Master of Science Thesis in Electrical Engineering

Department of Electrical Engineering, Linköping University, 2023

Gunshot Detection and

Direction of Arrival
Estimation Using Machine
Learning and Received
Signal Power
David Grahn & Timothy Cooper
 Master of Science Thesis in Electrical Engineering
Gunshot Detection and Direction of Arrival Estimation Using Machine Learning
and Received Signal Power
David Grahn & Timothy Cooper
LiTH-ISY-EX–23/5594–SE

Supervisor: Gustav Zetterqvist

isy, Linköping University

Examiner: Fredrik Gustafsson

isy, Linköping University

Division of Automatic Control

Department of Electrical Engineering
Linköping University
SE-581 83 Linköping, Sweden

Copyright © 2023 David Grahn & Timothy Cooper

Abstract
Poaching is a persistent issue that threatens many of earth’s species including the
rhino. The methods used by poachers are varied, but many use guns to carry out
their illegal activities. Gunfire is extremely loud and can be heard for kilometres.
This thesis investigates whether it is possible to aid anti-poaching efforts in Kenya
with a gunshot detection and estimation device using an array of microphones. If
successful, the device could be placed around the savannah or any exposed area
and warn if poaching is taking place in the nearby. If a shot is fired within the
audible range of the device’s microphones, a trained machine learning algorithm
detects the shot on the edge using a microprocessor. The detection runs in real
time and achieved an accuracy of 93% on an unbalanced data set, where the
majority class was the one without gunshots. Once a detection has been made, the
received signal power to each microphone is used to produce a direction of arrival
estimate. The estimate can produce an angle estimate with a standard deviation
of 66.78° for a gunshot, and with a standard deviation of 7.65° when testing the
model with white noise. Future implementations could use several devices that
detected the same event, and fuse their estimates to locate the shooter’s position.
All of this information, as well as the sound file, can be used to alert and assist
local wildlife services. The challenges of this project have been centred around
making a system run in real time with only a microprocessor on the edge, while
also prioritizing low cost components for future deployment.

iii
Acknowledgments
We want to extend our thanks to Gustav Zetterqvist who served as a supervisor
for this project, he has been instrumental with his guidance during the project.

We also want to thank Fredrik Gustafsson who has been the examiner for the
project. We want to thank him for the opportunity to partake in project Ngulia it
has been a fun challenge and a nice experience.

Carlos Vidal is also a person we want to thank. He has helped us a lot with the
hardware aspects of this thesis, without him the hardware would not have been
developed to the point it was.

A person who’s aid has been instrumental is major Fredrik Perlaky who is the
deputy head of education at Markstridsskolan Kvarn. He arranged for us to at-
tend a field exercise with our equipment, allowing us to record many shots being
fired. The data he allowed us to gathered served as all the positive data in our
training set without which the project would have suffered.

Another person deserving of our thanks is reserve lieutenant Benny Larsson at

the marine base in Karlskrona. He allowed us to attend one of his training ses-
sions with our equipment. The data he facilitated served as all the positive data
in our testing set as well as directional estimate to real gunshots.

Linköping, June 2023

Timothy Cooper and David Grahn

v
 Contents

Notation ix

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Delimitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Component overview . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Operational overview . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.6 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.7 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.8 Related research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.9 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Hardware and software implementation 7

2.1 Prototype development . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 deu hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Sony Spresense . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Spresense software . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Operational modes for the detection model . . . . . . . . . 10
2.3.2 Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Detection 13
3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Feature based methods . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Neural Net . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Evaluating machine learning models . . . . . . . . . . . . . . . . . 22
3.3.1 Unbalanced data sets effect on model precision . . . . . . . 23
3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

vii
viii Contents

4 DOA Estimation 31
4.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.1 Signal model . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.3 Fourier Series Model . . . . . . . . . . . . . . . . . . . . . . 33
4.1.4 Frequency dependency . . . . . . . . . . . . . . . . . . . . . 33
4.1.5 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.1 Known good calibration . . . . . . . . . . . . . . . . . . . . 35
4.2.2 Prototype 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.3 Prototype 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.1 Known good data . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.2 Prototype 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.3 Prototype 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.4 Solver error . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Conclusion 51
5.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.1 Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.1.3 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Bibliography 57
 Notation

Notations

Notation Meaning
yi Microphone signal for microphone i
si Received signal for microphone i
wi Microphone measurement noise for microphone i
l Measurement time
L Number of measurements
Pi Power of signal for microphone i
ei Measurement noise power for microphone i
ψ Angle to sound source
α Absolute received power
gi Gain for microphone i
h(ψ, θ i ) Microphone directional sensitivity of microphone i
θi Fourier series model parameters for microphone i
K Number of observed directions

ix
x Notation

Abbreviations

Abbreviation Explaination
bic Bayesian information criterion
deu Detection-estimation unit
doa Direction of arrival
dsp Digital signal processing
dt Decision Tree
fs Fourier Series
knn K-Nearest Neighbour
ls Least squares
nb Naive Bayes
nls Non-linear least squares
rss Received signal strength
sfs Sequential forward selection
svm Support Vector Machine
stft Short Time Fourier Transform
tdoa Time Delay of Arrival
 Introduction
1
This thesis which is part of the larger project Ngulia which aims to ensure the
future of wildlife preservation in Africa. This thesis will give aid by detecting
and estimating the direction to poachers. This is done by detecting shots fired
from poachers though machine learning, and then calculate the direction the
sound came from. An alert with the information can be sent to local wildlife
protection services, who hopefully can catch the poachers and perhaps even save
the animal. This thesis work consists of two large parts: detection and direction
of arrival (doa) estimation. Each part has its own chapter that describes relevant
theory and results, with an additional chapter discussing the hardware.

1.1 Background
Poaching has long been a problem for both African elephants and rhinos. African
forest elephants and black rhinos are classified as critically endangered [6, 9],
while the savannah elephant is classified as endangered [10]. White rhinos have
made a major population recovery since the end of the 19th century but are still
classified as vulnerable [7]. According to the International Union for Conserva-
tion of Nature (IUCN), poaching is the major cause for individual deaths and
population decline for all species. While sources disagree on the exact numbers,
it is likely that between 10,000 and 35,000 African elephants are killed every year
through poaching [1, 8, 25]. The number of rhinos killed annually is around 500
to 1,000 [2].

Poaching of large animals is most often carried out with firearms. Kalashnikov-
pattern rifles are commonly used since they are cheap and abundant but are gen-
erally not as effective as hunting rifles made for large game. Some poachers will
also employ sound suppressors or even forgo firearms entirely to avoid detection.

1
2 1 Introduction

Tranquilliser darts and poison-tipped arrows have been used where the likeli-
hood of detection by park rangers is high [26].

Downrange from a fired shot, one or two distinct sounds are produced depending
on whether the bullet is subsonic (slower than the speed of sound) or supersonic
(faster than the speed of sound). The first is the crack of the shock wave if the
bullet is supersonic, and second is the the muzzle blast that is created when the
bullet leaves the barrel. If the bullet is subsonic, no shock wave is produced. Ad-
ditionally, if the direction of fire is away from the measuring device the muzzle
blast is the only sound that can be observed regardless of the speed of the bullet.
A third type of sound is also produced by mechanical workings and residual gas
releases from the gun, however these are of much lower amplitude than the previ-
ous sounds and are therefore irrelevant to the project. It is important to note that
a shock wave and muzzle blast sound may appear to come from slightly different
directions [13]. An example of this occurring can be seen illustrated in Figure 1.1.

Figure 1.1: How a shock wave and muzzle blast may appear to come from
different direction to an observer downrange. The red arrow is the apparent
direction from the shock wave and the green arrow is the direction based on
the sound from the muzzle blast.

A system that can detect and locate users of firearms can allow park rangers to
respond quickly to poachers. Depending on the sensitivity and spacing of the
individual sensors, it is likely that even air or CO2 powered tranquilizer guns
could be located within seconds of a shot being fired.
1.2 Aim 3

1.2 Aim
This work aims to develop a novel method of detecting gunshots and estimating
the location of the shooter. To perform this task a device that we call Detection-
Estimation Unit (deu) has been constructed. The implementation of this device
will help to ensure that future poaching is difficult to the point of cessation. The
deu is a simple device consisting of a microcontroller, a microphone array and
a solar power source had been created to perform the detection and estimation.
The array contains 4 microphones, where currently one is used for detection and
all are used for estimation. A doa estimate is made by utilizing a new method
that only relies on the received power to the microphones [31].

1.3 Delimitations
The finished deu was made using a commercially available microcontroller and
microphones to simplify development. The external chassis was made using a 3D
printer to fit requirements specific to the project. The deu is powered by a solar
panel with accompanying 18650-batteries and a charge controller.

