Citation
Audio gunshot detection and localization systems

Material Information

Title:
Audio gunshot detection and localization systems history, basic design, and future possibilities
Creator:
Graves, Jordan R
Place of Publication:
Denver, Colo.
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file. : ;

Thesis/Dissertation Information

Degree:
Master's ( Master of Science)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
National Center for Media Forensics, CU Denver
Degree Disciplines:
Media Forensics
Committee Chair:
Smith, Jeff
Committee Members:
Bregitzer, Lorne
Ganzert, Charles

Subjects

Subjects / Keywords:
Sound -- Recording and reproducing -- History ( lcsh )
Detectors -- Technological innovations -- History ( lcsh )
Detectors -- Technological innovations ( fast )
Sound -- Recording and reproducing ( fast )
Genre:
History. ( fast )
non-fiction ( marcgt )
History ( fast )

Notes

Subject:
For decades, law enforcement organizations have increasingly utilized audio detection and localization systems to identify potential gunshot incidents and to respond accordingly. These systems have grown from simple microphone configurations used to estimate location into complex arrays that seem to pinpoint gunfire to within mere feet of its actual occurrence. Such technology comes from a long and dynamic history of developing equipment dating back to the First World War. Additionally, though basic designs require little in terms of programming or engineering experience, the mere presence of this tool invokes a firestorm of debate amongst economists, law enforcement groups, and the general public, which leads to questions about future possibilities for its use. The following pages will retell the history of these systems from theoretical conception to current capabilities. This work will also dissect these systems to reveal fundamental elements of their inner workings, in order to build a basic demonstrative system. Finally, this work will discuss some legal and moral points of dissension, and will explore these systems' roles in society now and in the future, in additional applications as well.
Thesis:
Thesis (M.S.)--University of Colorado Denver. Recording arts, media forensics
Bibliography:
Includes bibliographic references.
General Note:
College of Arts and Media
Statement of Responsibility:
by Jordan R. Graves.

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:
861979868 ( OCLC )
ocn861979868

Downloads

This item has the following downloads:


Full Text
AUDIO GUNSHOT DETECTION AND LOCALIZATION SYSTEMS:
HISTORY, BASIC DESIGN, AND FUTURE POSSIBILITIES
by
JORDAN R. GRAVES
B.S.Northern Michigan University2010
A thesis submitted to the
Faculty of the Graduate School of the
University of Colorado in partial fulfillment
of the requirements for the degree of
Master of Science
Recording Arts, with emphasis in Media Forensics Program
2012


This thesis for the Master of Science degree by
Jordan R. Graves
has been approved for the
Recording Arts, with emphasis in Media Forensics Program
by
Catalin Grigoras, Chair and Advisor
Jeff Smith
Lome Bregitzer
Charles Ganzert
Date: 11/16/2012


Graves, Jordan, R (MS, Recording Arts, with emphasis in Media Forensics)
Audio Gunshot Detection and Localization Systems:
History, Basic Design, and Future Possibilities
Thesis directed by Dr. Catalin Grigoras.
ABSTRACT
For decades, law enforcement organizations have increasingly utilized audio detection
and localization systems to identify potential gunshot incidents and to respond
accordingly. These systems have grown from simple microphone configurations used to
estimate location into complex arrays that seem to pinpoint gunfire to within mere feet of
its actual occurrence.
Such technology comes from a long and dynamic history of developing equipment dating
back to the First World War. Additionally, though basic designs require little in terms of
programming or engineering experience, the mere presence of this tool invokes a
firestorm of debate amongst economists, law enforcement groups, and the general public,
which leads to questions about future possibilities for its use.
The following pages will retell the history of these systems from theoretical conception to
current capabilities. This work will also dissect these systems to reveal fundamental
elements of their inner workings, in order to build a basic demonstrative system. Finally,
this work will discuss some legal and moral points of dissension, and will explore these
systems5 roles in society now and in the future, in additional applications as well.
The form and content of this abstract are approved. I recommend its publication.
Approved: Catalin Grigroas
iii


DEDICATION
I dedicate my work to my parents and my brother, Rick, Beth, and Connor,
because their support and love have directly translated to my success.
They taught me how to be kind, focused, and passionate
with all that I do, and I cannot thank them enough for that.
I would also like to dedicate my work to my friends and family,
who are one in the same. Knowing that what I do may
impact all the wonderful people in my life keeps me
ever-steady and ever-motivated to accept only the best of myself.
Finally, I would like to dedicate my work to my soulmate, Wendy.
She is the driver of my ambitions and pushes me to achieve my dreams,
both of which I gratefully do with her by my side.
Nothing about me is truly perfect
but the people in my life make living it seem
pretty spectacular.
IV


ACKNOWLEDGMENTS
I would like to thank my classmates in the MSRA-MF program,
my professors and teachers at UCD and at NMU,
the faculty and staff of the NCMF,
my coworkers, supervisors, and mentors at USACIL,
and the countless others who have offered guidance for this project.
I could not have even begun this on my own, and I appreciate
every contribution you all have made along the way.


TABLE OF CONTENTS
CHAPTER
L INTRODUCTION.......................................................1
II HISTORY...........................................................3
III. BASIC DESIGN....................................................8
Planning...........................................................8
Equipment and Configuration.......................................14
Programming.......................................................17
Testing and Results...............................................24
Discussion........................................................28
IV. FUTURE POSSIBILITIES............................................32
The Current State of Gunshot Detection and Localization Systems...32
Prospects for the Basic System....................................34
V. CLOSING REMARKS.................................................40
REFERENCES..........................................................41
vi


Table
LIST OF TABLES
1.1 Test Trial1 Configuration and Results........................................27
1.2 Test Trial 2 Configuration and Results........................................27
1.3 Test Trial 3 Configuration and Results........................................27
2.0 Test Trial Potentially Useful Evidence Results................................28
Vll


LIST OF FIGURES
Figure
1.1 Sound Ranging Diagram.......................................................4
1.2 'Boomerang5 Gunshot Detection System, Outfitted to US Army Humvee...........6
2.1 Waveform of .22 Caliber Rifle Gunshot with Reflection......................10
2.2 Angle Determination from Sound Delay between Two Microphones...........11
2.3 Angle Determination from Sound Delay between Two Microphones (II).........12
2.4 Wide Angle Determination from Sound Delay between Two Microphones.........12
2.5 Distanced Angle Determination from Two-Microphone Sound Delay.............13
3.1 Gunshot Detection and Localization System, Basic Design Overall..........15
3.2 Basic System Workflow.....................................................16
3.3 Gunshot Detection and Localization System, Arduino and Servo..............17
3.4 Audio Detection and Image Response Script.................................20
3.5 Thresholds in Audio Detection and Image Response Script...................20
3.6 Discriminatory Thresholds for Audio Event Exclusion.......................21
3.7 Channel Delay and Angle Calculation in Detection and Response Script......22
3.8 Delay and Angle Calculation Example 1.....................................23
3.9 Delay and Angle Calculation Example 2.....................................24
3.10 FFT of .22 Caliber Rifle Gunshot with Reflection.........................30
4.1 Source Determination from Three Microphones in Three Dimensions............35
4.2 Source Determination from Three Microphones in Two Dimensions.............36
viii


CHAPTER I
INTRODUCTION
Until now, no known measure exists outside of science fiction that can successfully
predict and prevent crime before it takes place. Therefore, law enforcement agencies
must maximize the efficiency and accuracy of the response effort to criminal activity.
The sooner law enforcement personnel can have a presence at the scene where a crime
took place, the sooner the area may return to an orderly state of safety. Agents can defuse
dangerous situations, gather evidence, and build a case used to apprehend those
responsible for the crime.
Many factors limit the amount of influence law enforcement agencies have on the general
public, and some of these factors introduce obstacles when striving for timely and
effective crime response. Budgets and available funding dictate the amount of staff and
equipment an agency can use, and legislation regulates agency power. Law enforcement
groups may intervene in a given situation, but only when logistically capable and legally
permitted. While positive intended results of this control include safe and cost-effective
law enforcement, unfortunately negative side-effects also may arise. Agencies may be
understaffed or ill-equipped, and may be restricted from responding as quickly or as
soundly as desired. It is then vital for law enforcement groups and the people they protect
to reach a compromisekeep the enforcers of the law within clear legal and financial
boundaries while providing them with enough tools to help maintain safe communities.
With that compromise in mind, law enforcement groups have begun utilizing the gunshot
detection and localization system. This technology provides information for law
1


enforcers in two regards: it identifies possible gunshot events based on audio information
acquired by microphones and interpreted by algorithmic processing, and it also provides
the perceived location of the sound source. The system is semi-automatic, which is to say
it operates largely by automated computer programming but still requires human
interface to complete its task as designed. Installed systems passively listen for specific
audio characteristics and alert operators of potential detected gunshot events, but the
decision to include or exclude an audio event as a gunshot (and requiring response at the
scene) still belongs to a human at the controls.
The following sections include a history of audio gunshot detection systems, a simple
design plan for a basic system, and a discussion of the potential problems facing the
implementation of these systems, with some speculation on their future use in law
enforcement and for other applications.
2


CHAPTER II
HISTORY
The origins of many technological advances are often traced back to innovations in
different fields, later made applicable through simple redesign. Modern-day gunshot
detection systems share similar roots.
The onset of World War I brought about a technique known as sound ranging which
provides information regarding the coordinates of artillery weaponry. Developed by
William Lawrence Bragg, a British military officer and physicist, initial sound ranging
techniques involved arrays of microphones carefully placed in the field of battle to detect
sound events from the fired weapons and report back to a monitor at an operating base, as
depicted in Figure 1.1. At times, the resulting information contained valuable clues about
the sound events origins. Though the techniques success was less-than-desirable in the
early years, nations from each of the opposing sides made adjustments to the process to
find increasingly useful results.1
3


By World War II, most major military players used sound ranging for mortar detection
and counter-artillery measures. In particular, British forces and United States Marines
made good use of sound ranging in defensive operations.2 Although sound ranging
equipment had been growing more sophisticated and less costly over the years, radar
systems and aerial surveillance took over as the primarily-used gun locating methods in
military operations.3 Radar operators were capable of locating large weaponry faster,
derived more conclusive data in settings of extreme terrain or overgrown vegetation,
equipment could be outfitted on more mobile units for determining the location of
airplanes and vehicles as well, and most importantly, radar could operate without waiting
for shots fired.4 Sound ranging still held a place in combat, but acted mostly as a backup
to rapidly expanding radar capabilities.
Techniques involving sound ranging for gunfire location receded in popularity until as
recently as the 1980s and 1990s. Researchers borrowed sound-ranging techniques from
seismologists studying earthquakes and began testing capabilities of detecting small arms
4


activity in urban areas.5 Organizations such as ShotSpotter Incorporated, now SST Inc.,
tested detection and localization systems in areas with high crime rates, and US police
departments along the Californian Pacific coast began working with the technology to
improve incident response time and subsequently help deter future crimes.6
Meanwhile, the military returned to using gunfire detection and location in combat zones,
mainly to assist in evading and countering enemy sniper attacks.7 Technology is now
mountable both to vehicles and personnel, and war fighting units currently rely on these
tools in the Middle East and other theaters worldwide.8 9 Figure 1.2 consists of a
Boomerang system outfitted to a US Army Humvee.
5


Figure 1.2 Boomerang Gunshot Detection SystemOutfitted to US Army Humvee
Photo courtesy Marine Corps Warfighting Lab via Office of Naval Research
Back in the United States, agencies nationwide have deployed gunshot detection and
localization systems in cities and other urban areas that are prone to gunfire-related
crimes and random gunfire incidents. These systems are receiving more consideration as
a large contributor to community safety and law enforcement success, and offer not only
increased response capabilities but potential video evidence as well, incorporating video
6


capture components in the system designs.10 While some critics raise concerns including
costs, privacy issues, and accuracy, gunshot detection and localization systems used in
American cities have had a significant impact in the way authorities identify and respond
to criminal activity.
7


