Iks' Lambda: 67726 approx F (2,20301=463 66 p<0 0000
Wiles' Partial F-ramo^ 1-To4ar.
Lambda Lambda 1,2030 p-level Toler. (R-Sqr.)
S3BS20 0.971636266 0 696865645 682 960063 0 0.968995512 0 031004486
S3NB20 0.676367119 0.99832195 3.412127495 0 064664375 0.968995512 0.031004486
Classification Functions, grouping. INCIDENT
G_1:0 G_2:1
p= 86326 p=.13674
S3BS20 0 262564399 0 1331478
S3NB20 3.564105015 3.7634764
Constant -6 891769409 -6.5716365
Discriminant Function Analysis Summary
Step 2, N of vars in model: 2; Grouping: INCIDENT (2 grps)
Wiles1 Lambda: .72138 approx F (2,1830)=353.41 pO OOOO
Wiles' Partial F-remova 1-Tolar.
Lambda Lambda 1.1830 p-ievel Tolar. (R-Sqr.)
S3BS30 0.970785675 0.743085364 632 7047729 0 0 947076797 0 052923203
S3NB30 0.721694661 0.999282777 1.313499212 0.251911906 0.947076797 0.052923203
Classification Functions; grouping. INCIDENT
G_1:0 G_2:1
p= 85325 p= 14675
S3BS30 0 29681376 0.1705546
S3NB30 11 47651672 11.737226
Constant -1267463521 -10.875577
Discriminant Function Analysis Summary
Step 7, N of vars It model: 7; Grouping INCIDENT (2 grps)
W*cs' Lambda: .34206 approx F (7,2216)=609 46 p<0 0000
Classification Functions, grouping. INCIDENT
G_1:0 G_21
p=.87152 p= 12846
Wiles' Partial F-remove 1-Toler. S4TVT1 0.002108175 0 0011697
Lambda Lambda 1,2218 p-leval Toler. (R-Sqr.) S3BS10 0 436441571 0.2787044
S4TVT1 0.347475141 0 964420896 35 10124588 3.61908E-09 0424666822 0.575333178 S4TV 0 004337872 0.000336
S3BS10 0 364675894 0 666756659 277.6156616 0 0 690276206 0.309723794 S3TVT2 0 003946286 0.0068607
S4TV 0.395523697 0 86463256 346 6562031 0 0.192194149 0.807805636 S3S4SD C 628067136 0 7922561
S3TVT2 0.40787679 0 836639975 426.7562092 0 0.404599369 0 595400631 S4TVT2 0002166137 0 0013392
S3S4SD 0 366316047 0 93376365 157 2825623 0 0 563254633 0 436745167 S4AO 0.636113964 -0.6252355
S4TVT2 0 347631395 0.98397645 36.11444655 2.16913E-09 0 492653101 0.507346809 Constant -42 51058576 -33.427639
S4AO 0 344666627 0.992436707 16.90322495 4.076686-05 0228232279 0.771767735
Discriminant Function Analysis Summary
Step 7, N of vars in model: 7; Grouping: INCIDENT (2 grps)
Wiles' Lambda: .34241 approx F (7,2025)=555 56 pO 0000
Classification Functions; grouping. INCIDENT
G_1:0 G_2:1
p= 86326 p= 13674
Wiles' Partial F-remova 1-Tolar. S4TVT1 0 002116212 0 0012144
Lambda Lambda 1,2025 p-leval Tolar. (R-Sqr.) S3BS20 0 42560944 02766351
S4TVT1 0.347754696 0.984635353 31.5988636 2.15778E-08 041758576 0 562414269 S4TV 0 004100337 0 0002816
S3BS20 0.36077569 0 699248004 226 8815765 0 0660371602 0 339628396 S3TVT2 0.003691636 0.0067023
S4TV 0.394724339 0 667470662 309 3726636 0 0.192204311 0.807795703 S3S4S0 0 637144206 0.7661364
S3TVT2 0.406449064 0 836321665 39C 5400696 0 0 405246286 0.594753742 S4TVT2 0.002110537 0 0012773
S3S4SD 0 364002496 0 940665272 127 6859665 9.35752E -29 0.541657627 0 458342373 S4AO *0.799617324 -0 5899548