The data gathered for this project does not contain any shock waves. To observe
a shock wave the observer needs to be downrange from a gun firing live ammo.
For the shock wave to be distinct from the muzzle blast sound the observer also
has to be at a distance from the shooter. Observing a shock blast requires a lot of
safety procedures which were not able to be performed for this thesis.

Only shots from guns fired without a suppressor will be attempted to be detected
in this project. As such no investigation into the efficacy of detecting with sound
suppressors has been made. Gathering data with suppressors would add a diffi-
cult step to the data collection process, which is why it is excluded. If a device
that detects and estimates non suppressed gun shot sounds could be made the
same principals should be able to make a device that can do the same for sup-
pressed gun shot sounds.

1.4 Component overview

During the project a series of prototypes were used. These varied in their layout,
mostly based on current knowledge and limitations at the time when they where
created. At the start of the thesis the plan was eight microphones in the array.
However, this was quickly lowered to six to save on the limited number of micro-
phones available and then to four when it was realized that the microcontroller
only supported up to that number of microphones. Details about the prototypes
and their development is found in Chapter 2.1. The final deployed deu can be
seen in Figure 1.2
4 1 Introduction

Figure 1.2: The deu deployed in Ngulia.

1.5 Operational overview

Since the shots are likely to be infrequent and it is not feasible to send a contin-
uous stream of audio, the detection itself must be performed ’on the edge’, i.e.
on the DEU itself. The device can operate in one of two modes, continuous and
non-continuous, with some variations in recording method. An explanation of
these modes can be found in Section 2.3.1. Had networking been operational the
device would then send the recording in a HTTP request over LTE together with
the time of detection and the identity of the device. At the receiving server, a
DOA estimate would be made and, if there were several detections made within
a short period of time, a fused estimate to find the location of the shot could also
be made. An approximate location could then be given as a latitude and longi-
tude on a map. All the relevant information and the sound file could then be sent
as an alert to wildlife protection services. The sound file is included so that a
ranger could listen to the file to see if it is plausible that a shot has been detected
and is not a false alarm.

1.6 Research questions

1. How can a gunshot be detected?

2. How can the location of a gunshot be estimated using the hardware setup?
3. How accurate is the detection/estimation?
4. How far away can a gunshot be detected?

5. What needs to be done so that the hardware can withstand the environment
in the savannah?
1.7 Contributions 5

1.7 Contributions
The project provided a quite obvious division of labour based on the expertise
of the individuals involved. Timothy studying a master in data driven analysis
and machine intelligence took responsibility of the detection part of the project.
David studies a master in mechatronics and took responsibility for the estimation
part of the project. The remaining work was divided to get hardware working,
where Timothy did most of the work on the final deu and David did most of
the work on the second prototype as well as the Arduino code running on the
microcontroller.

1.8 Related research

Another project that also is working on gunshot detection with machine learning
is Safe Reaction [3]. Their model does not only aim to detect gun shot sounds but
are trying to differentiate a lot of different sounds. Their website shows that they
have managed to get a gun shot recall of 86% which is impressive and it shows
that it is possible to detect gun shots well in some environments.

Way to estimate shooter location have seen research before such as in an article
written by David Lindgren, Olof Wilsson, Fredrik Gustafsson and Hans Habber-
stad that did this using a distributed network of microphones [18]. An interesting
discussion about the different sounds that a gun makes is included that may be
relevant for future work.

The estimation method used in this thesis was developed by Gustav Zetterqvist,
Fredrik Gustafsson and Gustaf Hendeby [31]. They had the idea of using received
signal strength to estimate direction of arrival. This thesis implementation also
tested how well it would work in practise.

Also related to novel ways of making doa estimates is a method that uses a Taylor
series expansion of the received signal [12]. The paper that describes this estima-
tion method was written by a team consisting of Fredrik Gustafsson, Gustaf Hen-
deby, David Lindgren, George Mathai, and Hans Habberstad. This method also
has benefits similar to the received signal power method such as it can have an
arbitrary array configuration and use a much lower sample rate than traditional
time delay of arrival methods.

1.9 Outline
This thesis consists of four large parts. The first part, Chapter 2, discusses the
hardware used in the thesis. The chapter details the configuration and devel-
opment of the prototypes and the components that make up the deu are listed
along with their purpose.
6 1 Introduction

The second part is the detection which is discussed in Chapter 3. The chapter
explores what sounds are produced when a gun is fired along with how to detect
them using machine learning. The data that was gathered and made available
to the models are also presented along with how the data was gathered. Also
included in the chapter are what features were constructed and how they were
chosen. How to evaluate a machine learning model along with an evaluation of
the different models is also part of the chapter. It ends with an analysis of the
detection algorithms performances as well as everything related to detection.

Chapter 4 presents the doa estimation. The chapter describes how to the signal
model is constructed and trained to preform the estimation using the received
signal power method. Results are presented for two of the prototypes as well as
from an array equipped with high quality microphones that was used to validate
the capabilities of the code and provide comparison.

The final part, Chapter 5, presents our conclusions. A final evaluation of the
different parts of the project is made here. A large portion is dedicated towards
what should be done in the future. There are parts of the project that we have left
unfinished and that should not be left so.
 Hardware and software
2
implementation

This chapter will discuss the hardware and software implementation that was
used in the thesis work.

2.1 Prototype development

Before a built prototype was finished, some testing of the estimation algorithm
was done on a "ReSpeaker USB Mic Array" equipped with four microphones.
However this device had its microphones on the top and with an omnidirec-
tional pickup-pattern which made it unsuitable for final implementation since
the method relies on directional sensitivity of the microphones to be different.
The first prototype, henceforth referred to as prototype 1, was a green hexagonal
box that used four unevenly placed microphones at 0°, 120°, 180° and 300°. The
first prototype was built when the availability of microphones was thought to be
limited and had room for six in the array. However when the microcontroller
arrived only four microphones where used as this was the limit.

The second prototype, henceforth referred to as prototype 1, was built with the
knowledge from the first and had four evenly spaced microphones in the octag-
onal shell and were connected with the same PCB:s that where later used on the
deu. The microphones were placed at 0°, 90°, 180° and 270°. Both of these first
two prototypes did not have internal room for the microcontroller or a power
source and where instead powered with an external battery. The final deu was
built to also house its battery and solar power manager internally which created
additional constraints on space. This was also the only one of the devices that
could be fully sealed. Pictures of all the prototypes can be found in Figure 2.1.

7
8 2 Hardware and software implementation

(a) ReSpeaker USB Mic Array (b) Prototype 1

(c) Prototype 2. Unfortunately only (d) Final deu with solar panel
one microphone is present in the picture.

Figure 2.1: The devices used for testing up to final implementation.

2.2 DEU hardware

The hardware in the final implementation of the deu consisted of the following
components:

• 1 Sony Spresense microcontroller with LTE extention board

• 4 Knowles Cornell II Digital MEMS microphones

• 1 SD-card 32GB

• 1 DFRobot Solar Power Manager 5V

2.2 deu hardware 9

• 1 Sunnysolar Semi Flexible Monocrystalline Solar Panel 5V

• 2 18650 Lithium-ion batteries

• 1 3D-printed shell with inserts for microphones

• Various cables and connectors

Some of the items on the list merit further examination and are discussed in the
following sections.

2.2.1 Sony Spresense

The Sony Spresense is an unusually powerful microcontroller with 6 ARM Cortex-
M4F cores and 1.5MB of RAM capable of running various embedded software
solutions including Arduino. With an LTE extension board there is room for 4
digital microphones, an SD-card reader and a SIM-card holder.

There are some oddities and issues that have arisen from using the microcon-
troller. One of the strangest things was the fact that the Spresense would not op-
erate if the board was exposed to direct sunlight. However rather than crashing
outright, which would require a manual restart, the board seemed to just stop as
if it had no power when in sunlight and resume working when back in the shade.
Exactly why this happened was not investigated further, but did result in some
lost recordings early in the project.

Other issues stem from what has been omitted from the Spresense. A WiFi mod-
ule would simplify testing of networking components significantly easier even if
the remote deployment location would instead use LTE. A battery monitor would
also allow for finer control for power management including telling the device to
sleep when battery levels are low. Another less urgent feature would be the abil-
ity to provide over the air updates to the firmware in order to test new models
and provide software support remotely without having to connect the Spresense
to a computer.

2.2.2 Shell
The device was left out during a night with rainfall and survived without damage
to any of the internal components. For future implementations some care should
be taken to make sure the microphones are well protected from the elements as
they are the most valuable and exposed components.
10 2 Hardware and software implementation

2.2.3 Connections
The connections where rushed and the soldering made without proper knowl-
edge of better methods which resulted in a system that risks being disabled by
small disturbances.

2.3 Spresense software

The software running on the Spresense was written in Arduino in order to utilize
the built in libraries exported from the Spresense SDK. This provided easy to
use libraries for audio recording, file management and LTE connectivity. The
model and its parameters were exported from the Edge Impulse project page
and imported as an Arduino library. A code example was modified to run the
inference process.