CHAPTER III
BASIC DESIGN
Planning
Before assembling and testing a basic audio detection system, a general strategy must be
outlined. The system to be designed in the following steps will detect sound of a certain
set of characteristics, will derive a direction of origin of the sound source, and will rotate
a camera to point in the derived direction. In a real-world setting, such a system would
activate when triggered and would automatically aim a camera towards the determined
sound source, in hopes of capturing potentially valuable video evidence to aid
investigators.
With these expectations in mind, the system should include microphones to capture
audio, a computer to process the incoming audio and send commands, a microcontroller
to receive the commands and send corresponding voltages, a servo to receive those
voltages and rotate a platform, and a camera affixed to the platform to quickly capture the
scene on video. The camera may then be wired back to the computer to display or record
the incoming video information. To keep things simple, the servo will only rotate the
camera along the horizontal x-axis, and will have a range of 180 degrees of rotation.
Along with the equipment planning, a strategy should be made for the system
programming. The two main questions to answer are: how will the system discriminate
gunshot-like sounds from other sounds? And how will the system determine the direction
of the sound source?
To discriminate gunshot-like sounds from others, the sounds of interest must be
characterized in terms of measurable traits. To the human ear, the most obvious of these
8


traits are the perceived loudness and short duration of the event. According to Michael
and Lucien Haag, a gunshot sound measured from 1 meter away often reports louder in
dB than chain saws, jackhammers, and even a jet taking off 100 feet away.11
Additionallythe rise time or time from the start of the event to the first peakis nearly
instantaneous. One study in particular found that the muzzle blast or explosive shock
wave and sound energy emanating from the weapons barreloften lasts for less than 3
milliseconds.12 This means the shape and relative intensity of a gunshots waveform or
visual representation of an audio signal or recording (used to show changes in amplitude
over time), can separate a gunshot sound from others. Though costs and timeframes limit
the materials used in this project, these audio characteristics can still be harnessed using
readily available components and intuitive programming.
9


Figure 2.1 Waveform of .22 Caliber Rifle Gunshot with Reflection
Audio courtesy of user gezortenplotz via FreeSound.org, recorded with Nady wireless
microphone to minidisc
The waveform in Figure 2.1 demonstrates the primary characteristics of a gunshot sound,
the high signal power and the near-instantaneous first peak from relative silence. This
recording in particular was purported to have taken place at an outdoor firing range.
Notice a pronounced reflection recorded very quickly after the originating event takes
placemost likely the response of the original sound event bouncing off the rearward
retaining wall or barrier used to stop incoming bullets.
Finally, the means of determining the direction of the sound source should be addressed.
In a planean objects velocity can be derived from the time elapsed over a known
distance, assuming the objects speed is constant. This is represented by v = d/t. When a
traveling wave maintains a constant speed over a known distance, the elapsed time will
be constant as well. However, when the wave begins at a third point and travels at a
constant speed along any trajectory other than perpendicular to the midpoint between two
10


microphonesthe velocity and distance can be constant and the arrival times to each
point can vary. The wave will reach the closer point first and the further point second.
Then, using the delay between the signals arriving at each channel,a source bearing can
be derived, with the source originating from a point along the bearing. These are the
working principles behind sound ranging, past and present, and are demonstrated in
Figures 2.2 through 2.5.
11


Figure 2.3 Angle Determination from Sound Delay between Two Microphones (II)
12


With sound source S, the produced sound propagates outward at a constant speed. In
Figure 2.2, because S is closer to L than R (SL < SR) the sound from S reaches L first,
then R. This also means radius a is shorter in length than radius b (ra < rb). The delay of
the arrival of the sound to each channel determines the calculated direction from which
the sound came (ang), which is assessed from the midpoint between L and R.
Like in Figure 2.2, the scenario in Figure 2.3 demonstrates how ang can be derived from
the difference between SL and SR. This time, since S is closer to R (or rb is shorter in
length than m), the corresponding angle is in the direction from the midpoint to the R-
side.
Example 2.4 shows that even extreme angles can be determined using the difference
between SL and SR, or ra and r b.
The resulting angles in all of these examples are independent of the distance from the
sound source to the microphones, since the derived angle is a bearing, not an absolute
13


point. Even though it involves the furthest distance from the microphones of all the given
examples, Figure 2.5 results in the same calculation process for delay and subsequent
angle.
Equipment and Configuration
The system begins with a pair of microphones. Microphones with high tolerance to loud
impulses would be ideal for a fully functioning system used in a real operation, but an
inexpensive pair of smaller microphones is suitable for this design. The microphones
used in this test are a pair of Olympus ME-15 microphones. These are considered a stereo
pair and both capture audio simultaneously. Next, the microphones are connected to a
laptop computer via a stereo input cable into the stereo mic-in port. This computer is
equipped with MatLab, which is a versatile computation and programming software.
MatLab handles both the audio input and the command output to the microcontroller. The
actual programming scripts used in MatLab and with the microcontroller will be
discussed later.
A microcontroller is then attached to the computer via serial interface. In this case, the
connection is via USB cable. The microcontroller of choice is the Arduino Uno, due to its
versatility and extensive open-source support. The Arduino accepts commands from the
computer and sends a corresponding voltage to a servo motor which rotates a mounted
webcam. The servo motor is a standard HS-422 servo, and the webcam is a 5-megapixel
USB webcam that is connected back to the laptop computer for display purposes.
14


Figure 3.1 Gunshot Detection and Localization System, Basic Design Overall
Figure 3.1 is a photo of the overall view of the system designed, assembled, and utilized
for the testing outlined in this project. The microphones acquire incoming sound and send
to MatLab for processing. Should the incoming signal meet the threshold requirements,
MatLab would process the signal delay and compute the angle. The Arduino receives the
angle rotation command via serial connection (the white cord on the right-hand side of
the laptop), and communicates to the servo motor with the camera mounted atop. Then,
the image information from the camera is sent to the laptop through another USB
connection for display and potential recording purposes. This workflow is outlined in
Figure 3.2 below.
15


Figure 3.2 Basic System Workflow
1)sound emanates from source, microphones capture upon arrival; 2) laptop receives
and processes incoming audio information, determines possible target signal
confirmation, delay, and resultant angle; 3) laptop sends command to Arduino;
4) Arduino sends voltage to servo motor with camera mount; 5) laptop receives resultant
image information from camera
16


Figure 3.3 Gunshot Detection and Localization System, Arduino and Servo
Close-up image of Arduino and servo assembly
Programming
With the system configured properly, the Arduino and MatLab need to be programmed
properly. The Arduino platform works fluently with MatLab, enough so that the
microcontroller can be programmed to work continuously, waiting for MatLab serial
commands, executing the commands, and returning results if asked. To prepare the
Arduino for this setup, the MATLAB Support Package for Arduino (aka ArduinoIO
Package) must be downloaded to the laptop computer. From this package, the
motorsrv.pde file must be uploaded to the Arduino IDE, and the appropriate AFMotor.h
17


and AFMotor.cpp files must be allocated properly. For instructions on how to perform
these stepsrefer to the corresponding forum at MatLabs home page.
Next, a script must be written for MatLab to automatically process the incoming audio
and send commands accordingly (Figure 3.4).
%% Automatic Audio Detection/Image Response (AADIR) system
% Mr. Jordan Graves, BS and Dr. Catalin Grigoras, PhD
% 2 012
%% Purpose:
% 1.This system will acquire live audio signals.
% 2. Based on defined thresholds, this system will discriminate
particular acquired audio events from others.
% 3. Using the perceived delay of incoming audio signals between the
pair of recording channels, this system will estimate the directional
source of the discriminated audio signal.
% 4. This system will command the servo motor to rotate the cameara
array towards the perceived source of the discriminated audio signal.
%% Materials Used:
% (1)Arduino microcontroller with serial connection to computer and
signal connection to servo motor
% (1)5V rotational servo motor (180-degree range) connected to Arduino
% (1)webcam attached to rotating mechanism of servo motor
% (2) omnidirectional microphones arranged to acquire stereo audio
signal,connected to computer via stereo microphone input, through Y-
adapter
%% Notes
% works with motorsrv and add AFMotor.cpp and AFMotor.h to path:
...Arduino\libraries\Servo
%% Script
%delete(a)
%connect to the board
a=arduino('COM13')
% define Pin#9 as output and attach the motor to it
a.pinMode(9,'output');
% Attach servo#2 to Pin#9
a.servoAttach(2);
a.servoWrite(2,90); %reset servo to center
% define the audio settings
% sampling frequency
fs=48000;
% resolution (bits)
18


nbits=16;
% no. of channels
ch=2;
% each ''extraction^ length in sec
t=0.5;
% signal power threshold
th=120
% window threshold size
win=200;
% define the audio object
recObj =audiorecorder(fs,nbits,ch) ; %begin recording
get(recObj) %collect/display values as they are recorded
disp 1***BEGINNING ACQUISITION***1 %status message
for k=l:2
% aquire the audio signal
recordblocking(recObj,t); %record without on-the-fly control until
recording is stopped
% Store data in double-precision array.
x=getaudiodata(recObj,'intl6'); %signed integers mapped to set
parameters (anything outside will be "rounded")
% find absolute value of incoming signal
xa=abs(x);
% extract L and R channels
L=double(x(:,1));
Redouble(x(:,2));
[k max(L) max(R)] %query for maximum values during sampling "window"
if max(L)>th && max(R)>th %set power threshold
if xa(k:k+win) % Plot the waveform
subplot(211),plot(L,
axis([0 length(L) -2
subplot(212),plot(R,
axis([0 length(R) -2
(grid on, tight to L/R)
1r1), grid on
15 2A15])
1g1), grid on
15 2A15])
disp 1***SYSTEM ARMED, DATA COLLECTED***1 %status message
[c,lags]=xcorr(L,R); %cross-correlation between vectors
(automatically adjusts for length differences) returns a ''lag vector^
[al,bl]=max(L); %fs/time of max values
[a2,b2]=max(R);
19


[a3,b3]=max(c); %define c's maximum values as a3,b3
delay2=fs/2-b3 %delay is half of sampling frequency minus b3
(maximum value for c), in samples
s=delay2;
if s<-127 %round values outside degree parameters to furthest
degree value left or right (to maintain 180 degree range)
s=-127;
elseif s>127
s=127;
end
% convert the delay s into degrees ang
ang=rund((s+128).*179/256)
% rotate angle ang
a.servoWrite(2,ang) ;pause(0.01);c;
end
end
end
delete(a)
Figure 3.4 Audio Detection and Image Response Script
As is standard for MatLab scripts, green lines of text with a percentage sign preceding the
content are considered notes and are not executed with the actual programming language.
The comments provide guidelines for each portion of the script.
Primary points of interest in the script include the win and th thresholds and the
delay calculation and angle conversion elements.
% signal power threshold
th=120
% window threshold size
win=200;
Figure 3.5 Thresholds in Audio Detection and Image Response Script
According to Figure 3.5, which is a reference to the threshold element of the script, the
th threshold is 1200 quantization levels of relative signal power. This is a setting
dependent on multiple factors, including microphone gain settings, expected distance
from sound source to microphones, and expected background noise. Due to these many
20


factorsthe th setting requires carefUl calibration for each deployment. Win
corresponds to a threshold of 200 samples. Note, the sampling frequency of the incoming
audio is 48kHz.
Figure 3.6 Discriminatory Thresholds for Audio Event Exclusion
Figure 3.6 illustrates how a set of thresholds might work in discriminating audio signals
by power and duration. At window /, the signal meets both the minimum power and
maximum duration thresholds. At //, the signal meets the minimum power threshold but
is too long in duration. At ///, the signal does not meet the power threshold. In the script
used in this project, the signal must meet the power threshold, then the duration
threshold, in order to provoke a system response.
The other significant portion of the script used in the project pertains to the delay
calculation and angle conversion portions (Figure 3.7).
21