S4TVT2 0.346220436 0.963319044 34 35201645 5.35623E-09 0 460327725 0.519672275 Constant -41.47926331 -32.633851
S4A0 0 345156312 0 992048442 16 23093224 5.81416E-05 0 230612639 0 769187331
Discriminant Function Analysis Summary
Step 7. N of vars In model: 7; Grouping. INCIDENT (2 grps)
Wiles' Lambda 35233 approx F (7,1825)=479 25 p<0 0000
Classification Functions, grouping. INCIDENT
<3-1:0 G_2:1
p=.65325 p= 14675
Wda' Partial F-ramova 1-Tolar, S4TVT1 0 002346567 0.0014458
Lambda Lambda 1,1626 p-leval Toier. (R-Sqr.) S3BS30 0.397460073 0.2811006
S4TVT1 0.356356711 0 983182192 31.21746445 2 65444E-06 0.410506427 0 689493573 S4TV 0.004042056 0.0002533
S3BS30 0 37645394 0.930976987 135 3062697 3.20171E-30 0 632957399 0 367042601 S3TVT2 0 003631428 0.0065456
S4TV 0410640717 0.658005285 302 0264893 0 0.193756297 0 806243716 S3S4SD 0 640705943 0.7869536
S3TVT2 0423616201 0.831724346 369 2365723 0 0 406305675 0.591694355 S4TVT2 0.002011716 0.0012154
S3S4SD 0 376325995 0.93624121 124 2839613 5.71324E-28 0 514173985 0.485826015 S4AO -0.815372407 -0.5594972
S4TVT2 0.356520359 0 962738912 32 05484009 1.73741E-08 0.464012365 0 515967635 Constant -40.6316452 -33.015331
S4AO 0 356970906 0 967004576 24 02894763 1 03301E-06 0 232166899 0 767613067

Generalized Linear Classifier
Modal Report Interval
Detector INCIDENT Parameter estmates INCIDENT Statist cs of goodness of fit
Dstrt>ution BINOMIAL Dstribution BINOMIAL
Unfc fund on: LOGIT Standard Link function LOGIT
Estimate Error Stat P Df Stat. Stat/Df
Interc -11 2744608 1.159789618 94 50028748 0 Deviance 5906 1178 27636 0199504671
&4TV 0 002564901 0.000242307 1120676302 0 Scaled Deviance 5906 1178 27636 0.196504971
S3AS 0.000558642 0.009144408 0 003732119 0 951286678 Pearson Cl# 5906 9525577046 16 12864383
&4AS 0 117024624 00191227 37 45036568 6 37691E-10 Scaled P. Cl# 5906 95255.77046 16 12864383
&4AO -0 000606057 0 076476446 5 96415E-05 0 993638163 Loglikellhood -589 1381801
S3&4SD -0 060994508 0.011715434 27.10595551 1 92604E-07
S3TVT1 9 13866E-0S 0 000129882 0 495073753 0.481672707
S4TVT1 0 000510465 0 000114691 19 80957568 8 5553E-06
S3TVT2 -0 001426472 0 000132449 115 9919841 0
&4TVT2 0 000178125 0 00011599 2358360718 0124613075
Scala 1 0
But 10 tec INCIDENT Parameter estimates INCIDENT Statistics of goodness of fit
Datrfcubon BINOMIAL Dstr button: BINOMIAL
Link function LOGIT Standard Link function. LOGIT
Estimate Error Stat P Df Stat Stat/Df
interc -1 329681338 0 264474025 25 27720779 4 96542E-07 Deviance 2223 1051.819028 0 473152956
S38S10 0108418069 0 005655026 367.5649505 0 Scaled Deviance 2223 1051 819028 0.473152956
S3NB10 -0 056033944 0 052127222 1 155508696 0 282398949 Pearson Cf# 2223 2679 269634 1 205249633
Scale 1 0 Scaled P Cf# 2223 2676 269934 1.205246633
Loglikellhood 525.9065138
But 20 tec INCIDENT Parameter estmates INCIDENT Statistics of goodness of fit
DatrtHJbon BINOMIAL Distribution BINOMIAL