2.3.1 Operational modes for the detection model

The DEU can operate in one of two detection modes, each with its own advan-
tages and disadvantages.

The more basic option is non-continuous mode where a recording is made into
an inference buffer until the a full window of data has been collected, then a clas-
sification is made on the data and a prediction is produced. The buffer is then
emptied and the process is repeated. During the time that the dsp and classifi-
cation are being performed, no audio is collected, resulting in gaps where events
may be missed. However, since the entire buffer is stored it can easily be saved
to a file if desired. This makes it simple to send the actual audio in the case of a
detection. However, this only saves one of the channels for reasons that will be
discussed later.

Continuous mode uses smaller sample buffers, called slices, that are some frac-
tion of the size of the full window. The slices are placed in an inference buffer
as a FIFO sequence that corresponding to the window size. When a new slice is
ready the oldest is removed from the back of the inference buffer and the new
one is added. After each iteration, the inference process is performed on the cur-
rent contents of the inference buffer. For example, a model with a model window
size of 2 seconds and 4 slices would have 500ms slices and would run the infer-
ence process on each 4 times. This mean that even if a shot occurs on the edge
between two slices it will not be missed. For the actual implementation a double
buffering method is used. Two slice buffers are created, one is used for the au-
dio sampling process and the other is used for the inference process. When the
system is started, one of the buffers is filled with audio samples, while the infer-
ence process waits for this buffer to be full. When the buffer is full the inference
process takes over and the sampling process starts sending the data to the other
buffer. Each time a buffer is full, the buffers are swapped to ensure that there
is always a place to collect the audio samples. Depending on the hardware and
2.3 Spresense software 11

model setup, the parameters for the number of slices and sample buffer size must
be tuned to ensure smooth operation, otherwise crashes are likely to occur.

There are however other drawbacks to this approach. The entire window sized
buffer is handled and stored internally by Edge Impulse and is seemingly inacces-
sible. Instead, a recording is instead started after a detection is made while the
inference process is suspended. Storing the data separately in its own window
length buffer is both inefficient from a data management perspective and may
not be possible due to limited RAM.

A shared problem in both approaches is that only one of the microphones can
send data to the model. In the case of non-continuous mode this also means that
only one channel is recovered when saving the buffer. The first issue may be
solved by sending each microphone to a separate subcore for processing. This
will however require an inference buffer for each microphone, or two if running
in continuous mode. If storing the whole window at the same time in some cir-
cular buffer is desired, there simply is not enough memory. Some things can
be done to mitigate this issue: a shorter window, lower sampling rate and more
slices can all be used to reduce memory usage. More slices will however lead to
more time spent on processing and classifying the data.

2.3.2 Networking
The Spresense only officially supports Truphone and Soracom as LTE operators
in a limited number of countries. These are operators that specialise in IoT ap-
plications. There was limited documentation as to how to connect with other op-
erators including the one used in Kenya, Airtel. Despite using the same settings
as other LTE-enabled devices the Spresense was unable to establish a connection
to the cellular network. Whether this was due to an unknown incorrect setting,
weak signal strength, compatibility error or some other reason, the source of the
issue could not be found. As a result the project proceeded without networking
capabilities.
 3
Detection

When a firearm is discharged it produces an extremely loud sound that can travel
several kilometers. Microphones can be used to record this sound and a detection
algorithm can attempt to determine whether or not a recording contains a gun-
shot. This section will describe how to construct a detection algorithm that can
determine if a recording contains a gunshot and what the results were.

As mentioned previously in Section 1.1, when a gun shot is fired two sounds are
produced that could be of interest when designing a detection algorithm. Those
two sounds are the muzzle blast and the shock wave from the bullet traveling
past the observer. For this thesis the shock wave is ignored as it was not able
to be observed. The muzzle blast contains more power and can always be heard
when a bullet is fired, rather than just in a cone in the direction of the bullet. All
the methods later described are therefore always only trying to detect the muzzle
blast. A spectrogram of a shot can be viewed in Figure 3.1.

13
14 3 Detection

Figure 3.1: A two second spectrogram of a gunshot.

The figure shows the spectrogram of a gunshot that was recorded in Karlskrona,
Sweden. As mentioned it only contains the sound of the muzzle blast. In the
figure it is clear when the sound of the gunshot starts. The sound then dissipates,
but where exactly the sounds is determined to end is a difficult question as the
rumbling of a bullet can last for a few seconds. Most of the energy is present at
the first few milliseconds after which some echos that contain a bit less energy
and the rumbling that slowly dissipates.

Two second recordings were chosen with the idea of having one second overlap,
with which there would be at least an uninterrupted second of any shot. If the
shot started toward the very end of one recording that might not be enough to
make a detection. Without overlap the next sound segment might not either be
able to make a detection as most of the energy would not be present in it. With
overlap the next recording would have the start of the sound and the following
3.1 Data 15

second, which ought to be plenty to make a detection.

The other reason two second recording were chosen is that when a detection is
made the recording could be sent to the rangers allowing them to listen to the
sound and determine if it was a false alarm or if it was a shot. If the recording is
too short it can be hard for a human to determine what is present in the audio.

3.1 Data
There are three types of data in the context of machine learning: training, vali-
dation and test data. Training data is used to train the model, validation data is
then used to see how well the model performs when evaluating different methods
or making feature selections. When the model is fully complete, it is evaluated
against the test data to determine its overall performance.

The best kind of test data is made up entirely new data, with no dependence on
the data in the training set. It is therefor important to not change the model af-
ter evaluating it on the test data, doing so could introduce a bias in the models
toward the test data and the data is no longer independent. The purpose of the
test data is to show how the models would perform when deployed where it will
process new data. To this end, the test data in this thesis was gathered at a differ-
ent occasion from the rest of the data to avoid any dependence to the rest of the
data.

The files were manually labelled to indicate whether they contained a shot or not.
If a shot could be discerned in the recording, a two second segment was extracted
and labelled as a shot. Since the recording equipment always had four channels,
each microphone was listened to and labelled separately. This also meant that
there could be a scenario where one a shot could be heard on one microphone
and was labelled as such, while another microphone at the same time could not
discern the same shot. Since these sound segment cannot receive the same label
it quadrupled the amount of labeling work.

3.1.1 Data collection

The first proper data set was collected during a military exercise at MSS Kvarn
using Prototype 1 and an unfinished Prototype 2. Since the exercise involved the
combatants hiding and moving through the forest there was little to no metadata,
but the distance varied approximately from 0.1km to 1.5km as the audio collec-
tion equipment was moved further away from the training area. Also due to the
nature of the exercise, sometimes more than shots were fired at the same time or
in rapid succession.
The method to collecting the data without knowing exactly when a shot would be
fired was to place the recording equipment in one spot and wait for the combat-
ants to find each other, then moving some distance further away when shooting
16 3 Detection

ceased and repeating until the road ran out. This gave a data set with a wide
range of distances to gunshots.

It was also at Kvarn that it was realised that the Sony Spresense would stop work-
ing when exposed to light as discussed in Section 2.2.1. Since neither prototype
had a lid at the time to cover the micro controller, special care was taken to en-
sure that the device was kept in the shade, but this was not always possible when
moving and as such some data was lost. The data from this occasion was chosen
to be training and validation data with 67% of the points randomly determined
to be training data.

Another set of data was collected from a lone rifleman on a training field in Karl-
skrona. Both Prototype 1 and Prototype 2 were used to collect the data. As it was
possible to communicate without interfering with an exercise it was requested
that the shooter fire a groups of 5-7 shots, then to wait a minute for us to move
further way. This meant that the shots could be more clearly distinguished. Only
distances up to 850 meters are present in this data set, beyond which shots could
not be distinguished as the place was hilly and forested. The shooter used a
Swedish Automat Karbin 4 (AK4) which fired 7.62 x 51 mm NATO ammunition.
This data set was selected as test data.

There are no shock waves present in the data set as they could not be observed
safely. At Kvarn they were not firing live ammo, so a shock wave was not created
since no bullet left the barrel. At Karlskrona live ammo was used but it was
not deemed safe to be in front of the gun. Furthermore the shooting range was
quite short, this means that even if the recording equipment could have been
placed down range from the shooter it is certain that a shock wave would could
be distinguished from the muzzle blast. The device being placed close to the gun
means that the two sounds would have been so close together that the sound of
the muzzle blast would probably have over powered the sound of the shock wave.
All the negative data was gathered in Kenya at two occasions. One occasion was
in the Ngulia rhino sanctuary with two deu devices. The data from one device
was made to be training and validation data, with the same split as the positive
data. The other devices data was used as testing data. The other occasion was on
the Kenyan savannah but not in the Ngulia rhino sanctuary, as the data from the
sanctuary better depicts conditions in the rhino park this data was not suitable
as testing data and was used as training and validation data.