[c,lags]=xcorr(L,R); %cross-correlation between vectors (automatically
adjusts for length differences) returns a ''lag vector^
[al,bl]=max(L); %fs/time of max values
[a2,b2]=max(R);
[a3,b3]=max(c) ; %define c's maximum values as a3, b3
delay2=fs/2-b3 %delay is half of sampling frequency minus b3
(maximum value for c), in samples
s=delay2;
if s<-127 %round values outside degree parameters to furthest
degree value left or right (to maintain 180 degree range)
s=-127;
elseif s>127
s=127;
end
% convert the delay s into degrees ang
ang=rund((s+128).*179/256)
Figure 3.7 Channel Delay and Angle Calculation in Detection and Response Script
The lower portion of the script determines the delay of the incoming audio event between
channels, and then produces the corresponding angle for the sample delay. The Arduino
uses 0 degrees as a valid degree integer, so the 180-degree range actually includes 0 and
spans from 0-179.
Some further explanation is necessary for the delay and angle calculation portions of the
script. The xcorr, max(L), and max(R) portions of the script mark the initial peak values
of the incoming signal in each sampling windowdefined earlier in / as .5 seconds in
length). Each initial peak is marked in the numerical sample it was measured to take
place. The delay is then determined from the difference in those sample values; if the
value fell outside the allotted range, it would be rounded to the high or low extreme,
depending on whether it was above or below those extremes, respectively.
Assuming the speed of sound is approximately 350 meters per second, a sound wave
would travel 6 feet (or approximately 1.829 meters) in around .0053 seconds. 6 feet is the
22


prescribed distance between microphones used in the system, .0053 seconds is the
maximum delay between channels. Since the sampling frequency defined above is
48kHz, or 48000 samples per second, the maximum delay between channels can also be
measured as approximately 256 samples. The delay is then added to 128 to account for
the delay reference to the R channel instead of L, and then compared to the ratio of
samples to angles. The resultant value is rounded to the nearest whole integer and is the
calculated angle for rotation.
So, if the delay were 98 samples (signal reaching R 98 samples faster than L) \
Figure 3.8 Delay and Angle Calculation Example 1
ang = (98+128)*179/256
ang = (226)*179/256
ang =158 degrees
23


On the other hand, a delay of -110 samples (signal reaching Z 110 samples faster than R):
Figure 3.9 Delay and Angle Calculation Example 2
ang = (-110+128)*179/256
ang = (18)*179/256
ang =13 degrees
The angle calculated from this script is measured from left to right of the midpoint
between microphones. This means angles between 0 and 88 degrees will rotate the
camera counterclockwise from neutral, while angles between 90 and 179 degrees will
rotate the camera clockwise from neutral. Of course, the user may reverse which side, L
or R, to assign reference to, among many other customizable features (thresholds,
microphone distance calibration, etc).
Testing and Results
To evaluate the system5 s functionality, a simple test was formulated and executed. The
system was assembled as described above, in a series of open, outdoor tennis courts. This
24


place was chosen with the intent of minimizing potential interfering reverberations, as
well as other variables introduced in more crowded areas. The test was executed at night,
to reduce the chance of external noise interference from wind or passerby. The air
temperature was approximately 37 degrees Fahrenheit. This is significant because,
although relatively small changes in temperature would not affect the speed of sound in a
drastic way, it is well known that larger temperature variations could introduce
complications in calculating the speed of sound.13
As mentioned earlier, the microphones used were elevated and spaced at approximately
72 inches apart, with the notion that spacing should be towards the wider end of the
spectrum to emphasize the delay between incoming audio channels.
Markers were placed at 5,10, and 15-foot distances from the center point between
microphonesall distances at 103050and 70-degree angles from that same center
point in either direction. In total,24 markers were made. These markers indicated the
intended positions from where the test sound would originate.
At the time of the test, an actual firearm was not an available sound source. Instead, a
loud, sharp clap of the hands was utilized at each marker. The overall waveform shape of
a hand clap could properly simulate a gunshot because both events can be characterized
with high intensity and short duration. Though the claps were kept at consistent volume,
some variation in signal intensity must be acknowledged. However, the variations were
considered acceptable because of the multiple factors that introduce variations in sound in
a real-life situation. The test itself was designed to be controlled in most reasonable
aspects, yet allowed for some semblance of a realistic environment.
25


After the system was assembled ana initiated, the testing began. After each instance of a
hand clap at each marker, the system was reviewed for a response and possible camera
movement. At each marker, the possible responses for the system were:
1. Rotation of the camera towards the sound source, stopping with the
marker in the center of the camera frame (represented by | in the tables
below)
2. Rotation of the camera towards the sound source, stopping with the
marker in the frame but not in the center (represented by O in the tables
below)
3. Rotation of the camera, stopping without the sound source in frame
(represented by X in the tables below)
4. No camera movement in the response (represented by in the tables
below)
The responses were determined after some camera movement in response to the claps, or
after a maximum 5 clap attempts at the marker.
Each marker was tested in a trial, with three total trials making up the test. The order for
marker tests varied by trial; the first two trials were in order of each degree at one
distance followed by the remaining two distances, while the last trial proceeded in a more
staggered pattern. Tables 1.1 through 1.3 illustrate each trial and set of results.
26


Table 1.1 Test Trial1 Configuration and Results
Trial1-Results 5 feet 10 feet 15 feet
70 Stage Left (SL) 4 4 4
50 SL - - 1
30 SL - - 0
10 SL 0 X 0
10 Stage Right (SR) 0 X 0
30 SR X X X
50 SR - - 0
70 SR 4 - 0
Table 1.2 Test Trial 2 Configuration and Results
Trial 2 Results 5 feet 10 feet 15 feet
70 Stage Right (SR) i i i
50 SR - - 0
30 SR X X X
10 SR 0 0 -
10 Stage Left (SL) 1 0 0
30 SL X X X
50 SL X - i
70 SL - - i
Table 1.3 Test Trial 3 Configuration and Results
Trial 3 Results 5 feet 10 feet 15 feet
70 Stage Right (SR) i i i
10 Stage Left (SL) 0 i 0
50 SR 0 X X
30 SL i X X
30 SR i X X
50 SL i 0 0
10 SR 0 0 0
70 SL 1 1 1
Each test proceeded through the first column, then the middle column, then the last
column, each column from top to bottom ('Stage5 directions refer to the direction from
the viewpoint of the camera outward towards the markers).
27


Discussion
At first glance, the results from the test as a whole seem mixed and inconsistent, with
only Trial 3 producing responses to each marker with camera movement. To make the
results more relevant and truly evaluate the test results for inconsistent system responses
each trial was evaluated in terms of how many Potentially Useful Evidences (or
PUEs) were created. These are defined within the realm of the test as camera responses
that end with the location of the sound source somewhere in frame, either centered or not
centered. This is to simulate a real-world scenario, where a video recording of an incident
would be submitted for evidence. Potentially useful evidence in such a scenario would
require the event itself or the immediate aftermath to be captured somewhere in frame,
either centered or not. Within those parameters, the test resulted with Table 2.0.
Table 2.0 Test Trial Potentially Useful Evidence Results
PUE Results 5 feet 10 feet 15 feet
Trial1 4 of possible 8 1 of possible 8 7 of possible 8
Trial 2 3 of possible 8 3 of possible 8 5 of possible 8
Trial 3 8 of possible 8 5 of possible 8 5 of possible 8
The 8 possible instances for PUE response refer to each marker at the given distance.
This table of PUE results helps clarify exactly what about the system5 s responses are
inconsistent. The inconsistency does not necessarily exist on a trial-by-trial basis, but
more within the changes between distances. While the potential for useful evidence
increases in Trial 3 from the others, Trial1 and 2 exhibit increases in PUEs as the
distance grows. The opposite results occur in Trial 3, where the PUE decreases and stays
the same at each respective distance. This is a counterintuitive result, since one would
assume that events at further distances are more likely to take place within the frame of
view of a camera.
28


The overall inconsistency in the system can be attributed to multiple factors. First of all,
the microphones used may not be most suitable for the task of discrimination. Due to the
very short duration of most gunshot sounds, microphones used should be sensitive
enough to accurately define the incoming audio information, to the point in which a
series of loud impulses spaced closely together in rapid succession would be recorded as
such, instead of one long, loud impulse. The microphones used in this test were not
specifically designed with that task in mind.
Second, the surrounding environment plays a large role in the effectiveness of these
systems. Even though the testing location and time of day were chosen with the intent
that uncontrollable variables would be minimized, not everything could possibly be
accounted forand minor changes in the test conditions could introduce fluctuations in
data.
Finally, the programming language used in MatLab itself could use some review and
potential upgrading. In order to derive the correct thresholds at a particular settingtesting
must be done to determine a combination that works best. A level of automation for
thresholds might be worthy of some attention, where the system could be designed to
automatically adjust for changes in the noise floor, etc. Real-time adaptive filters might
also work in terms of limiting the amount of extraneous and useless sound information
that would only hinder progress, especially sounds of frequencies below around 400Hz
and above around 2.5kHz, which are the primary frequencies exhibited by gunfire.14
These instructionstestsand results are usefUl in describing the process for basic
detection and response, but the design does carry an underlying flaw. Any impulse that is
short and loud enough would trigger a camera movement, not necessarily a gunshot. This
29


is because further discrimination of gunshot sounds from other noises of the same shape
involves higher-level filtering and analysis of the sounds, via the programming.
Unfortunately, gunshot audio usually exhibits peaks around 630Hz.15 These ranges are
common for a wide variety of other sounds, making discrimination much more
complicated.
1000Hz 2000Hz 3000Hz 4000Hz SOOOHz 6000Hz 7000Hz 8000Hz 9000Hz 11000Hz 13000Hz 15000Hz 17000Hz 19000Hz 21000Hz
Algorithm [ Spectrum_^ Size: [ 2048 __y [ Export...] Rep j
Function: [ Hamming window_cj Axis: [ Linear frequency y Close _] Grids
Figure 3.10 FFT of .22 Caliber Rifle Gunshot with Reflection
Using Fast Fourier Transform to bring a signal into the frequency domain, Figure 3.10
allows for some frequency analysis of the example recording shown earlier as a
waveform in Figure 2.1. The frequencies of Highest intensity range from the lowest up to
2kHz, with maximum values below 500Hz. These results do incorporate some
environmental acoustics due to the reflection in the recordingbut that would be
expected for most audio events in realistic scenarios.
To properly discriminate from other sounds, the best approach at this point is an
algorithmic learning strategy, such as those proposed by Morton and Collins16 or
30


Valenzise et al.17 Algorithmic learning strategies are not used in this project, but will be
discussed later.
To properly simulate the products used in the field and attempt to replicate their
functionality, vendors such as ShotSpotter should provide scientists with some data in
regards to their product specifications and schematics, simply for research purposes. This
conflicts with the vendors rights to withhold proprietary information, but more
importantly encourages unbiased review and testing in a scientific and peer-reviewed
forum. Those kinds of procedures would help alleviate doubts about these systems
capabilities, and may improve their general public image, as observed below. At the time
of this test, representatives from both ShotSpotter and Boomerang elected not to field
questions about specific elements of their respective systems5 designs, functionalities,
and test data.
31