Unfc function LOGIT Standard Ur* function LOGIT
Estimate Error Stat P Df Stat Stat/Df
Interc -1.18403054 0.227381178 27.11544791 1.9166E-07 Deviance 2030 1003.400755 0 494286086
S3BS20 0 106251957 0005446161 380 6212217 0 Scaled Deviance 2030 1003 400755 0.494286066
S3N820 -0 137786029 0.086677548 2.526952779 0.111916156 Pearson Cf# 2030 2644 313381 1.302617429
Scale 1 0 Scaled P. Cl# 2030 2644.313381 1.302617426
Loglikellhood -501.7003777
But 30 tec INCIDENT Parameter estmates INCIDENT Statist cs of goodness of fit
Definition BINOMIAL Distrbubon. BINOMIAL
Link function: LOGIT Standard Link function LOGIT
Estimate Error Stat P Df Stat Stat/Df
Interc -1 043109538 0 278061856 13 98202853 0 000184566 Deviance 1830 1044 323732 0.570668706
S3BS30 0 095711349 0 005326088 3229312922 0 Scaled Deviance 1830 1044 323732 0 570668706
S3NB30 -0.117824756 0.16927996 0484464379 0 486407065 Pearson Cf# 1830 2152 560433 1.176262532
Scale 1 0 Scaled P. Cf# 1830 2152 560433 1.176262532
Lofllikelbood -522 161666

Generalized Linear Classifier Continued
Model Report Interval
Det&Bue 10 MC INCIDENT - Perametei estmates INCIDENT Statutes of goodness of fit
Distribution BINOMIAL Distribution: BINOMIAL
Lrk find ion LOGIT Standard Link function LOGIT
Estrnate Error Stat P Df Stat Stet/Df
interc -10.2419568 1 488132882 47 36771697 5.8844E-12 Deviance 2218 320.2445357 0144384371
S4TV 0 002627717 0 000384634 46 67253556 6 36662E-12 Scaled Deviance 2218 320.2445357 0.144384371
S4AO 0 0388467 0110009786 0124694021 0 723998175 Pearson ChP 2218 2071 050357 0.93374678
S3S4SD -0.011348355 0.013580791 0 69825763 0.403369776 Scaled P CM* 2218 2071 050367 0 63374678
S3GS10 C 046142428 001113147 17.18287519 3 39483E-05 Loglkehhood -160.1222679
S4TVT1 0 001173546 0 000262349 20 0097848 7.70486E-06
S3TVT2 -0 00166691 0.000226788 54 02464326 1.67953E-13
&4TVT2 0 000519083 0 000251683 4 254371835 0.039149424
Scale 1 0
Det&Bua 20 sec INCIDENT - Parameter estmates INCIDENT Stetstic* of goodness of fit
Distribution BINOMIAL Distribution BINOMIAL
Lrfc function lOG(T Link function LOGfT
Estfrnate Error Stat P Df Stat Stat/Df
Interc -9 735260074 1.473205514 43 66851713 3 88979E-11 Deviance 2025 319 6784865 0157865919
S4TV 0 002629403 0 00C383244 47 07201631 6 &4253E-12 Scaled Davianca 2025 319.6784865 0157865919
S4AO 0 006618002 01C7350959 0 00380051 0 950842928 Pearson ChP 2025 1932 569669 0 954355392
S3S4S0 -0 01306229$ 0 014130568 0 854515584 0 355278032 Seeled P. ChP 2025 1932 569669 C 954355392
S36S20 0 039530204 0.011735739 11 34585531 0.000756165 LoglfceNhood -159 8392433
&4TVT1 0 001124531 0.000260865 18 58274888 1 62686E-05
S3TVT2 -0.001664247 0 000229457 52 60551589 4077B5E-13
S4TVT2 0 000557212 0 000252196 4 881503707 0.027145961
Scale 1 0
DetABue 30 sec INCIDENT Parameter estimates INCIDENT Statistics of goodness of nt
Distribution BINOMIAL Distribution : BINOMIAL