3.2 Method
Two different approaches were explored to produce different detection algorithms.
The first approach was a simple machine learning solution and the other was a
neural network solution. To detect whether a shot has been fired, a supervised
approach has been used in both cases. The supervised approach allows labelled
data could be produced and fed through the algorithm, and can exclude sounds
3.2 Method 17

that might sound similar to gunshots. An overview of the detection algorithm

when using simple machine learning can be viewed in Figure 3.2.

Figure 3.2: Overview of the detection system.

3.2.1 Feature based methods

The algorithms that has been explored for the feature based method are: Deci-
sion Trees, Naive Bayes Random forest, K-nearest neighbour and Support Vector
Machines. More information on Decision Trees, Naive Bayes and Support Vector
Machines can be viewed in [16].

Decision trees, dt, are models that sort the data using queries of the features. The
order in which they query the features are arranged in a tree like structure, hence
the name. Each node in a dt represents a feature and the branches from the node
represent different values the node can assume. A classification is made starting
at the root node where the first feature is queried and then its branches followed,
once a leaf has been reached it will have a result and a classification for the file
can be made. An example of a DT can be seen in Figure 3.3.
18 3 Detection

Figure 3.3: A Decision tree

The image is taken from [16], in the figure there are four nodes with different at-
tributes or feature and from them there are branches with different alternatives
depending on the value of the feature. In this classification a yes or a no classifi-
cation is made.
P (i|X)
Naive Bayes, nb, models work with a very simple equation R = P (j|X) , where i
and j are different classifications and X are the features. If R is greater than 1 it
predicts class i otherwise class j. nb has short computational time since it is so
simple.

Random forest models are an ensemble learning method [5]. During training sev-
eral decision trees are constructed. When classifying, all the dt are tested and
the class that most of the dt:s predicted is chosen. Random forests are not as
prone to over fitting but can require more computing power than a decision tree
since it needs to train multiple.

The K-nearest neighbour, knn, method relies on comparing new data points po-
sition in the feature space to the position of the points in the training set [14].
The K points that are closest to the new point will decide the classification. The
majority class of these K points will be the class for the new data point. In the
event of a tie there are many different approaches to make a classification like
increasing or decreasing k, randomizing and many more.
3.2 Method 19

Support vector machines, svm, relies on separating the classes using a vector or a
hyperplane. The optimal plane should separate the classes with as large a margin
as possible, all the training points should be far from the plane and the classes
should be maximally separated. When a new data point is to be classified its
position in the feature space is compared to the hyper plane that separates the
classes. svm is a very robust method that is not prone to over fitting due to the
fact that the hyper plane cannot bend.

Feature generation
The part that most determines how well these kinds of algorithms perform is
what features they have access to. The features are what the algorithms use to
make their predictions. The raw data is the received signal from a microphone, in
a two second window there are a lot of samples. To make the data more workable,
a pre-processing step is performed. To reduce dimensionality, feature extraction
is performed on the two second sample window. The features take the samples
in the window into some mapping to make a new value, this results in a lower
dimensionality. Features has been constructed using the principles described in
[20].

There are two kinds of features explored in this thesis. First kind is the feature
that takes some common mapping of the raw data and create their features from
this mapping. This mapping usually highlights some aspect of the data aiding
feature generation. The other kind of features are generated from the raw data
directly and tries to describe some aspect or characteristic of the data.

Described in this first list are the features that are generated from some common
mapping.
• Standard Deviation [17]. The feature is the standard deviation over the
entire window. The idea is that the feature can estimate how noisy the
window is.
• Mel Frequency Cepstral Coefficients, MFCC, [22]. MFCC measures how
the sound changes over time. These features should have similar values for
similar sounds making it possible to distinguish gun shot sounds.
• Spectrogram. A lot of different features are generated from the spectro-
gram. Some are which frequencies has the most energy and the highest
energy peak. There is also a features that describes if the low frequencies
had their energy peak at the same time, the idea is that a shot has energy in
all frequencies and if they occur at the same time it might indicate a shot.
These are also features describing how intense the energy peak is compared
to its surroundings.
• Short time Fourier transform, stft. The stft has a lot of similarities to the
spectrogram as such the same features are extracted from it.
20 3 Detection

• Zero crossing and zero crossing rate [29]. These features show how often
the signal changes sign. The idea with these features are that gun shots
ought to have similar values.

• Principal Component Analysis [4]. This measures what is most important

in the signal. Similar sounds, such as gunshots, should have the similar
values.

In this second list are features that are generated from the raw data

• Max difference. These features takes the differences between all points in
the window that are separated by some number of steps. Different features
have a different number of steps. The largest difference for each step size
becomes the feature. The idea behind this feature is that when a shot occurs
there will be a large and sudden change.

• Piecewise standard deviation. The window is divided into parts and the
standard deviation is calculated for each of these parts. The resulting values
are then used as input to calculate the standard deviation, the mean and the
maximum.

• Piecewise mean. These features are similar to Piecewise standard deviation

but for each of the smaller parts the mean is calculated instead of the stan-
dard deviation.

• Peak time. These features takes the points with the larges value and checks
the surrounding points if they also have a high value. The idea here is that
shot sounds are intense and dissipate quickly.

In total 62 features were generated from each window. The new feature space is
significantly smaller than the raw input from the microphone but calculating all
these features takes more than two seconds. This means that, if the system should
run in real time, not all of these features could be used. Further dimensionality
reduction was therefore necessary.

Feature selection

Once the new features have been generated, a sequential forward selection (sfs)
scheme is used to determine which features were the best [15]. sfs is used be-
cause it is assumed that the best feature space does not contain a lot of features,
which allows sfs to perform well. An illustration of how sfs works is shown in
Figure 3.4. How well suited a combination of features in a model are for shot
classification is determined by sfs in value called F1 score, that is described in
section 3.3.
3.2 Method 21

Figure 3.4: Example of how sfs iterates, the accuracy of the features can be
viewed in the rightmost box.

To determine which feature based model structure is best for the task, all of them
are tested. For each model, sfs is performed to determine the features to each
model. It was decided to perform seven iterations of the sfs for each model.

3.2.2 Neural Net

The neural net was built using an online resource called Edge Impulse. It can
build a neural net when provided with data. Once the model is built code can
be generated, it can convert the code to an Arduino library code, which is more
compatible to run on a micro processor [30].

Neural networks are more capable of handling large feature spaces. when the
spectrogram feature was used for the regular machine learning models it had to
be further processed to reduce dimensionality. The neural net model takes the
entire spectrogram as its input. This is possible since the layers in the neural
network reduce the dimensionality automatically. Since the dimensionality does
not have to be reduced no information is lost, as a result all the information the
neural network needs is in the spectrogram and it does not require any further
input [23].

The spectrogram is treated like an image, therefore an image classification net-

work is constructed. The implementation does not use standard image classifi-
cation methods such as residual layers or batch normalization, because the final
model can not use a lot of processing power as it has to run on a micro processor.
Without a large network many standard ways of increasing model performance
22 3 Detection

cannot be implemented as they require to many layers or they are not necessary
since there are not so many layers.

3.3 Evaluating machine learning models

When evaluating machine learning models the confusion matrix is very useful.
A confusion matrix shows the truth and what the model predicted [27]. Each
column shows what the model predicted and each row shows what the true label
was. With two classes there is often one that of interest while the class is just
everything else, like in this thesis where gun shot sounds are are being separated
from all other sound. The class that is being separated is usually called the pos-
itive class and the other the negative class. The positive class is usually shown
first in each row and column. The structure of a confusion matrix will be like in
table 3.1.

Predicted class 1 Predicted class 2

True class 1 True Positive (TP) False Negative (FN)
True class 2 False Positive (FP) True Negative (TN)
Table 3.1: Confusion matrix

When the positive class is labeled correctly it is called true positive and when
labeled incorrectly it is called false negative, same for the negative class.

From the confusion matrix some useful matrices can be calculated [24]. These
are shown in the following list.

• Accuracy, shows how well the model is at separating the classes. Accuracy
T P +T N
is calculated as T P +FN +FP +T N .

• Precision, tells how correct the model is when predicting a specific class. To
calculate precision the following equation is used T PT+FP
P
.
• Recall, shows how many data points of a class the model predicted correctly
of how many were expected. The following equation is used to calculate
recall T PT+FN
P
.
• F-score or F1-score, is an attempt to combine precision and recall into one
precision · recall
metric. F-score is calculated by 2 · precision+recall .

An example of a confusion matrix can be viewed in table 3.2.