CHAPTER IV
FUTURE POSSIBILITIES
The Current State of Gunshot Detection and Localization Systems
Gunshot detection and localization systems, ShotSpotter in particular, have been used to
aid law enforcement for some time now, and have been met with mixed reactions. In
some situations, law enforcement agencies report dramatic improvements in gunfire
incident response time. In others however, the system is reported as inaccurate, costly,
and overly demanding of valuable resources, including taxpayer money.
In 2009, the New York Times followed ShotSpotter use in the greater New York, New
Jersey, and Connecticut areas. Officials praised the system for helping reduce shooting
response time, which led to faster aid for victims and better suspect apprehension.
According to one Sergeant, the system proves especially helpful in areas where citizens
are so accustomed to gunfire that 9-1-1 calls simply do not take place any longer.18
In 2011, city officials of Wilmington, North Carolina approved a 2-year contract with
ShotSpotter, with the mindset that personnel is the greatest expense, and the technology
could be used to better manage staff in the field. The contract expires in mid-December
of 2012, and will be reevaluated by the city to determine whether or not to extend the
19
service.
Other agencies have seen less success with ShotSpotter. In March of 2012, the New
Haven Police Department adopted a new protocol of sending audio data to ShotSpotter
headquarters prior to redirection to the New Haven 9-1-1 center. Though the change in
routing only delays information from reaching dispatch by mere seconds, the system has
reported enough false positives to cause concern and necessitate the change.20 More
32


worrisome are the gunshot events that take place but are not detected, which has
happened in the area.
That very situation factored into a decision made by city council members of Trenton,
New Jersey to reject a ShotSpotter expansion in January 2012. The previous Christmas, a
man was fatally shot and left to die on a sidewalk of a major avenue. Although a
ShotSpotter sensor was set up only blocks away, no alarms were triggered. The South
Ward Councilman, a former police officer in Trenton, concluded that u[ShotSpotter] does
not workat least not for Trenton.21
To further cloud judgment on system value, the city of Detroit rejected ShotSpotter
installation because it was too taxing on available personnel. The Broward County
Sheriff in Florida previously used ShotSpotter but decided to remove it, citing the system
was not cost effective and came with too many false alarms. However, the Rochester
Police Department in New York swears by ShotSpotterasserting ...its value is always a
relative question. We think its valuable or we wouldnt have done it. Furthera
Criminal Justice professor and 26-year RPD veteran asksIf it gets police to one more
victim sooner, how do you put a price on that? If it adds to the evidence to convict
someonehow would you add value to that?22
To better decide whether or not to employ a gunshot detection and localization system
like ShotSpotter, authorities need reliable, unbiased information in the form of extensive
testing and reporting. While efficacy studies and evaluations are available, most do not
publish complete statistical results.2j r Some studies are conducted with tangible and
measurable results, but are not current within the past decade.25 A long-term study with
complete transparency of test materials and results will be the best means of assessing the
33


effectiveness of gunshot detection and localization systems, determining strengths and
weaknesses, and choosing whether to utilize the system or not. Until the day comes when
leading manufacturers reveal their design plans and schematics, researchers must
continue to build replicas as similar as possible to those used in the field, so results
gathered will be relevant and useful.
Prospects for the Basic System
Within the scope of this project, the first future developments available should be
improvements and further testing. As mentioned earlier, additional testing using superior
equipment such as high-quality microphones, cameras, etc. would logically introduce
different results than those found with the current configuration. In particular,
microphones more sensitive to changing dynamics in audio signals should theoretically
offer more precise measurements and findings. Not only could additional upgrades offer
improvements on current functions, new additions to the current microphone arrangement
could literally add another dimension to the system response capabilities. A third
microphone could allow for three-dimensional localization, either in the form of
triangulation in a single plane or in space.
In much the same way that two-microphone systems would give a direction in a single
plane, three microphones would accomplish the same featboth vertically and
horizontally. If the corresponding microphones are arranged as described in Figure 4.1,
such a system could theoretically provide azimuth and elevation information about a
signal source.
34


Figure 4.1 Source Determination from Three Microphones in Three Dimensions
With two microphones, such as in LR earlier, two-dimensional hemispherical sensing is
feasible. Three microphones could allow for three-dimensional sensing, along an X, Y,
and Z axis accordingly. Each resulting angle produced from each pair of microphones
would converge on a point in space and thereby provide a location for the source of
sound. This is shown in Figure 4.1.
Triangulation, often used outdoors by navigators or cartographers with compasses,
involves the use of geometry and known relationships to estimate an objects location
within two-dimensional space. By spreading the three microphones in such a way that the
sound event is contained within the triangle formed by the sensorsthe resulting angles
derived from each pair would localize and pinpoint the perceived source to a specific
location. Technically, two angles could produce a location for the sound source.
However, a third angle would not only be available by default after using three different
35


microphones, the third angle would provide additional correction to minimize error in the
results. Triangulation in this context is described in Figure 4.2.
Using the a-b, b-c, and a-c differences and resultant angles for the XY, YZ, and XZ pairs
respectively, the relative location for the sound source can be calculated. Therefore,
assuming each sensor is fixed in a known coordinate, the tme coordinate of the sound
source can be derived from its relationship to the sensors. This particular example shows
how the source can be derived in a plane.
Additional microphones would not only increase the capabilities of the system, but would
minimize potential error while maximizing sensing capabilities. Sound events, such as
muzzle blasts emanating from the firearm may be highly directional and vary
significantly at different angles.26 The greater the microphone coverage available, the
greater the potential for capturing an adequate amount of audio information.
Once these upgrades have been made, further testing would undoubtedly take place. To
better estimate accuracy, the camera may be replaced with a laser pointer, which will
36


provide a more quantitative means of evaluating system response. The laser would
produce a beamterminating in a point. The distance of this point from the originating
sound source could offer measurable results for analysis and future calibrations.
On the topic of calibrations, the programming language could also be adjusted to
accommodate for different microphone spacing. This would be used to determine a
minimum or perhaps even an optimum distance between microphones, which would
become more important as more microphones are added and the system complexity
grows.
Since this system in particular was adjusted to react to gunshot-like audio events, it
would not be recommended for other applications as-is. However, the signal power and
window thresholds allow for a variety of potential target sounds. For instance
researchers might consider using a system described above to monitor animal activity in
their natural habitats. This could even entail the building and installation of systems in
extreme environments, such as seafloors or mountaintops, so scientists may carry out
their studies from the safety of a remote laboratory. Research using these sorts of systems
in national parks has already commenced, in hopes of combating outbreaks of poaching
on protected territory.27
A system such as this might also be used by an instructor teaching an on-line courseso
that he or she might feel free to move about the setting without worrying about whether
the camera will remain fixated upon him or her while talking. While the teacher shifts to
draw attention to a demonstrative object at his or her side, the detection system, attuned
to his or her voice, would compensate accordingly with a camera adjustment.
37


Of course, such possibilities would depend heavily on far superior programming to prove
successful in a long-term scenario. These kinds of uses would almost certainly necessitate
the use of learning algorithms to allow the system to more accurately suggest whether to
accept or reject incoming signals, perhaps to the point where it may make the decision for
itself and allow the user to review the decisions made and correct any errors.
Though some might suspect a complicated set of equations and commands must steer a
discriminatory algorithm, one study in particular28 actually listed a series of relatively
simple correlations as one of the more robust methods tested. Templates of gunshot
signals taken from 30 metros (Spanish for meters in English) and 90 metros are
averaged and stored in vectors 1000 samples in length. At 39 samples per iteration, a
correlation between the incoming signal and averaged templates is calculated for each
iteration and stored in a pair of vectors for comparison against given thresholds. Testing
revealed a Tme Positive Rate (true positives detected divided by total number of
positives) of .91, with a False Positive Rate (false positives detected divided by total
number of negatives) of 0.0. Using this sort of system would require little in terms of
additional programming skill or hardware resources, and could be easily customizable for
a users need.
Similar to the time-based correlation method is another adapted from a document
released by ShotSpotter to inform about gunshot location systems.29 With sensitive
equipment and comprehensive programming, an accurate frequency envelope can be
developed for the gunshot event. This envelope could be stored as a series of data points,
along with a large volume of multiple series of data points generated in similar fashion,
all normalized to the same length. Storing such data would require much less in terms of
38


relative capacity when compared to storing numerous uncompressed recordings. An
incoming signal could be broken down into the same sort of envelope, and then compared
at target frequencies to the stored templates for correlation values. If the incoming
signal shared a high enough correlation in frequency to known templates, an automated
decision could be made to alert the system and direct a response. On the other hand, data
points could be stored for false positives (heavy machinery, backfiring automobiles, and
the like), so that if the incoming signal shared high enough correlation with false
positives, the system could notify the operator of a potential false alarm needing review
and confirmation. As each event takes place, the incoming signal values could be
designated as true or false positiveand then the system would subsequently learn. This
sort of strategy would function best using highly sensitive microphones and large
databases for comparison, which could restrict processing speeds and inflate costs.
While some might question the effectiveness of this type of programming against false
positives as deceptive as other small explosions like fireworks, the opposing viewpoint is
that a system prone to more false positives is much more acceptable than one prone to
error in false negatives. Nevertheless, researchers have shown that even firecrackers
differ from gunshots in terms of frequency domain bandwidth,30 meaning this proposal in
particular requires further development but shows some promise.
39


CHAPTER V
CLOSING REMARKS
Gunshot detection and localization systems have attracted attention of nearly every sort
the past few years. While some praise the systems5 implementation in law enforcement
scenarios, others remain skeptical of their effectiveness versus costs. At this time, it is
unclear whether or not gunshot detection and localization systems deserve widespread
installation or banishment, but the most reasonable course of action is to continue testing
and development, with hopes of constant improvement and simultaneous research
transparency. Fortunately, basic designs such as the one outlined in this thesis, combined
with the enormous potential for their use in gunshot response or other related
applications, allow for progressive scientific development and advancements. This in turn
may give rise to a universally reliable system to use in law enforcement. No matter the
means, the bottom line is this: authorities and law enforcement personnel must continue
to pursue a healthy partnership with technology to combat the ever-changing threat of
crime and violence. In the educated words of Criminal Justice professor Dennis Kenney,
Guns are more ubiquitous than they used to be. Theres inevitably going to be more
gunplay on the streetsand its inevitable that police will want to begin to use technology
to help address that.31
40


REFERENCES
1. Van der Kloot W. Lawrence Bragg5 s role in the development of sound-ranging in
World War I. Notes and Records of the Royal Society. 2005;(59):273-284.
2. Melson CD. Organization and Equipment for the Defense Battalion. In: Condition
Red: Marine Defense Battalions in World War II United States Marine Corps-
Editing and Design Section, History and Museums Division. 1996.
3. Locating Surveillance and Target Acquisition Association. Locating Artillery
Overview.
http://www.locatingartillery.org/overview.htm
4. Varshney L. Ground Surveillance Radars and Military Intelligence. North
Syracuse (NY): Syracuse Research Corporation; 2002 Dec.
5. Lahr JC, Ward PL, Stauffer PH, Hendley JW. Earthquake Technology Fights
Crime. United States Geological Survey-Earthquake Hazard Reduction Program.
http://jclahr.com/science/psn/gunshots/factsheet/index.html
6. ShotSpotter, History.
http: // shot spotter. com/ comp any/hi story
7. Hsu J. Tracing a Bullet Back to Its Gun. ScienceLine.org. 2007 Jan.
http://scienceline.org/2007/01/tech-hsu-guns/
8. Defense Industry Daily. Sniping at US Forces Beginning to Boomerang.
DefenseIndustryDaily.com. 2011 Feb.
http://www.defenseindustrydaily.com/sniping-at-us-forces-beginning-to-
boomerang-01128/
9. Raytheon BBN Technologies. Boomerang Warrior-X. 2010 Jul.
http://bbn.com/resources/pdf/Boomerang-Warrior-X-072210.pdf
10. Al-Muslim A. Gunfire detection coming to Hempstead Village. Newsday.com.
2012 Jul.
http://www.newsday.com/long-island/nassau/gunfire-detection-coming-to-
hempstead-village-1.3828324
11. Haag MGHaag LC. Shooting Incident Reconstruction. 2nd rev. ed. San Diego:
Academic Press, 2011.
12. Maher RC. Acoustical Characterization of Gunshots. Institute of Electronics and
Electrical Engineers. 2007 Apr. 109-113.
41


13. Dean EA. Atmospheric Effects on the Speed of Sound. El Paso (TX): United
States Army Electronics Research and Development Command; 1979. Report
No.: ASL-CR-79-0100-4.
14. Haag MG, Haag LC.
15.Ibid.
16. Morton KD, Collins L. Bayesian Detection of Acoustic Muzzle Blasts. Sensors
and Command, Control, Communications, and Intelligence Technologies for
Homeland Security and Homeland Defense VIII (proceedings of SPIE). 2009;
7305.
17. Valenzise G, Gerosa L, Tagliasacchi M, Antonacci F, Sarti A. Scream and
Gunshot Detection and Localization for Audio-Surveillance Systems. Institute of
Electronics and Electrical Engineers. 2007.
18. Buckley C. High Tech Ears Listen for Shots. NYTimes.com. 2009 Nov.
http://www.nytimes.com/2009/ll/22/nyregion/22shot.html?pagewanted=l&r=2
&emc=etal
19. Feskos B. Wilmington police hope gunfire sensor will help response times.
StarNewsOnline.com. 2011 Jun.
http://www.starnewsonline.com/article/20110609/ARTICLES/110609693?p=l&t
c=pg
20. Kaempffer W. New Havens ShotSpotter reliability in question. NHRegister.com.
2012 Mar.
http://www.nhregister.com/articles/2012/03/12/news/new_haven/doc4f5d6e0426b
cfl 01484344.txt?viewmode=fullstory
21. Fair M. Trenton council rejects expansion of'ShotSpotter5 gunshot detection
system. NJ.com. 2012 Jan.
http://www.nj.com/mercer/index.ssf/2012/01/trentoncouncilrejectsexpans.htm
1 _ _
22. Hong A. RPD: ShotSpotter Valuable to Crime Solving. 13WHAM.com. 2012
Jun.
http://www.13wham.com/news/local/story/rochester-shotspotter-police-alerts-
sensors/uSDQnaKgw0i-CRyzn4HQ_w.cspx
23. Selby N, Henderson D, Tayyabkahn T. ShotSpotter Gunshot Location System
Efficacy Study. CSG Analysis, endorsed by the National Organization of Black
Law Enforcement Executives. 2011.
42