Lrti function LOGIT Standard Link function: LOGIT
Estimate Error Stat P Df Stat Stet/Df
Interc -9 380545672 1.457834973 41.40373073 1 23821E-10 Deviance 1625 317.6112767 0.174033576
S4TV 0.002732805 0 000389639 49.19184537 2 32114E-12 Scaled Deviance 1625 317 6112767 0.174033576
S4AO 0 026754668 0.107956452 0.061416607 0 804270785 Pearson ChP 162$ 1969.136182 1.07897873
S3S4SD -0 020459469 0 014079075 2111739918 0.146173038 Scaled P Ctf 1825 1966.136182 1 07897873
S36S30 0 023555034 0 01242397 3 59456766 0057668705 LoglikeHhood -158.8056383
S4TVT1 0 001109793 0.000258448 18.43696224 1.75434E-06
S3TVT2 0.001665469 0.000231074 53 20360277 3 00759E-13
S4TVT2 0 000622138 0 000255046 5 950246979 0.014715208
Scale 1 0
Neural Networks
Training: Back Propagation
Epochs: 100
Learning Rate: 0.1
Momentum: 0.1
Shuffle Cases

1. Ahmed, S. A., and A. R. Cook. Application of time-series analysis techniques to freeway
incident detection. In Transportation Research Record 841, TRB, National Research
Council, Washington, D.C., 1982, pp. 19-21.
2. Persaud, B., and F. L. Hall. Catastrophe theory and pattern in 30 second freeway traffic
data-implication for incident detection. Transportation Research Part A, Vol. 23, No. 2,
1989, pp. 103-113.
3. Ritchie, S. G., and R. L. Cheu. Simulation of freeway incident detection using artificial
neural networks. Transportation Research-C, Vol. 1, No. 3, 1993, pp. 203-217.
4. Cheu, R. L, and S. G. Ritchie. Automated detection of lane blocking freeway incidents
using artificial neural networks. Transprotation Research Part C, Vol. 3, No. 6, 1995, pp.
5. Abdulhai, B. A Neuro-genetic-based universally transferable freeway incident detection
framework. Ph.D. Dissertation, University of California, Irvine, 1996.
6. Teng, H., D. R. Martinelli, and P. Jiang. Enhanced freeway incident detection utilizing
traffic measurements from two contiguous detectors. Paper presented at the
Transportation Research Board Annual Meeting 1998, 1998.
7. Payne, H. J., and S. C. Tignor. Freeway incident-detection algorithms based on decision
tree with states. In Transportation Research Record 682, TRB, National Research
Council, Washington, D.C., 1978, pp. 30-37.
8. Stephenedes, Y. J., and Chassiakos. Application of filtering for incident detection.
Journal of Transportation Engineering, Vol. 119, No. 1, 1993, pp. 13-26.
9. Sethi, V., N. Bhandah, F. S. Koppelman, and J. L. Schofer. Arterial incident detection
using fixed detector data and probe vehicle data. Transportation Research-C, Vol. 3, No.
2, 1995, pp. 99-112.
10. Ivan, J. N. Neural network representation for arterial street incident detection data fusion.
Transportation Research-C, Vol. 5, No. 3/4, 1997, pp. 245-254.
11. Castle Rock Consultants, Denver regional transportation district automatic vehicle
location system, in Evaluation Final Report Prepared for the Federal Transit
Administration. 1998.