Predicted class 1 Predicted class 2

True class 1 16 4
True class 2 1 9
Table 3.2: Example of confusion matrix
3.3 Evaluating machine learning models 23

Row one shows the true labels of class 1, the positive class. There are a total of
16 + 4 = 20 points in class 1. Of the 20 points in class 1, 16 were labeled as class
1 which is correct, i.e. TP. Similarly class 2 contains 10 points, of which 9 were
labeled correctly, i.e. TN. The metrics for this class can be viewed in the table 3.3.

Precision Recall F-score Accuracy

Class 1 16 16
2 0.94 · 0.8
16+1 = 0.94 16+4 = 0.8 0.94+0.8 = 0.86
Class 2 9 9
2 0.69 · 0.9
9+4 = 0.69 9+1 = 0.9 0.69+0.9 = 0.78
16+9
16+4+1+9 = 0.833
Table 3.3: example of model quality

3.3.1 Unbalanced data sets effect on model precision

An unbalanced data set has the effect that the majority class precision increases
while the minority class precision decreases. To illustrate this imagine a model
with 90% accuracy for each class. Giving the model an even testing data set con-
taining 10 class A points and 10 class B points results in the following confusion
matrix.

Predicted class A Predicted class B

True class A 9 1
True class B 1 9
Table 3.4: Example of a confusion matrix on a balanced data set

This results in 90% precision and recall for both classes. If the testing data set
given to the model is instead 10 class A points and 100 class B points the follow-
ing matrix would emerge:

Predicted class A Predicted class B

True class A 9 1
True class B 10 90
Table 3.5: Example of a confusion matrix on an unbalanced data set

Now, the same model has achieved a class A precision of 47% and a class B pre-
cision of 99%. The minority class precision has been decreased and the majority
class precision has increased even though the model has not changed. Note that
both still have a recall of 90%.
24 3 Detection

3.4 Results
The training and validation data available to the models were roughly two hours
split equally between shot and non-shot sounds. For the testing data set more
than five hours were available but only 7 minutes were shot sounds. The confu-
sion matrices for the different model can be viewed in Figure 3.5.

(a) Confusion matrix for (b) Confusion matrix for the

the decision tree model K-nearest neighbour model

(c) Confusion matrix for (d) Confusion matrix for

the naive Bayes model the random forest model

(e) Confusion matrix for the support

vector machine model

Figure 3.5: Confusion matrices of the feature based models.

3.4 Results 25

The neural network was also trained on the same data and achieved the results
seen in 3.6.

Predicted shot Predicted sound

True shot 130 99
True sound 1434 8231
Table 3.6: Confusion matrix for the neural network model

All the performance metrics are tabulated in Table 3.7.

Precision Recall F-score Accuracy

Decision Tree Gun shot 10.2% 68.1% 17.7%
Sound 99.1% 85.7% 91.9%
85.3%
K-Nearest Neighbour Gun shot 8.2% 58.5% 14.4%
Sound 98.8% 84.5% 91.1%
83.9%
Naive Bayes Gun shot 17.0% 52.0% 25.6%
Sound 98.8% 93.9% 96.3%
93.0%
Random Forest Gun shot 9.5% 69.9% 16.7%
Sound 99.1% 84.3% 91.1%
83.9%
Support Vector Machine Gun shot 12.5% 73.4% 21.4%
Sound 99.3% 87.9% 93.3%
87.5%
Neural Net Gun shot 8.3% 56.8% 14.5%
Sound 98.8% 85.2% 91.5%
84.5%
Table 3.7: All the performance matrix for the different models

The feature based models all had an sfs performed on them. The features that
best described the data can be viewed in Table 3.8
26 3 Detection

Decision Tree 3 MFCC

1 zero crossing
1 picewise mean
1 picewise std
1 peak time
K-Nearest Neigbour 5 MFCC
1 peak time
1 max differance
Naive Bayes 6 MFCC
1 peak time
Random Forest 5 MFCC
1 peak time
1 max difference
Support Vector Machine 6 MFCC
1 picewise std
Table 3.8: Features chosen by the sfs for the different models

3.5 Discussion
The testing data set is unbalanced, and this is that to better mimic the real world
where background noise is orders of magnitude more common than sound of gun-
fire. There were 229 files containing shots and 9665 that did not, in other words
there is a lot more negative data than positive in the test set. Even this uneven
distribution is rather balanced compared to real world data. The distribution is
such that there is 42 shots for every hour of background noise which is a lot more
then there would be in reality, where one shooting a year would be a lot. The im-
pact of the unbalanced data set is that all gunshot precision is lower and sound
precision is higher, but recall is unaffected. The unbalanced data set also has the
effect of making sound recall and accuracy highly correlated.

The naive Bayes model was the model that had the best overall accuracy. It had
the best precision on shots, because it had a high sound recall, and it had the
worst recall for shots. Since the test data set is very unbalanced this means that
to achieve a good overall accuracy the recall on the negative class is the most im-
portant thing. However, of the models tested it is still believed to be the best. If a
system gives false alarms excessively often it will stop getting used even if it had a
100% recall for positive data. A system that only has 50% recall for positive data
but with very low false alarm rate would be better, since every warning would
be taken seriously and acted on. If one assumes that a poacher fires more than
one shot even low recall gets a high chance of detecting at least one. For example
with two shots the chance of catching at least one becomes 77% even with just
a 52% recall. Even so, the naive Bayes model which had the lowest false alarm
rate still would give almost 4000 false alarms each day which is far to many to be
useful.
3.5 Discussion 27

The svm model stands out from the rest with a very good gun shot precision as
well as the highest shot recall and the second highest accuracy. It is surprising
that naive Bayes model was the model with the highest accuracy since the model
is generally outclassed by the other models. It only achieved this because it heav-
ily favored sound classification at the expense of shot classification, shown by the
fact that it has the worst shot recall of any model. Even so, low false positive is
beneficial for this task, making naive Bayes the best. With a new set of features
or better features it is plausible that svm would still have among the highest ac-
curacy while naive Bayes would not.

To improve the performance of the system the first step is always more data. For
this project there was plenty of negative data but more positive data would im-
prove the models. Another way of improving the performance is to construct
better features. This is a hard step as it is difficult to predict what kind of fea-
tures will be good, as they need to represent some aspect of the data well.

The model with the lowest false positive rate, naive Bayes, still had to many false
positives to be useful. When evaluating model performance in the sfs shot F-
score was used. To reduce false positives, another metric like shot precision or
sound recall might be better, since this would create models that priorities keep-
ing the false positive rate low. This might however have made the models worse
in some other metric. The sfs might also be further improved with some tuning.
Currently, it performs seven iterations for all models, and increasing this number
may improve the quality of the features selected. sfs is performed because there
is not enough time to calculate all the features in real time. To improve the sfs
the number of iterations performed could depend on the computing demand of
the chosen features, with another feature only being chosen if doing so would
not exceed the computing power available. Another approach is to only perform
another iteration if the previous iteration yielded a large enough improvement,
i.e. having a stopping criterion like an optimization algorithm. The final feature
environment would need to be verified that it did not require more computing
power than were available. A the very least, the number of iterations performed
ought not be static.

Another way of reducing false positives is to change the classification threshold,

giving the model a larger bias for one class. Giving a model a larger bias for the
sound class would make it chose the sound class more often. However, doing this
as a rule reduces the overall accuracy and since the shot recall is not great for any
of the models it might not be a good trade off. It is probably better to improve
the data, change the features or tweak the sfs.

The device is intended to be installed in Kenya to catch poachers, currently all

the positive data is gathered in Sweden further more gathered in Swedish forests
and dense pine forests are not a feature of the Kenyan savannah. The assumption
has been made that the only impact from the forest is to make the shot appear
28 3 Detection

further away, however this assumption could be false. It could be the case that
the sound of a shot traveling through forest significantly alters the characteristic
of the shot. Because of this, the device should be tested on positive data from the
savannah or savannah like environment prior to deployment.

Something observed while collection data was that at the furthest ranges the hu-
man ear can quite clearly hear the gun shots. However, in the data it was hard or
sometimes impossible to distinguish the gun shot from the background or even
to hear if the sound was present at all. If a microphone with the ability to record
lower amplitude sounds, it could increase the range of the detection. Such a
microphone was not chosen for this project as it had to be physically small and
cheap to allow for large scale deployment.

As stated in section 3.1.1 the furthest ranges the a gun shot was observed was esti-
mated to be 1.5km through forest. For this thesis we have assumed that the only
impact is that the sound of a shot traveling through forest makes the shot appear
to be further away. If this assumption holds a shot should be able to be identified
at longer ranges on the savannah, we would expect at least double. When collect-
ing the data we noticed that the sound level dropped sharply when just a small
grove of trees was placed in between the shooter and the device. The caliber of
the weapons used will also affect the range at which they can be distinguished.
At Kvarn we did not know which type of weapon was fired but a clear difference
in sound level could be observed between some on the guns. It is most certainly
the case that at the furthest ranges only the most loud guns were audible. The
ranges for the less audible guns are probably lower than 1.5km but we have no
way of knowing which these guns are or what the ranges are.