24. Platt J. ShotSpotter Stat Update. TheHuntingtonian.com. 2012 May.
http://thehuntingtonian.com/2012/05/25/shotspotter-stat-update/
25. Mazerolle LG, Watkins C, Rogan D, Frank J. Random Gunfire Problems and
Gunshot Detection Systems. Washington (DC): National Institute of Justice,
United States Department of Justice; 1999 Dec. NCJNo.: 179274.
26. Beck SD, Nakasone H, Marr KW. An Introduction to Forensic Gunshot
Acoustics. Proceedings of the 162nd Meeting of the Acoustical Society of
America; 2011 Oct 31-Nov 4; San Diego (CA).
27. Gonzalez-Castano FJ, Alonso JV, Costa-Montenegro E, Lopez-Matencio P,
Vicente-Carrasco F, Parrado-Garcia F, et al.Acoustic Sensor Planning for
Gunshot Location in National Parks: A Pareto Front Approach. Sensors 2009
Nov;(9):9493-9512.
28. Chacon-Rodriguez A, Julian P. Evaluation of Gunshot Detection Algorithms.
Proceedings of the Argentine School of Micro-Nanoelectronics, Technology and
Applications; 2008:49-54.
29. Showen RL, Calhoun RB, Chu WC, Dunham JW. Gunshot Location in Complex
Environments-Concepts and Results. In: Carapezza, EM, editor. Acoustic
Sensors, and Command, Control, Communications, and Intelligence Technologies
for Homeland Security and Homeland Defense VII (proceedings of SPIE). 2008;
6943.
30. Ibid.
31. Buckley C.
43


Full Text

PAGE 1

AUDIO GUNSHOT DETECTION AND LOCALIZATION SYSTEMS: HISTORY, BASIC DESIGN, AND FUTURE POSSIBILITIES by JORDAN R. GRAVES B.S., Northern Michigan University, 2010 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Master of Science Recording Arts, with emphasis in Media Forensics Program 2012

PAGE 2

ii This thesis for the Master of Science degree by Jordan R. Graves has been approved for the Recording Arts, with emphasis in Media Forensics Program by Catalin Grigoras, Chair and Advisor Jeff Smith Lorne Bregitzer Charles Ganzert Date: 11/16/2012

PAGE 3

iii Graves, Jordan, R (MS, Recording Arts, with emphasis in Media Forensics) Audio Gunshot Detection and Localization Systems: History, Basic Design, and Future Possibilities Thesis directed by Dr. Catalin Grigoras. ABSTRACT For decades, law enforcement organizations have increasingly utilized audio detection and localization systems to identify potential gunshot incidents and to respond accordingly. These systems have grown from simple microphone configurations used to estimate location into complex arra ys that seem to pinpoint gunfire to within mere feet of its actual occurrence. Such technology comes from a long and dynamic history of developing equipment dating back to the First World War. Additionally, though basic designs require little in terms of p rogramming or engineering experience, the mere presence of this tool invokes a firestorm of debate amongst economists law enforcement groups and the general public, which leads to questions about future possibilities for its use. The following pages will retell the history of these systems from theoretical conception to current capabilities. This work will also dissect these systems to reveal fundamental elements of their inner workings, in order to build a basic demonstrative system. Finally, this w ork will discuss some legal and moral points of dissension, and will explore these now and in the future, in additional applications as well. The form and content of this abstract are approve d. I recommend its publication Approved: Catalin Grigroas

PAGE 4

iv DEDICATION I dedicate my work to my parents and my brother, Rick, Beth, and Connor, because their support and love have directly translated to my success. They taught me how to be kind, focused, and passionate with all that I do, and I cannot thank them enough for that. I would also like to dedicate my work to my friends and family, who are one in the same. Knowing that what I do may impact all the wonderful people in my life keeps me ever steady and ever motivated to accept o nly the best of myself. Finally, I would like to dedicate my work to my soulmate, Wendy. She is the driver of my ambitions and pushes me to achieve my dreams, both of which I gratefully do with her by my side. Nothing about me is truly perfect, but the people in my life make living it seem pretty spectacular.

PAGE 5

v ACKNOWLEDGMENTS I would like to thank my classmates in the MSRA MF program, my professors and teachers at UCD and at NMU, the faculty and staff of the NCMF, my coworkers, supervisors, and mentors at USACIL, and the countless others who have offered guidance for this project. I could not have even begun this on my own, and I appreciate every contribution you all have made along the way.

PAGE 6

vi TABLE OF CONTENTS CHAPTER I. INTRODUCTION ................................ ................................ ................................ ........ 1 II. HISTORY ................................ ................................ ................................ ................... 3 III. BASIC DESIGN ................................ ................................ ................................ ........ 8 Planning ................................ ................................ ................................ ....................... 8 Equipment and Configuration ................................ ................................ .................... 14 Programming ................................ ................................ ................................ ............. 17 Testing and Results ................................ ................................ ................................ .... 24 Discussion ................................ ................................ ................................ .................. 28 IV. FUTURE POSSIBILITIES ................................ ................................ ...................... 32 The Current State of Gunshot Detection and Localization Systems ............................. 32 Prospects for the Basic System ................................ ................................ ................... 34 V. CLOSING REMARKS ................................ ................................ ............................. 40 REFERENCES ................................ ................................ ................................ .............. 41

PAGE 7

vii LIST OF TABLES Table 1.1 Test Trial 1 Configuration and Results ................................ ................................ ..... 27 1.2 Test Trial 2 Configuration and Results ................................ ................................ ..... 27 1.3 Test Trial 3 Configuration and Results ................................ ................................ ..... 2 7 2.0 Test Trial Potentially Useful Evidence Results ................................ ......................... 28

PAGE 8

viii LIST OF FIGURES Figure 1.1 Sound Ranging Diagram ................................ ................................ ............................ 4 .................. 6 2.1 Waveform of .22 Calib er Rifle Gunshot with Reflection ................................ .......... 10 2.2 Angle Determination from Sound Delay between Two Microphones ....................... 11 2.3 Angle Determination from Sound Delay between Two Microphones (II) ................. 12 2.4 Wide Angle Determination from Sound Delay between Two Microphones .............. 12 2.5 Distanced Angle Deter mination from Two Microphone Sound Delay ...................... 13 3.1 Gunshot Detection and Localization System, Basic Design Overall ....................... 15 3.2 Basic System Workflow ................................ ................................ ........................... 16 3.3 Gunshot Detection and Localization System, Arduino and Servo ............................. 17 3.4 Audio Detection and Image Response Script ................................ ............................ 20 3 .5 Thresholds in Audio Detection and Image Response Script ................................ ...... 20 3.6 Discriminatory Thresholds for A udio Event Exclusion ................................ ............. 21 3.7 Channel Delay and Angle Calculation in Detection and Response Script ................. 22 3.8 Delay and Angle Calculation Example 1 ................................ ................................ .. 23 3.9 Delay and Angle Calculation Example 2 ................................ ................................ .. 24 3.10 FFT of .22 Caliber Rifle Gunshot with Reflection ................................ .................. 30 4.1 Source Determination from Three Micro phones in Three Dimensions ..................... 35 4.2 Source Determination from Three Microphones in Two Dimensions ....................... 36

PAGE 9

1 CHAPTER I INTRODUCTION Until now, no known measure exists outside of science fiction that can successfully predict and prevent crime before it takes place. Therefore, law enforcement agencies must maximize the efficiency and accuracy of the response effort to criminal activity. The sooner law enforcement personnel can have a presence at the scene where a crime took place, the sooner the area may return to an orderly state of safety. Agents can defuse dangerous situations, gather evidence, and build a case used to apprehend t hose responsible for the crime. Many factors limit the amount of influence law enforcement agencies have on the general public, and some o f these factors introduce obstacles when striving for timely and effective crime response. Budgets and available funding dictate the amount of staff and equipment an agency can use, and legislation regulates agency power. Law enforcement groups may interve ne in a given situation, but only when logistically capable and legally permitted. While positive intended results of this control include safe and cost effective law enforcement, unfortunately negative side effects also may arise. Agencies may be understa ffed or ill equipped, and may be restricted from responding as quickly or as soundly as desired. It is then vital for law enforcement groups and the people they protect to reach a compromise -keep the enforcers of the law within clear legal and financial b oundaries while providing them with enough tools to help maintain safe communities. With that compromise in mind, law enforcement groups have begun utilizing the gunshot detection and localization system. This technology provides information for law

PAGE 10

2 enforc ers in two regards: it identifies possible gunshot events based on audio information acquired by microphones and interpreted by algorithmic processing, and it also provides the perceived location of the sound source. The system is semi automatic, which is to say it operates largely by automated computer programming but still requires human audio characteristics and alert operators of potential detected gunshot even ts, but the decision to include or exclude an audio event as a gunshot (and requiring response at the scene) still belo ngs to a human at the controls. The following sections include a history of audio gunshot detection systems, a simple design plan for a b asic system, and a discussion of the potential problems facing the implementation of these systems, with some speculation on their future use in law enforcement and for other applications.

PAGE 11

3 CHAPTER II HISTORY The origins of many technological advances are often traced back to innovations in different fields, later made applicable through simple redesign. Modern day gunshot detecti on systems share similar roots. provides information regarding the coordinates of artillery weaponry. Developed by William Lawrence Bragg, a British military officer and physicist initial sound ranging te chniques involved arrays of m icrophones carefully placed in the field of battle to detect sound events from the fired weapons and report back to a monitor at an operating base as d epicted in Figure 1.1 At times the resulting information contained valuable clues about was less than desirable in the early years, nations from each of the opposing sides made adjustments to the process to find increasingly useful results. 1

PAGE 12

4 Figure 1 .1 Sound Ranging Diagram By World War II, most major military players used sound ranging for mortar detection and counter artillery measures. In particular, British forces and United States Marines made good use of sound r anging in defensive operations. 2 Although sound ranging equipment had been growing more sophisticated and less costly over the years, radar systems and aerial surveillance took over as the primarily used gun locating methods in military operations. 3 Radar operators were capable of locati ng large weaponry faster, derived more conclusive data in settings of extreme terrain or overgrown vegetation, equipment could be outfitted on more mobile units for determining the location of airplanes and vehicles as well and most importantly, radar cou ld operate without waiting for shots fired 4 Sound ranging still held a place in combat, but acted mostly as a backup to rapidl y expanding radar capabilities. Techniques involving sound ranging for gunfire location receded in popularity until as recently a s the 1980s and 1990s. Researchers borrowed sound ranging techniques from seismologists studying earthquakes and began testing capabilities of detecting small arms

PAGE 13

5 activity in urban ar eas. 5 Organizations such as ShotSpotter Incorporated, now SST Inc., test ed detection and localization systems in areas with high crime rates, and US police departments along the Californian Pacific coast began working with the technology to improve incident response time and subsequ ently help deter future crimes. 6 Meanwhile, t he military returned to using gunfire detection and location in combat zones, mainly to assist in evading and countering enemy sniper attacks. 7 Technology is now mountable both to vehicles and personnel, and war fighting units currently rely on these tools in the Middle Eas t and other theaters worldwide. 8 9 Figure 1.2 consists of a Boomerang system outfitted to a US Army Humvee.