12. FHWA. CORSIM User's Manual. Traffic Software Integrated System 97 User's Guide.
Federal Highway Administration, U S. Department of Transportation., 1997
13. FHWA (1995) FRESIM User Guide, Office of Traffic Safety and Operations R&D, Federal
Highway Administration.
14. ITT-Kaman Sciences, TSIS User Manual, Version 4.2, 1998.
15. ORNL, ITRAF User Guide., ITRAF Version2.0, 1997.
16. Leonard, J., Active X Routines., 1998.
17. Image Sensing Systems, Inc., Econolite Control Products, Inc. Autoscope Wide Area
Video Vehicle Detection System. Supervisor User's Guide., 1998.
18. ESRI. ArcView GIS 3.1, 1998.
19. Cheu, R. L., W. W. Recker, and S. G. Ritchie. Calibration ofINTRAS for simulation of
30-sec loop detector output. In Transportation Research Record 1457, TRB, National
Research Council, Washington D.C., 1994, pp. 208-215.

20. Khan, S. /., and B. Hoeschen. Calibrating a microsimulation model using global
positioning system in buses. Transportation Research Board Annual Meetingfsubmitted
for publication), 1999.
21. StatSoft, Inc. (1999). Electronic Statistics Textbook. Tulsa, OK: StatSoft. WEB:
22. Heicht-Nielsen, R., 1990. Neuro Computing. Addison-Wesley, Reading, MA..
23. Lindeman, R. H., P. F. Merenda, and R. Gold. Introduction to bivariate and multivariate
analysis., New York, 1980.
24. Thomas, N.E., Multi-state and multi-sensor incident detection systems for arterial streets.
Transportation Research-C, 1999. 6(5-6): p. 337-357.
25. Ivan, J.N., J.L. Schofer, and F.S. Koppelman. Real-time data fusion for arterial street
incident detection using neural networks, in Proceeding of Transportation Research
Board, 74^ Annual Meeting. 1995. Washington, D.C.
26. Khan, S I. and S.G. Ritchie, Statistical and neural classifiers to detect traffic operational
problems on urban arterials. Transportation Research-C, 1999. 6(5-6): p. 291-314.
27. Bullock, D. and A. Catarella, A real-time simulation environment for evaluating traffic
signal systems, in Paper presented at the Transportation Research Board Annual
Meeting. 1998: Washington, D.C.
28. Cragg, C.A. and M.J. Demetsky, Simulation analysis of route diversion strategies for
freeway incident management,. 1995, Virginia Transportation Research Council.
29. Rouphail, N.M. and B.S. Eads, Pedestrain impedance of turning-movement saturation
flow rates: comparison of simulation, analytical, and field observation, in Transportation
Research Record 1578. 1997, TRB, National Research Council: Washington, D.C. p. 56-
30. Administration, F.H., CORSIM User Guide, 1996, Federal FUghway Administration, US-
DOT: McLean, Virginia.
31. Benekohal, R.F., Development and validation of a car following model for simulation of
traffic flow and traffic wave studies at bottlenecks, 1986. p. xix, 477 leaves.
32. Benekohal, R.F., Procedure for validation of microscopic traffic flow simulation models, in
Transportation Research Record 1320. 1991, TRB, National Research Council:
Washington D.C. p. 190-202.
33. Yang, Q. and FI.N. Koutsopoulos, A microscopic traffic simulator for evaluation of
dynamic traffic management systems. Transportation Research-C, 1996. 4(3): p. 129.
34. Owen, L.E. and G. McFlale, A framework for traffic model validation, in Paper presented
at the Transportation Research Board Annual Meeting 1997: Washington D C.
35. Mastbrook, S.A. and L.E. Owen, An illustration of the TRAF-FRESIM validation process,
in Paper presented at the Transportation Research Board Annual Meeting. 1996:
Washington, D.C.
36. Chang, G.-L. and A. Kanaan, Variability assessment for TRAF-NETSIM. Journal of
Transportation Engineering, 1990. 116(5): p. 636-657.