Something that is worrying is that at the furthest distances the direction of the
incoming sound seemed to matter a lot in regards to how distinguishable the shot
was. It could be the case that the microphone in the shots direction could quite
clearly hear the shot while the microphone at the other end of the device could
not hear the shot at all. In our current implementation the detection algorithm
only detects on one microphone to save processing time, this means that if we
manage to get a detection range of 3 km it will not be 3 km in all directions. A
solution is to run detection on alternating microphones, this will not give a bias
toward a direction but it could still mean that the device could miss a fired shot
within range. Another solution is to make a weaker detection algorithm that can
run faster, enabling two detection algorithms to run in parallel on different mi-
crophones. This could result in greater coverage area depending on how much
worse the algorithms has to be. Increasing the processing power could also be a
solution.

A limitation of our testing set is that it does not contain shots occurring at the
same time or very close to one another as the data is all gathered from a single
man shooting, and not firing full automatic the shots always appear one at a time.
In the training set we have plenty of data where shots are appearing at the same
3.5 Discussion 29

time or very close to one another making one set of two second contain multiple
shots. This fact is not something that increases the accuracy of the models as the
files which contain multiple shots in them are easier for the model to detect. The
accuracy ought to be a bit better if such data was also present in the testing set.

There was a high presence of wind in the negative testing data that was not
present in the rest of the data, this is probably a factor to why the results are
not better. The station that produced all the negative data was placed in precari-
ous situation with meant a lot more wind, if more wind would be present in the
rest of the sets it could aid detection. The reason for this station being chosen
to produce test data and not training data was because it was the single device
that had produced the most data, if it had not been so the test set would be sig-
nificantly smaller. Another way to mitigate this is to place some cover on the
microphones, such a cover was planed for but it was not implemented.

Our data has no meta data of distance to shooter. Rough distances in a set are
known but any individual shot there is no estimation for. It is probably the case
that it is easier to detect shots that are closer as there is more energy present and
the sound of the shot is not overpowered by other sounds in the surroundings.
We can not know how well the models performs at different distances only how
well it performs for all of them. Before deployment some data with good meta
data should be gathered to see for which ranges the models can detect and with
what performance.

We were not able to collect any data that contained a shock blast as this requires
you to be near the bullets path, which we could not do safely. We do not believe
that this effected the project much as poachers are close to the animals when they
shoot, for a shock wave to be observed the bullet must miss and travel in the di-
rection of the deu. Even if they are not close to the animal they have to shoot
over the deu for a shock wave to be observed. This fact ought not effect the per-
formance too much as, even when it is present, the shock wave is usually not as
audible as the sound of the muzzle blast. Furthermore, a shock wave will not
always be present in the practice either and a detection model should be able to
detect a shot without the shock wave being present.

Whenever a shot detection is made an alert along with the sound file the was
flagged as containing gun shot should be sent to the rangers. The reason to send
the sound file is so that they can listen to it and determine if it is a false alarm
or if some action should be taken. For a human it is easier to determine if it is a
real gunshot if the file is a bit longer so that they can hear what is going on. The
actual sound of the shot happens quickly and as such a short file could be better
for the detection, however too short will hinder human verfication. It is possi-
ble that a shorter file might be easier for the detection algorithm to classify, this
would need to be investigated. If the optimal file length is to short for a human
to interpret, more audio around the detection could also be sent. There was not
enough time in this thesis to investigate this.
30 3 Detection

When inspecting Table 3.8 we can see that some of the MFCC features are always
chosen and the none of the spectrogram or STFT features are chosen. The three
methods are similar so it makes sense that only the best would be chosen, but
it is interesting that the MFCC was chosen so often while the other were not. It
might be the case that the feature extracted from the spectrogram or the STFT
are poorly designed and needs to be looked over. Furthermore we can see that a
peak time feature was often chosen, which is understandable as it describes the
largest peak in a file, which should be very distinct for a gunshot. If the loudest
sound in a file is not a shot, a peak time feature could help a model identify that.
 DOA Estimation
4
Once a gunshot has been detected the direction of arrival needs to be calculated.
Conventionally, doa estimation is achieved by measuring the slight difference
in arrival time of a signal to an array of measurement devices. These are called
time-difference of arrival (tdoa) measurements and give an estimate by solving a
non-linear least squares problem (nls) problem. Beamforming methods can also
be used for doa since the phase of the incoming signal means that frequency do-
main analysis is also possible, assuming a coherent wave[11, 31]. These methods
require a precise and relatively large spacing in the array to create differentiable
signals. For a tdoa method the distance between the sensors need to be at least
half the wave length and are therefore unsuited for a small device that needs to
detect a broad range of frequencies. In this thesis a new method of doa estima-
tion is explored that only needs the received power to make an estimate. The
chapter will first explain the new method used for doa estimation, the results
for a number of arrays and a discussion of the results.

4.1 Method
Each device performs its own doa estimation using received power of the mi-
crophones with a method developed by Gustav Zetterqvist, Fredrik Gustafsson
and Gustaf Hendeby[31]. The method consists of a training phase to compensate
for the different gains of the microphones, frequency dependency and model the
directional sensitivity. From this directional sensitivity an estimate can be per-
formed. Note that while this method also relies on signal strength it is not to be
confused with more typical received signal strength (rss) estimations that mea-
sure distance.

31
32 4 DOA Estimation

4.1.1 Signal model

The measured signal from microphone i can be expressed as

yi (t) = si (t) + wi (t) (4.1)

where yi is the measured signal from microphone i = 1, 2, . . . S that consists of the

received signal si (t) and measurement noise wi (t) that is assumed to be normally
distributed N (0, σi2 ). The power of the signal from microphone i is measured at
a discrete time denoted l, and can be calculated as

L
1X
Pi = yi (l)2 , (4.2)
L
l=1

where L is the number of samples. Inserting (4.1) into (4.2) means the expression
can be rewritten with three terms
L L L
1X 2 1X 1X 2
Pi = si (l) + 2si (l)wi (l) + wi (l) (4.3)
L L L
l=1 l=1 l=1
| {z } | {z } | {z }
Pis Pisw ei

where Pis is power of the received signal, Pisw is power of the cross-term between
signal and noise, and ei is power of the measurement noise. The number of sam-
ples L is assumed large and as such the cross-term Pisw will approach zero. The
normal distribution of wi (l) means that ei will have a chi-squared distribution
with L degrees of freedom. Again, since the number of samples is large this chi-
squared distribution can be approximated by a normal distribution.

4
L 2 Approx 2 2σi
e i ∼ χ L −
−−−−−−
→ e i ∼ N (σ i , ) (4.4)
σi2 L

The aim is to use this information to get a doa estimate ψ, i.e. the angle to the
object being tracked. Each microphone is assumed to have a directional sensitiv-
ity in its power attenuation, either by design or construction of the array. Since
the microphones are placed close together in the array the absolute level of the
power received is assumed to be equal for all microphones. The absolute level of
power is denoted α and is considered a nuisance parameter to the estimation. Pi
can now instead be written as a function dependent on ψ,

Pi (ψ) = αgi h(ψ, θ i ) + ei (4.5)

where gi is the microphone gain, h(ψ, θi ) is the directional sensitivity of the mi-
crophone that depends on the angle ψ and parameters θi , and ei is the error
described in equation (4.4).
4.1 Method 33

4.1.2 Training
The array is exposed to a signal of wide-band noise from a number of directions.
The angle to this sound source and signal received by each microphone is ob-
served. The power measured by each microphone is then calculated using equa-
tion (4.2). The parameters to the directional sensitivity θ and gain gi is then
estimated by solving the following non-linear optimization problem. In this the-
sis, the parameters to the optimization problem were found using YALMIP [21]
with the FMINCON solver.

minimize V (x)
x
subject to α > 0,
gi > 0,
(4.6)
h(ψi , θ i ) = 1 ∀i = 1, 2, . . . , S,
S
X
gi2 = 1
i=1

where ψi is the angle when the microphone i directly faces the source of the
sound and S is the number of microphones. x contains the optimization variables:
α, {g1 , . . . , gS } and {θ 1 . . . θ S }. The loss function V (x) is defined as

S K
X L X
V (x) = 4
(Pi (ψk ) − (αgi h(ψk , θ i ) + σi2 ))2 . (4.7)
i=1
2σ i k=1

where K is the number of observed directions.