PAGE 14

6 Figure 1. 2 Boomerang Outfitted to US Army Humvee Photo courtesy Marine Corps Warfighting Lab via Office of Naval Research Back in the United States, agencies nationwide have deployed gunshot detection and localization systems in cities and other urban areas that are prone to gunfire related crimes and random gunfire incidents These systems are recei ving more consideration as a large contributor to community safety and law enforcement success, and offer not only increased response capabilities but potential video evidence as well, incorporating video

PAGE 15

7 capture components in the system designs. 10 While s ome critics raise concerns including costs, privacy issues, and accuracy, gunshot detection and localization systems used in American cities have had a significant impact in the way authorities identify and respond to criminal activity.

PAGE 16

8 CHAPTER III BASIC DESIGN Planning Before assembling and testing a basic audio detection system, a general strategy must be outlined. The system to be designed in the following steps will detect sound of a certain set of characteristics, will derive a direction of origin of the sound source, and will rotate a camera to point in the derived direction. In a real world setting, such a system would sound source, in hopes of capturing p otentially valuable video evidence to aid investigators. With these expectations in mind, the system should include microphones to capture audio, a computer to process the incoming audio and send commands, a microcontroller to receive the commands and send corresponding voltages, a servo to receive those voltages and rotate a platform, and a camera affixed to the platform to quickly capture the scene on video. The camera may then be wired back to the computer to display or record the incoming video informat ion. To keep things simple, the servo will only rotate the camera along the horizontal x axis, and will have a range of 180 degrees of rotation. Along with the equipment planning, a strategy should be made for the system programming. The two main questions to answer are: how will the system discriminate gunshot like sounds from other sounds? A nd how will the system determine the direction of the sound source? To discriminate gunshot like sounds from others, the sounds of interest must be characterized in te rms of measurable traits. To the human ear, the most obvious of these

PAGE 17

9 traits are the perceived loudness and short duration of the event. According to Michael and Lucien Haag, a gunshot sound measured from 1 meter away often reports louder in dB than chain saws, jackhammers, and even a jet taking off 100 feet away. 11 instantaneous. One study wave and milliseconds. 12 r visual representation of an audio signal or recording (used to show changes in amplitude over time) can separate a gunshot sound from others. Though costs and timeframes limit the materials used in this project, these audio characteristics can still be harnessed using readily available compon ents and intuitive programming.

PAGE 18

10 Figure 2.1 Waveform of 22 Caliber Rifle Gunshot with Reflection Audio courtesy of user gezortenplotz via FreeSound.org, recorded with Nady wireless microphone to minidisc The waveform in Figure 2.1 demonstrates the primary characteristics of a gunshot sound, the hi gh signal power and the near instantaneous first peak from relative silence This recording in particular was purported to have taken place at an outdoor firing range. Notice a pronounced reflection recorded very quickly after th e originating event takes p retaining wall or barrier used to stop incoming bullets. Finally, the means of determining the direction of the sound source should be addressed. In a velocity can be derived from the time elapsed over a known distance, assuming the speed is constant. This is represented by v = d/t. When a traveling wave maintains a constant speed over a known distance, the elapsed time will b e constant as well. However, when the wave begins at a third point and travels at a constant speed along any trajectory other than perpendicular to the midpoint between two

PAGE 19

1 1 microphones the velocity and distance can be constant and the times to e ach point can vary. The wave will reach the closer point first and the further point second. Then, using the delay between the signals arriving at each channel, a source bearing can be derived, with the source originating from a point along the bearing. Th ese are the working principle s behind sound ranging, past and present, and are demonstrated in Figures 2.2 through 2.5 Figure 2.2 Angle Determination from Sound Delay between Two Microphones

PAGE 20

12 Figure 2.3 Angle Determination from Sound Delay between Two Microphones (II) Figure 2.4 Wide Angle Determination from Sound Delay between Two Microphones

PAGE 21

13 Figure 2.5 Distance d Angle Determination from Two Microphone Sound Delay With sound source S the produced sound propagates outward at a constant speed. In Figure 2.2, b ecause S is closer to L than R ( SL < SR ) the sound from S reaches L first, then R This also means radius a is shorter in length than radius b ( ra < rb ) The delay of the arrival of the sound to each channel determines the calculated direction from which the sound came ( ang ), which is assessed from the midpoint between L and R Like in Figure 2.2, the scenario in Figure 2.3 demonstrates how ang can be derived from the difference between SL and SR This time, since S is closer to R (or r b is shorter in length than r a ), the corresponding angle is in the direction from the midpoint to the R side. Example 2.4 shows that even extreme angles can be de termined using the difference between SL and SR or r a and r b The resulting angles in all of these examples are independent of the distance from the sound source to the microphones, since the derived angle is a bearing, not an absolute

PAGE 22

14 point. E ven though it involves the furthest distance from the m icrophones of all the given examples, Figure 2.5 results in the same calculation process for delay and subsequent angle. Equipment and Configuration The system begins with a pair of microphones. Microphones with high tolerance to loud impulses would be ideal for a fully functioning system used in a real operation but an inexpensive pair of smaller microphones is suitable for this design. The microphones used in this test are a pair of Olympus ME 15 microphones. These are considered a stereo pair and both capture audio simultaneously. Next, the microphones are connected to a laptop computer via a stereo input cable into the stereo mic in port. This computer is equipped with MatLab, which is a versatile computation and programming software. MatLab handles both the audio input and the command output to the microcontroller. The actual programming scripts used in MatLab and with the microcontr oller will be discussed later. A microcontroller is then attached to the computer via serial interface. In this case, the connection is via USB cable. The microcontroller of choice is the Arduino Uno, due to its versatility and extensive open source support. The Arduino accepts commands from the computer and sends a corresponding voltage to a servo motor which rotates a mounted webcam. The servo motor i s a standard HS 422 servo, and the webcam is a 5 megapixel USB webcam that is connected back to the laptop computer for display purposes.

PAGE 23

15 Figure 3.1 Gunshot Detection and Localization System, Basic Design Overall Figure 3.1 is a photo of the overall view of the system designed, assembled, and utilized for the testing outlined in this project The microphones acquire incoming sound and send to MatLab for processing. Should the incoming signal meet the threshold requirements, MatLab would process the si gnal delay and compute the angle. T he Arduino receives the angle rotation command v ia serial connection (the white cord on the right hand side of the laptop), and communicates to the servo motor with the camera mounted atop. Then, the image information from the camera is sent to the laptop through another USB connection for display and potential recording purposes. This workflow is outlined in Figure 3.2 below.

PAGE 24

16 Figure 3.2 Basic System Workflow 1) sound em anates from source, microphones capture upon arrival; 2) laptop receives and processes incoming audio information, determines possible target signal confirmation, delay, and resultant angle; 3) laptop sends command to Arduino; 4) Arduino sends voltage to servo motor with camera mount; 5) laptop rece ives resultant image information from camera

PAGE 25

17 Figure 3.3 Gunshot Detection and Localization System, Arduino and Servo Close up image of Arduino and servo assembly Programming With the system configured properly, the Arduino and MatLab need to be programmed properly. The Arduino platform works fluently with MatLab, enough so that the microcontroller can be programmed to work continuously, waiting for MatLab serial commands, exec uting the commands, and returning results if asked. To prepare the Arduino for this setup, the MATLAB Support Package for Arduino (aka ArduinoIO Package ) must be downloaded to the laptop computer. From this package, the motorsrv.pde file must be uploaded t o the Arduino IDE, and the appropriate AFMotor.h

PAGE 26

18 and AFMotor.cpp files must be allocated properly. For instructions on how to perform these steps, refer to the corresponding age. Next, a script must be written for MatLab to automat ically process the incoming audi o and send commands accordingly (Figure 3.4). %% Automatic Audio Detection/Image Response (AADIR) system % Mr. Jordan Graves, BS and Dr. Catalin Grigoras, PhD % 2012 %% Purpose: % 1. This system will acquire live audio signals. % 2. Based on defined thresholds, this system will discriminate particular acquired audio events from others. % 3. Using the perceived delay of incoming audio signals between the pair of recording channels this system will estimate the directional source of the discriminated audio signal. % 4. This system will command the servo motor to rotate the cameara array towards the perceived source of the discriminated audio signal. %% Materials Used: % (1) Ardu ino microcontroller with serial connection to computer and signal connection to servo motor % (1) 5V rotational servo motor (180 degree range) connected to Arduino % (1) webcam attached to rotating mechanism of servo motor % (2) omnidirectional microphones arranged to acquire stereo audio signal, connected to computer via stereo microphone input, through Y adapter %% Notes % works with motorsrv and add AFMotor.cpp and AFMotor.h to path: ...Arduino \ libraries \ Servo %% Script %delete(a) %connect to the board a=arduino( 'COM13' ) % define Pin#9 as output and attach the motor to it a.pinMode(9, 'output' ); % Attach servo#2 to Pin#9 a.servoAttach(2); a.servoWrite(2,90); %reset servo to center % define the audio settings % sampling frequency fs=48000; % resolution (bits)

PAGE 27

19 nbits=16; % no. of channels ch=2; % length in sec t=0.5; % signal power threshold th=1200; % window threshold size win=200; % define the audio object recObj=audiorecorder(fs,nbits,ch); %begin recording get(recObj) %collect/display values as they are recorded disp '***BEGINNING ACQUISITION***' %status message for k=1:2000 % aquire the audio signal recordblocking(recObj,t); %record without on the fly control until recording is stopped % Store data in double precision array. x=getaudiodata(recObj, 'int16' ); %signed integers mapped to set parameters (anything outside will be "rounded") % find absolute value of incoming signal xa=abs(x); % extract L and R channels L=double(x(:,1)); R=double(x(:,2)); [k max(L) max(R)] %query for maximum values during sampling "window" if max(L)>th && max(R)>th %set power threshold if xa(k:k+win)
PAGE 28

20 [a3,b3]=max(c); %define c's maximum values as a3,b3 delay2=fs/2 b3 %delay is half of sampling frequency minus b3 (maximum value for c) in samples s=delay2; if s< 127 %round values outside degree parameters to furthest degree value left or right (to maintain 180 degree range) s= 127; elseif s>127 s=127; end % convert the delay s into degrees ang ang=round((s+128).*179/256) % rotate angle ang a.servoWrite(2,ang);pause(0.01); c; end end end delete(a) Figure 3. 4 Audio Detection and Image Response Script As is standard for MatLab scripts, green lines of text with a percentage sign preceding the content are considered notes and are not executed with t he actual programming language. The comments provide guidelines for each portion of the script Primary poin ts of interest in the script include the delay calculation and angle conversion elements. % signal power threshold th=1200; % window threshold size win=200; Figure 3.5 Thresholds in Audio Detection and Image Response Script According to Figure 3.5, which is a reference to the threshold element of the script the This is a setting dependent on multipl e factors, including microphone gain settings, expected distance from sound source to microphones, and expected background noise. Due to these many

PAGE 29

21 corresponds to a threshold of 200 samples Note, the sampling frequency of the incoming audio is 48kHz. Figure 3.6 Discriminatory Threshold s for Audio Event Exclusion Figure 3.6 illust rates how a set of thresholds might work in discriminating audio signals by power and duration. At window i the signal meets both the minimum power and maximum duration thresholds. At ii the signal meets the minimum power threshold but is too long in duration. At iii the signal does not meet the power threshold. In the script used in this project, the signal must meet the power threshold, then the duration threshold, in order to provoke a system response. The other significant portion of the script used in the project pertains to the delay calculation and angle conversion portions (Figure 3.7)

PAGE 30

22 [c,lags]=xcorr(L,R); %cross correlation between vectors (automatically [a1,b1]=max(L); %fs/time of max values [a2,b2]=max(R); [a3,b3]=max(c); %define c's maximum values as a3,b3 de lay2=fs/2 b3 %delay is half of sampling frequency minus b3 (maximum value for c), in samples s=delay2; if s< 127 %round values outside degree parameters to furthest degree value left or right (to maintain 180 degree range) s= 127; elseif s>127 s=127; end % convert the delay s into degrees ang ang=round((s+128).*179/256) Figure 3. 7 Channel Delay and Angle Calculation in Detection and Response Script The lower portion of the script determines the delay of the incoming audio event between channels, and then produces the corresponding angle for the sample delay. The Arduino uses 0 degrees as a valid degree integer, so the 180 degree range actually includes 0 and spans from 0 179. Some further explanation is necessary for the delay and angle calculation portion s of the script. The xcorr max(L) and max(R) portions of the script mark the initial peak values earlier in t as .5 seconds in length). Each initial peak is marked in the numerical sample it was measured to take place. The delay is then determined from the difference in those sample val ues; if the value fell outside the allotted range, it would be rounded to the high or low extreme, depending on whether it was above or below those extremes, respectively. Assuming the speed of sound is approximately 350 meters per second, a sound wave wo uld travel 6 feet (or approximately 1.829 meters) in around .0053 seconds. 6 feet is the

PAGE 31

23 prescribed distance between microphones used in the system, .0053 seconds is the maximum delay between channels. Since the sampling frequency defined above is 48 kHz, o r 48000 samples per second, the maximum delay between channels can also be measured as approximately 256 samples. The delay is then added to 128 to account for the delay reference to the R channel instead of L and then compared to the ratio of samples to angles. The resultant value is rounded to the nearest whole integer and is the calculated angle for rotation. So, if the delay were 98 samples ( signal reaching R 98 samples faster than L ) : Figure 3. 8 Delay and Angle Calculation Example 1 ang = (98+128)* 179/256 ang = (226)*179/256 ang = 158 degrees

PAGE 32

24 On the other hand, a delay of 110 samples ( signal reaching L 110 samples faster than R ): Figure 3. 9 Delay and Angle Calculation Example 2 ang = ( 110+128)*179/256 ang = (18)*179/256 ang = 13 degrees The angle calculated from this script is measured from left to right of the midpoint between microphones. This means angles between 0 and 88 degrees will rotate the camera counterclockwise from neutral while angles between 90 and 179 degrees will rotate the camera clockwise from neutral Of course, the user may reverse which side, L or R to assign reference to, among m any other customizable features (thresholds, microphone distance calibration, etc). Testing and Results functionality a simple test was formulated and executed. The system was assembled as described above, in a series of open, outdoor tennis courts. This

PAGE 33

25 place was chosen with the intent of minimizing potential interfering reverberations, as well as other variables intro duced in more crowded areas. The test was executed at night, to reduce the chance of external noise interference from wind or passerby The air temperature was appr oximately 37 degrees Fahrenheit. This is significant because although relatively small chan ges in temperature would not affect the speed of sound in a drastic way, it is well known that larger temperature variations could introduce complications in calculating the speed of sound 13 As mentioned earlier, t he microphones used were elevated and spa ced at approximately 72 inches apart, with the notion that spacing should be towards the wider end of the spectrum to emphasize the delay b etween incoming audio channels. Markers were placed at 5, 10, and 15 foot distances from the center point between mic rophones, all distances at 10, 30, 50, and 70 degree angles from that same center point in either direction. In total, 24 markers were made. These markers indicated the intended positions from where the test sound would originate. At the time of the test, an actual firearm was not an available sound source. Instead, a loud, sharp clap of the hands was utilized at each marker. The overall waveform shape of a hand clap could properly simulate a gunshot because both events can be chara cterized with high intensity and short duration. Though the claps were kept at consistent volume, some variation in signal intensity must be acknowledged. However, the variations were conside red acceptable because of the multiple factors that introduce var iations in sound in a real life situation. The test itself was desi gned to be controlled in most reasonable aspect s yet allowed for some semblance of a realistic environment.

PAGE 34

26 After the system was assembled and initiated, the testing began After each inst ance of a hand clap at each marker, the system was reviewed for a response and possible camera movement. At each marker, the possible responses for the system were: 1. Rotation of the camera towards the sound source, stopping with the marker in the center of the camera frame (represented by in the tables below) 2. Rotation of the camera towards the sound source, stopping with the marker in the frame but not in the center (represented by O in the tables below) 3. Rotation of the camera, stopping without t he sound source in frame (represented by X in the tables below) 4. No camera movement in the response (represented by in the tables below) The responses were determined after some camera movement in response to the claps, or after a maximum 5 clap attemp ts at the marker. Each marker was tested in a trial, with three total trials making up the test. The order for marker tests varied by trial; the first two trials were in order of each degree at one distance followed by the remaining two distances, while th e last trial proceeded i n a more staggered pattern. Tables 1.1 through 1.3 illustrat e each trial and set of results.

PAGE 35

27 Table 1.1 Test Trial 1 Configuration and Results Trial 1 Results 5 feet 10 feet 15 feet 70 Stage Left (SL) 50 SL 30 SL O 10 SL O X O 10 Stage Right (SR) O X O 30 SR X X X 50 SR O 70 SR O Table 1.2 Test Trial 2 Configuration and Results Trial 2 Results 5 feet 10 feet 15 feet 70 Stage Right (SR) 50 SR O 30 SR X X X 10 SR O O 10 Stage Left (SL) O O 30 SL X X X 50 SL X 70 SL Table 1.3 Test Trial 3 Configuration and Results Trial 3 Results 5 feet 10 feet 15 feet 70 Stage Right (SR) 10 Stage Left (SL) O O 50 SR O X X 30 SL X X 30 SR X X 50 SL O O 10 SR O O O 70 SL Each test proceeded through the first column, then the middle column, then the last column, each column the viewpoint of the camera outward towards the markers)

PAGE 36

28 Discussion At first glance, the results from the test as a whole seem mixed and inconsistent with only Trial 3 producing respo nses to each marker with camera movement To make the results more relevant and truly evaluate the test results for inconsistent system responses (or were created. These are defined within the realm of the test as camera responses that end with the location of the sound source somewhere in frame, either centered or not centered. This is to simulate a real world scenario, where a video recording of an i ncident would be submitted for evidence. Potentially useful evidence in such a scenario would require the event itself or the immediate aftermath to be captured somewhere in frame, either centered or not. With in those parameters, the test resulted with Tab le 2.0. Table 2 .0 Test Trial Potentially Useful Evidence Results PUE Results 5 feet 10 feet 15 feet Trial 1 4 of possible 8 1 of possible 8 7 of possible 8 Trial 2 3 of possible 8 3 of possible 8 5 of possible 8 Trial 3 8 of possible 8 5 of possible 8 5 of possible 8 The 8 possible instances for PUE response refer to each marker at the given distance This table of PUE inconsistent. The inconsistency do es not necessarily exist on a trial by trial basis, but more within the changes between distances. While the potential for useful evidence increases in Trial 3 from the others, Trial 1 and 2 exhibit increases in PUEs as the distance grows. The opposite results occur in Trial 3, where the PUE decreases and stays the same at each respective distance This is a counterintuitive result, since one would assume that events at further distances are more likely to take place within the frame of view of a camera.

PAGE 37

29 The overall inconsistency in the system can be attributed to multiple factors. First of all, the microphones used may not be most suitable for the task of discrimination. Due to the very short duration of most gunshot sounds, microphones used should be se nsitive enough to accurately define the incoming audio information, to the point in which a series of loud impulses spaced closely together in rapid succession would be recorded as such, instead of one long, loud impulse. T he microphones used in this test were not specifically designed with that task in mind. Second, the surrounding environment plays a large role in the effectiveness of these systems. Even though the testing location and time of day were chosen with the intent that uncontrollable variables would be minimized, not everything could possibly be accounted for, and minor changes in the test conditions could introduce fluctuations in data. Finally, the programming language used in Mat Lab itself could use some review and potential upgrading In order to derive the correct thresholds at a particular setting, testing must be done to determine a combination that works best. A level of automation for thresholds might be worthy of some attention, where the system could be designed to automaticall y adjust for changes in the noise floor, etc. Real time adaptive filters might also work in terms of limiting the amount of extraneous and useless sound information that would only hinder progress, especially sounds of frequencies below around 400Hz and ab ove around 2.5kHz, which are the primary frequencies exhibited by gunfire. 14 These instructions, tests, and results are useful in describing the process for basic detection and response, but the design does carry a n underlying flaw. Any impulse that is sho rt and loud enough would trigger a camera movement, not necessarily a gunshot. This

PAGE 38

30 is because further discrimination of gunshot sounds from other noises of the same shape involves higher level filtering and analysis of the sounds, via the programming. Unf ortunately, gunshot audio usually exhibits peaks around 630Hz. 1 5 These ranges are common for a wide variety of other sounds, making discrimination much more complicated. Figure 3. 10 FFT of 22 Caliber Rifle Gunshot with Reflection Using Fast Fourier Transform to bring a signal into the frequency domain Figure 3.10 allows for some frequency analysis of the example recording shown earlier as a waveform in Figure 2.1 The frequ encies of highest intensity range from the lowest up to 2kHz, with maximum values below 500Hz. The se results do incorporate some expected for most audio events in realistic scenarios. To properly discriminate from other sounds, t he best approach at this point is an algorithmic learning strategy, such as those proposed by Morton and Collins 1 6 or

PAGE 39

31 Valenzise et al 1 7 Algorithmic learning strategies are not used in this project, but will be discussed later To properly simulate the products used in the field and attempt to replicate their functionality, vendors such as ShotSpotter should provide scientists with some data in regards to their product specifications and schematics, simply for research purposes. This but more importantly encourages unbiased review and testing in a scientific and peer reviewed cap abilities, and may improve their general public image, as observed below. At the time of this test, representatives from both ShotSpotter and Boomerang elected not to field questions about design s, functionali ties and test data.

PAGE 40

32 CHAPTER IV FUTURE POSSIBILITIES The Current State of Gunshot Detection and Localization Systems Gunshot detection and localization systems, ShotSpotter in particular, have been used to aid law enforcement for some time now, and have been met with mixed reactions. In some situations, law enforcement agencies report dramatic improvements in gunfire incident response time. In others however, the system is reported as inaccurate, cos tly, and overly demanding of valuable resources, in cluding taxpayer money In 2009, the New York Times followed ShotSpotter use in the greater New York, New Jersey, and Connecticut areas Officials praised the system for helping reduce shooting response time, which led to faster aid for victims and better suspect apprehension. According to one Sergeant, the system proves especially helpful in areas where citizens are so accustomed to gunfire that 9 1 1 calls simply do not take place any longer. 1 8 In 2011, city officials of Wilmington, North Carolina approve d a 2 year contract with ShotSpotter, with the mindset that personnel is the greatest expense, and the technology could be used to better manage staff in the field. The contract expires in mid December of 2012, and will be reevaluated by the city to determ ine whethe r or not to extend the service. 1 9 Other agencies have seen less success with ShotSpotter. In March of 2012, the New Haven Police Department adopted a new protocol of sending audio data to ShotSpotter headquarters prior to redirection to the New H aven 9 1 1 center. Though the change in routing only delays information from reaching dispatch by mere seconds, the system has reported enough false positives to cause concern and necessitate the change 20 More

PAGE 41

33 worrisome are the gunshot events that take place but are not detected, which has happened in the area. That very situation factored into a decision made by city council members of Trenton, New Jersey to reject a ShotSpotter expansion in January 2012. The p revious Christmas, a man was fatally shot and left to die on a sidewalk of a major avenue. Although a ShotSpotter sensor was set up only blocks away, no alarms were triggered. The South ShotSpotter] does 21 To further cloud judgment on system value, the city of Detroit rejected ShotSpotter installation because it was too taxing on available personnel. The Broward County Sheriff in Florida previously us ed ShotSpotter but decided to remove it, citing the system was not cost effective and came with too many false alarms. However, the Rochester Police Department in New York swear s relative question. We thi Criminal Justice professor and 26 If it gets police to one more victim sooner, how do you put a price on that? If it adds to the evidence to convict someone, how would you ad 22 To better decide whether or not to employ a gunshot detection and localization system like ShotSpotter, authorities need reliable, unbiased information in the form of extensive testing and reporting. While efficacy studies and evaluatio ns are available, most do not publis h complete statistical results. 2 3 2 4 Some studies are conducted with tangible and measurable results, but are not current within the past decade 25 A long term study with complete transparency of test materials and results will be the best means of assessing the

PAGE 42

34 effectiveness of gunshot detection and localization systems, determining strengths and weaknesses, and choosing whether to utilize the system or not. Until the day comes when leading manufacturers reveal their design plans and schematics, researchers must continue to build replicas as similar as possible to those used in the field, so results gathe red will be relevant and useful. Prospects for the Basic System Within the scope of this project, the first future developments available should be improvements and further testing. As mentioned earlier, additional testing using superior equipment such as high quality microphones, cameras, etc. would logically introduce different results than those found with the current configuration. In particular, microphones more sensitive to changing dynamics in audio signals should theoretically offer more precise measurements and findings Not only could additional upgrades offer improvements on current functions, new additions to the current microphone arrangement could literally add another dimension to th e system response capabilities. A third microphone could allow for three d imensional localization, either in the f orm of triangulation in a single plane or in space In much the same way that two microphone systems would give a direction in a single plane, three microphones would accomplish the same feat both vertically and hor izontally. If the corresponding microphones are arranged as described in F igure 4.1 such a system could theoretically provide azimuth and elevation information about a signal source.