4.1.3 Fourier Series Model

In order to model the directional sensitivity h(ψ, θi ) for each microphone a Fourier
series model
D
X
i i i
h(ψ, θ i ) = θ0 + θd,c cos dψ + θd,s sin dψ (4.8)
d=1

is utilized where D is the order of the FS. This order is determined by using the
Bayesian information criterion (bic) which aims at minimizing the code needed
to store data [19]. !
log(K)
BI C = V (x) 1 + (2D + 1) (4.9)
K

4.1.4 Frequency dependency

Empirical testing shows that frequency content of the signal affects the direc-
tional sensitivity and the gain of the individual microphones. To compensate, a
frequency dependency was introduced to the parameter of the directional sensi-
tivity, θ i (f ), and to the microphone gain g(f ), where f is the frequency content
34 4 DOA Estimation

of the signal. The training phase is altered to include the frequency dependency.
The chosen solution is to use adjacent band-pass filters to separate the wide-band
noise signal into discrete frequency bands over the whole spectrum. Directional
sensitivity for each band-pass and microphone can then be determined by solv-
ing the optimization problem (4.6) in that range of frequencies. To account for
the increased complexity this introduces the bic is slightly redefined to also sum
over the band-pass filters
F !
X log(K)
BI C = V (x) 1 + (2D + 1) (4.10)
K
m=1

where F is the number of filters.

4.1.5 Estimation
Following the training the signal power model for all the S microphones can be
written on vector form as

P̂ (ψ) = αG(f )h(ψ, θ̂(f )) + e (4.11)

where h( · ) contains the directional sensitivity of the microphones, θ̂(f ) contains

all of the estimates θ̂i (f ) and G(f ) is a gain matrix with the estimated microphone
gains ĝi (f ) on the diagonal. For estimation data the frequency content of the
received signal is calculated and used as a gain vector to the band-pass filters of
each microphone. Power of the received signal in the frequency range is denoted
pm for filter m, and the range is [fm − B/2, fm + B/2] where B is the bandwidth of
the filter. The power can then be estimated as
F
X
Pˆ (ψ) = pm G(fm )h(ψ, θ̂(fm )) + e. (4.12)
m=1

Least squares can then be used to calculate the doa estimate of ψ.

2
P Pˆ (ψ)
ψ̂ = arg min − . (4.13)
ψ ∥P ∥ ∥Pˆ (ψ)∥
4.2 Results 35

4.2 Results
Continuous testing with the prototypes was done as they became available. Three
sets of results where produced from three different array setups and calibration
environments.

4.2.1 Known good calibration

To verify the capabilities of the code, calibration data previously collected in an
anechoic chamber by our supervisor Gustav was used. This original array used
eight microphones and to mimic the layout of the deu half where removed in
the analysis, leaving only every second microphone. The data set contained 34
angle data points with 1° increments between 0 and 5 followed by 5° increments
between 5° and 45°, and finally 15° increments for the rest. To determine a good
range for the band pass filters the signal was made into spectrogram to determine
its frequency contents.

Spectrogram of calibration data, channel 0, 0°

10

20 0

-10
Frequency (kHz)

15

-20

10
-30

5 -40

-50
0
1 2 3 4 5 6 7 8 9
Time (s)

Figure 4.1: The spectrogram of the signal from the first microphone at 0°
using the high quality array.

Figure 4.1 shows that a signal of white noise with a cutoff frequency around
10kHz. Based on this result, the signal was divided into bandpasses 200Hz wide
from 200Hz to 8000Hz. The upper limit was chosen so that future calibrations
that used lower sampling rates such as 16kHz could be compared. Otherwise this
would cause aliasing when a lower sampling frequency is used [28]. To determine
the best Fourier model order, the bic defined in Equation 4.10 was used. Model
orders between 1 and 12 were tested as running the YALMIP optimizer on higher
orders took increasingly longer time.
36 4 DOA Estimation

BIC BIC
2500 200

180

2000 160

140

1500 120

100

1000 80

60

500 40

20

0 0
1 2 3 4 5 6 7 8 9 10 11 12 3 4 5 6 7 8 9 10 11 12
Model order Model order

(a) A bar plot of the bic compared to (b) Zoomed version of the previous figure
the model order

Figure 4.2: (a) and (b) show the value of the bic compared to the model order
for the high quality array.

Figure 4.2 shows that the bic continues to decrease with increased model order.
Lower bic is better but takes a long time to find a solution for. Therefore a model
order of 1, 4, 7 and 12 was tested to see if this affected the performance of the
estimation.

After model training was complete, the same white noise audio was fed back into
the model to assess its ability to estimate known sound sources. The true mea-
surements and the models built with different model order are shown in Figure
4.3. The maximum error and standard deviation are presented in Table 4.1, while
the raw angle errors depending on angle can be seen in Figure 4.4.

Model Order 1 4 7 12
Standard deviation [°] 5.4645 1.7425 1.7498 1.7973
Maximum error [°] 10.92000 4.4400 4.8000 5.880
Calculation time [s] 21 46 124 603
Table 4.1: The performance data of the estimation for the different model
orders when using the high quality array.
4.2 Results 37

Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 0.5 60° 8
120° 60°
0.4
0.1 6

150° 30° 150° 0.3 30° 150° 30°

4
0.2
0.05 Mic 1
2 Model 1
0.1
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(a) Band 1 model order 1 (b) Band 7 model order 1 (c) Band 39 model order 1
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 0.5 60° 120° 60°
8
0.4
0.1
6
150° 30° 150° 0.3 30° 150° 30°

0.2 4
0.05 Mic 1
Model 1
0.1 2
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(d) Band 1 model order 4 (e) Band 7 model order 4 (f) Band 39 model order 4
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 0.5 60° 120° 60°
8
0.4
0.1
6
150° 30° 150° 0.3 30° 150° 30°

0.2 4
0.05 Mic 1
Model 1
0.1 2
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(g) Band 1 model order 7 (h) Band 7 model order 7 (i) Band 39 model order 7
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 0.5 60° 120° 60°
8
0.4
0.1
6
150° 30° 150° 0.3 30° 150° 30°

0.2 4
0.05 Mic 1
Model 1
0.1 2
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(j) Band 1 model order 12 (k) Band 7 model order 12 (l) Band 39 model order 12

Figure 4.3: The Fourier series approximation for different model orders and
frequency bands when using the high quality array.
38 4 DOA Estimation

Model angle error Model angle error

12 18

16
10
14
Number of estimates

Number of estimates
8 12

10
6
8

4 6

4
2
2

0 0
-20 -15 -10 -5 0 5 10 15 20 -6 -4 -2 0 2 4
Error(°) Error(°)

(a) Model order 1 (b) Model order 4

Model angle error Model angle error
14 20

18
12
16

10 14
Number of estimates

Number of estimates

12
8
10
6
8

4 6

4
2
2

0 0
-6 -4 -2 0 2 4 -6 -4 -2 0 2 4
Error(°) Error(°)

(c) Model order 7 (d) Model order 12

Figure 4.4: Histogram of the magnitude of the angle errors when using the
high quality array. A normal distribution has been fitted to the histogram.
4.2 Results 39

4.2.2 Prototype 1

This prototype is the same prototype 1 as referred to in Chapter 2, and was cali-
brated with a Bluetooth speaker inside an empty room. The calibration data set
contained 18 data points from angles with 20° increments from 0° to 340°.

Spectrogram of calibration data, mic 1, 0°

10

0
20

-10
Frequency (kHz)

15
-20

10 -30

-40

5
-50

0
1 2 3 4 5 6 7 8 9
Time (s)

Figure 4.5: The spectrogram to the signal recorded by the first microphone
at 0° using prototype 1.

Since the sound file was different than the one used previously there is no cut off
frequency. The received power varies quite significantly based on frequency but
the model is made to account for that. To get consistent and comparable results
the same bandpass intervals of 200Hz wide and a range of 200Hz to 8000Hz were
used.

Again bic was used to determine the best model order. Results can be seen in
Figure 4.6.
40 4 DOA Estimation

BIC 10-10 BIC

3500 1.5

3000

2500
1

2000

1500

0.5
1000

500

0 0
1 2 3 4 5 6 7 8 9 10 11 12 9 10 11 12
Model order Model order

(a) A bar plot of the bic compared to (b) Zoomed version of the previous figure
the model order

Figure 4.6: (a) and (b) show the value of the bic compared to the model order
when using prototype 1.

Doing this analysis quickly revealed a problem with the calibration. Going over
a model order of 8 caused the YALMIP solver to assign NaN values to one of the
θ factors in the directional sensitivity model. Extensive testing with different op-
timization parameters could not resolve the problem and the exact cause for the
error could not be identified. Therefore the maximum possible of 8 as well 1 and
4 where chosen as model orders in the analysis.

Like before the sound used for training was also used to evaluate the estimation
ability. The true measurements and the models built with different model order
are shown in Figure 4.7. The maximum error and standard deviation are pre-
sented in Table 4.2, while the raw angle errors depending on angle can be seen in
Figure 4.8.