PAGE 43

35 Figure 4.1 Source Determination from Three Microphones in Three Dimensions With two microphones such as in LR earlier two dimensional hemispherical sensing is feasible. Three microphones could allow for three dimensional sensing, along an X, Y, and Z axis accordingly. Each resulting angle produced from each pair of microphones would converge on a point in space and thereby provide a location for the source of sound. This is shown in Figure 4.1. Triangulation, often used outdoors by navigators or cartographers with comp asses, involves the within two dimensional space. By spreading the three microphones in such a way that the esulting angles derived from each pair would localize and pinpoint the perceived source to a specific location. Technically, two angles could produce a location for the sound source. However, a third angle would not only be available by default after using three different

PAGE 44

36 microphones, the third angle would provide additional correction t o minimize error in the results. Triangulation in this context is described in Figure 4.2. Figure 4.2 Source Determination from Three Microphones in Two Dimensions Using the a b b c and a c differences and resultant angles for the XY YZ and XZ pairs respectively the relative location for the sound source can be calculated. Therefore, assuming each sensor is fixed in a known coordinate, the true coordinate of the sound source can be derived from its relationship to the sensors. This particular example shows how the source can be derived in a plane. Additional microphones would not only increase the capabilities of the system, but would minimize potential error while maximizing sensing capabi lities. Soun d events, such as muzzle blasts emanating from the firearm may be highly directional and vary significantly at different angles. 2 6 The greater the microphone coverage available, the greater the potential for captu ring an adequate amount of audio information. Once these upgrades have been made, further testing would undoubtedly take place. To better estimate accuracy, the camera may be replaced with a laser pointer, which will

PAGE 45

37 provide a more quantitative means of ev aluating system response. The laser would sound source could offer measurable results for analysis and future calibrations. On the topic of calibrations, the programm ing language could also be adjusted to accommodate for different microphone spacing. This would be used to determine a minimum or perhaps even an optimum distance between microphones, which would become more important as more microphones are added and the system complexity grows. Since this system in particular was adjusted to react to gunshot like audio events, it would not be recommended for other applications as is. However, the signal power and re searchers might consider using a system described above to monitor animal activity in their natural habitats. This could even entail the building and installation of systems in extreme environments such as seafloors or mountaintops so scientists may carr y out their studies from the safety of a remote laboratory. Research using these sorts of systems in national parks has already commenced, in hopes of combating outbreaks of poaching on protected territory. 2 7 A system such as this might also be used by an instruct or teaching an on line course, so that he or she might feel free to move about the setting without worrying about whether the camera will remain fixated upon him or her while talking. While the teacher shifts to draw attention to a demonstrative object at his or her side, the detection system, attuned to his or her voic e, would compensate accordingly with a camera adjustment.

PAGE 46

38 Of course, such possibilities would depend heavily on far su perior programming to prove successful in a long term scenario. These kinds of uses would almost certainly necessitate the use of learning algorithms to allow the system to more accurately suggest whether to accept or reject incoming signals, perhaps to th e point where it may make the decision for itself and allow the user to review the decisions made and correct any errors. Though some might suspect a complicated set of equations and commands must steer a discriminatory algorithm, one study in particular 28 actually listed a series of relatively simple correlation s as one of the more robust methods tested. Templates of gunshot averaged and stored in vectors 1000 samples in lengt h. At 39 samples per iteration, a correlation between the incoming signal and averaged templates is calculated for each iteration and stored in a pair of vectors for comparison against given thresholds. Testing revealed a True Positive Rate (true positives detected divided by total number of positives) of .91, with a False Positive Rate (false positives detected divided by total number of negatives) of 0.0. Using this sort of system would require little in terms of additional programming skill or hardware r esources, and could be easily customizable for Similar to the time based correlation method is another adapted from a document released by ShotSpotter to inform about gunshot location systems. 29 With sensitive equipment and comprehensive pro gramming an accurate frequency envelope can be developed for the gunshot event. This envelope could be stored as a series of data points, along with a large volume of multiple series of data points generated in similar fashion all normalized to the same length Storing such data would require much less in terms of

PAGE 47

39 relative capacity when compared to storing numerous uncompressed recordings An incoming signal could be broken down into the same sort of envelope, and then compared at target frequencies to th If the incoming signal shared a high enough correlation in frequency to known templates, an automated decision could be made to alert the system and direct a response. On the other hand, data points could be sto red for false positives ( heavy machinery backfiring automobiles, and the like), so that if the incoming signal shared high enough correlation with false positives, the system could notify the operator of a poten tial false alarm needing review and confirmation. As each eve nt takes place, the incoming signal values could be designated as true or false positive, and then the system would subsequently This sort of strategy would function best using highly sensitive microphones and large d atabases for comparison which could restrict processing speeds and inflate costs While some might question the effectiveness of this type of programming against false positives as deceptive as other small explosions like fireworks, the opposing viewpoin t is that a system prone to more false positives is much more acceptable than one prone to error in false negatives. Nevertheless, researchers have shown that even firecrackers differ from gunshots in terms of frequency domain bandwidth 30 meaning t his proposal in particular r equire s further development but shows some promise.

PAGE 48

40 CHAPTER V CLOSING REMARKS Gunshot detection and localization s ystems have attracted attention of nearl y every sort implementation in law enforcement scenarios, others remain skeptical of t heir effectiveness versus costs At this time, it is unclear whether or not gunshot detection and localization systems deserve widespread installation or banishment, but the most reas onable course of action is to continue testing and development, with hopes of constant improvement and sim ultaneous research transparency. Fortunately, basic designs such as the one outlined in this thesis combined with the enormous potential for their us e in gunshot response or other related applications, allow for progressive scientific development and advancements This in turn may give rise to a universally reliable system to use in law enforcement. No matter the means, the bottom line is this: authori ties and law enforcement personnel must continue to pursue a healthy partnership with technology to combat the ever changing threat of crime and violence In the educated words of Criminal Justice professor Dennis Kenney, 31

PAGE 49

41 REFERENCES 1. rang ing in World War I. Notes and Records of the Royal Society. 2005;(59):273 284. 2. Melson CD. Organization and Equipment for the Defense Battalion. In: Condition Red: Marine Defense Battalions in World War II. United States Marine Corps Editing and Design Sec tion, History and Museums Division. 1996. 3. Locating Surveillance and Target Acquisition Association. Locating Artillery Overview. http://www.locatingartillery.org/overview.htm 4. Varshney L. Ground Surveillance Radars and Military Intelligence. North Syracu se (NY): Syracuse Research Corporation; 2002 Dec. 5. Lahr JC, Ward PL, Stauffer PH, Hendley JW. Earthquake Technology Fights Crime. United States Geological Survey Earthquake Hazard Reduction Program. http://jclahr.com/science/psn/gunshots/factsheet/index.ht ml 6. ShotSpotter, History. http://shotspotter.com/company/history 7. Hsu J. Tracing a Bullet Back to Its Gun. ScienceLine.org. 2007 Jan. http://scienceline.org/2007/01/tech hsu guns/ 8. Defense Industry Daily. Sniping at US Forces Beginning to Boomerang. DefenseIndustryDaily.com. 2011 Feb. http://www.defenseindustrydaily.com/sniping at us forces beginning to boomerang 01128/ 9. Raytheon BBN Technologies. Boomerang Warrior X. 2010 Jul. http://bbn.com/resources/pdf/Boomerang Warrior X 072210.pdf 10. Al Muslim A. Gunfire detection coming to Hempstead Village. Newsday.com. 2012 Jul. http://www.newsday.com/long island/nassau/gunfire detection coming to hempstead village 1.3828324 11. Haag MG, Haag LC. Shooting Incident Reconstruction. 2 nd rev. ed. San Diego: Academic Pr ess, 2011. 12. Maher RC. Acoustical Characterization of Gunshots. Institute of Electronics and Electrical Engineers. 2007 Apr. 109 113.

PAGE 50

42 13. Dean EA. Atmospheric Effects on the Speed of Sound. El Paso (TX): United States Army Electronics Research and Development Command; 1979. Report No.: ASL CR 79 0100 4. 14. Haag MG, Haag LC. 15. Ibid. 16. Morton KD, Collins L. Bayesian Detection of Acoustic Muzzle Blasts. Sensors and Command, Control, Communications, and Intelligence Technologies for Homeland Security and Homeland Defen se VIII (proceedings of SPIE). 2009; 7305. 17. Valenzise G, Gerosa L, Tagliasacchi M, Antonacci F, Sarti A. Scream and Gunshot Detection and Localization for Audio Surveillance Systems. Institute of Electronics and Electrical Engineers. 2007. 18. Buckley C. High http://www.nytimes.com/2009/11/22/nyregion/22shot.html?pagewanted=1&_r=2 &emc=eta1 19. Feskos B. Wilmington police hope gunfire sensor will help response times. StarNewsOnline.com. 2011 Jun. http://www.starnewsonline.com/article/20110609/ARTICLES/110609693?p=1&t c=pg 20. 2012 Mar. http://www.nhregister.com/articles/2012/03/12/news/new_haven/doc4f5d6e0426b cf101484344.txt? viewmode=fullstory 21. system. NJ.com. 2012 Jan. http://www.nj.com/mercer/index.ssf/2012/01/trenton_council_rejects_expans.htm l 22. Hong A. RPD: ShotSpotter Valuable to Crime Solving. 1 3WHAM.com. 2012 Jun. http://www.13wham.com/news/local/story/rochester shotspotter police alerts sensors/uSDQnaKgw0i CRyzn4HQ_w.cspx 23. Selby N, Henderson D, Tayyabkahn T. ShotSpotter Gunshot Location System Efficacy Study. CSG Analysis, endorsed by the Nati onal Organization of Black Law Enforcement Executives. 2011.

PAGE 51

43 24. Platt J. ShotSpotter Stat Update. TheHuntingtonian.com. 2012 May. http://thehuntingtonian.com/2012/05/25/shotspotter stat update/ 25. Mazerolle LG, Watkins C, Rogan D, Frank J. Random Gunfire Pro blems and Gunshot Detection Systems. Washington (DC): National Institute of Justice, United States Department of Justice; 1999 Dec. NCJ No.: 179274. 26. Beck SD, Nakasone H, Marr KW. An Introduction to Forensic Gunshot Acoustics. Proceedings of the 162 nd Meet ing of the Acoustical Society of America; 2011 Oct 31 Nov 4; San Diego (CA) 27. Gonzlez Castao FJ, Alonso JV, Costa Montenegro E, Lpez Matencio P, Vicente Carrasco F, Parrado Garca F, et al. Acoustic Sensor Planning for Gunshot Location in National Parks: A Pareto Front Approach. Sensors 2009 Nov;(9):9493 9512. 28. Chacn Rodrguez A, Julin P. Evaluation of Gunshot Detection Algorithms. Proceedings of the Argentine School of Micro Nanoelectronics, Technology and Applications; 2008:49 54. 29. Showen RL, Ca lhoun RB, Chu WC, Dunham JW. Gunshot Location in Complex Environments Concepts and Results. In: Carapezza, EM, editor. Acoustic Sensors, and Command, Control, Communications, and Intelligence Technologies for Homeland Security and Homeland Defense VII (pro ceedings of SPIE). 2008; 6943. 30. Ibid. 31. Buckley C.