Model Order 1 4 8
Standard deviation [°] 23.2460 29.7073 32.2087
Maximum error [°] 41.7200 76.5200 79.4000
Calculation time [s] 15 41 98
Table 4.2: The performance data of the estimation for the different model
orders when using prototype 1.
4.2 Results 41

Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 2 60° 120° 3 60° 10
120° 60°
8
1.5
2
150° 30° 150° 30° 150° 6 30°
1
4 Mic 1
1
0.5 Model 1
2 Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(a) Band 1 model order 1 (b) Band 7 model order 1 (c) Band 39 model order 1
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 2 60° 120° 3 60° 10
120° 60°
8
1.5
2
150° 30° 150° 30° 150° 6 30°
1
4 Mic 1
1
0.5 Model 1
2 Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(d) Band 1 model order 4 (e) Band 7 model order 4 (f) Band 39 model order 4
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 2 60° 120° 3 60° 10
120° 60°
8
1.5
2
150° 30° 150° 30° 150° 6 30°
1
4 Mic 1
1
0.5 Model 1
2 Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(g) Band 1 model order 8 (h) Band 7 model order 8 (i) Band 39 model order 8

Figure 4.7: The Fourier series approximation for different model orders and
frequency bands when using prototype 1.
42 4 DOA Estimation

Model angle error Model angle error

6 7

6
5

5
Number of estimates

Number of estimates
4

4
3
3

2
2

1
1

0 0
-80 -60 -40 -20 0 20 40 60 80 -100 -80 -60 -40 -20 0 20 40 60 80 100
Error(°) Error(°)

(a) Model order 4 (b) Model order 7

Model angle error
7

5
Number of estimates

0
-100 -80 -60 -40 -20 0 20 40 60 80 100
Error(°)

(c) Model order 8

Figure 4.8: Histogram of the magnitude of the angle errors when using pro-
totype 1. A normal distribution has been fitted to the histogram.
4.2 Results 43

4.2.3 Prototype 2

Additional time was scheduled to test the next prototype in an anechoic chamber
to see if this would improve the results. For this purpose, the new and improved
prototype 2 described in 2.1 was used. The calibration data set included 12 data
points from angles with 30° increments from 0° to 330°. The same sound file that
the high quality array was exposed to was used for the calibration.

Spectrogram of calibration data, mic 1, 0°

-10

-15
20
-20

-25
Frequency (kHz)

15
-30

-35
10
-40

-45

5 -50

-55

-60
0
1 2 3 4 5 6 7 8 9
Time (s)

Figure 4.9: The spectrogram to the signal recorded by the first microphone
at 0°.

Like the calibration data for the high quality array, Figure 4.9 shows a signal of
white noise with a cutoff frequency around 10kHz. The signal was again divided
into bandpasses 200Hz wide from 200Hz to 8000Hz to get comparable results.

Again bic was used to determine the best model order. Results can be seen in
Figure 4.10.
44 4 DOA Estimation

BIC 10-10 BIC

10000 1.5

9000

8000

7000
1
6000

5000

4000
0.5
3000

2000

1000

0 0
1 2 3 4 5 6 7 8 9 10 11 12 6 7 8 9 10 11 12
Model order Model order

(a) A bar plot of the bic compared to (b) Zoomed version of the previous figure
the model order

Figure 4.10: (a) and (b) show the value of the bic compared to the model
order

This analysis also had problems when using a high model order for the calibra-
tion. Going over a model order of 5 again caused the YALMIP solver to assign
NaN values in the directional sensitivity model. Performance testing was done
with model order 1, 3, and the maximum possible 5.

As with the previous results the sound used for training was also used to evaluate
the estimation ability. The true measurements and the models built with differ-
ent model order are shown in Figure 4.11. The maximum error and standard
deviation are presented in Table 4.3, while the raw angle errors depending on
angle can be seen in Figure 4.12.

Model Order 1 3 5
Standard deviation [°] 22.8470 7.6448 11.7895
Maximum error [°] 41.0400 15.1200 21.6000
Calculation time [s] 12 26 47
Table 4.3: The performance data of the estimation for the different model
orders when using prototype 2.
4.2 Results 45

Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 60° 120° 60°
0.6 1.5 30

150° 0.4 30° 150° 1 30° 150° 20 30°

Mic 1
0.2 0.5 10 Model 1
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(a) Band 1 model order 1 (b) Band 7 model order 1 (c) Band 39 model order 1
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 60° 120° 60°
0.6 1.5 30

150° 0.4 30° 150° 1 30° 150° 20 30°

Mic 1
0.2 0.5 10 Model 1
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(d) Band 1 model order 3 (e) Band 7 model order 3 (f) Band 39 model order 3
Band numbr: 1 Range: 200 - 400 Hz Band numbr: 7 Range: 1400 - 1600 Hz Band numbr: 39 Range: 7800 - 8000 Hz
90° 90° 90°
120° 60° 120° 60° 120° 60°
0.6 1.5 30

150° 0.4 30° 150° 1 30° 150° 20 30°

Mic 1
0.2 0.5 10 Model 1
Mic 2
Model 2
180° 0 0° 180° 0 0° 180° 0 0°
Mic 3
Model 3
Mic 4
Model 4

210° 330° 210° 330° 210° 330°

240° 300° 240° 300° 240° 300°

270° 270° 270°

(g) Band 1 model order 5 (h) Band 7 model order 5 (i) Band 39 model order 5

Figure 4.11: The Fourier series approximation for different model orders
and frequency bands when using prototype 2.
46 4 DOA Estimation

Model angle error Model angle error

4.5 4.5

4 4

3.5 3.5
Number of estimates

Number of estimates
3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
-80 -60 -40 -20 0 20 40 60 80 -20 -15 -10 -5 0 5 10 15 20 25
Error(°) Error(°)

(a) Model order 1 (b) Model order 3

Model angle error
5

4.5

3.5
Number of estimates

2.5

1.5

0.5

0
-21 -9 3 15
Error(°)

(c) Model order 5

Figure 4.12: Histogram of the magnitude of the angle errors when using
prototype 2. A normal distribution has been fitted to the histogram.
4.2 Results 47

Some live testing with real gunshot was made to ascertain what real world perfor-
mance could be expected. The shooter was around 20 meters from the recording
device and to get the true angle a protractor attached to the device was used. The
angle error histogram can be seen in Figure 4.14 and the performance results can
be seen in Table 4.4. A spectrogram of one of the gunshots can be seen in Figure
4.13.

Spectrogram of calibration data, mic 1, 0°

20

20 10

0
Frequency (kHz)

15
-10

-20
10

-30

5
-40

-50
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time (s)

Figure 4.13: The spectrogram of one of the gunshots used in the estimation.
The second peak is an echo, likely from the embankment that catches the
bullet.

Model Order 1 3 5
Standard deviation [°] 74.9128 69.7265 66.7821
Maximum error [°] 144.6000 94.8000 93.6400
Table 4.4: The performance data of the estimation to real gunshots for dif-
ferent model orders.
48 4 DOA Estimation

Model angle error Model angle error

5 5

4.5 4.5

4 4

3.5 3.5
Number of estimates

Number of estimates
3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
-155 -45 65 -80 0 80
Error(°) Error(°)

(a) Model order 1 (b) Model order 3

Model angle error
5

4.5

3.5
Number of estimates

2.5

1.5

0.5

0
-85 -15 55
Error(°)

(c) Model order 5

Figure 4.14: Histogram of the magnitude of the angle errors to gunshots. A

normal distribution has been fitted to the histogram.
4.3 Discussion 49

4.3 Discussion
4.3.1 Known good data
The calibration data supplied by our supervisor provided a good testing ground
to tune the performance and verify functionality of the code. Further analysis
gave some surprising results. While a low bic indicates that the model is well
optimized in terms of number of parameters and accurate to the testing data
it has some limitations because of how it is currently implemented. Since then
bic is not tested against the complete summarized model used for the actual
estimation but rather on each individual frequency band, it results in a model
that is only optimized to follow these individual frequency bands. Overall the
results show that a low bic does not always correspond to a more accurate model
when the bands are summed up and used for estimation.

4.3.2 Prototype 1
Calibration of prototype 1 proved difficult, both because of the environment
where calibration data collection took place and because of the uneven behav-
ior of the microphones. The recording was made in an empty and echoing room
which caused the sound to bounce a lot and possibly interfere with itself. This
is likely the explanation to the "double bubble" look to the directional sensitivity,
i.e. specific echos that the device pick up way more sound from unexpected di-
rections. The uneven placement of the microphones around the array may also
have affected performance of the estimation since it left some directions with less
overlap. Two of the microphones had a consistent intrinsic lower received power
despite all microphones being of the same make and model as well as being set to
the same gain. Completely removing the weakest microphone from the analysis,
i.e. microphone number 3, would barely affect the estimate in most situations ex-
cept for a few specific angles. This indicates that for the model to work optimally,
the microphones gain should be roughly equal and that echos will adversely af-
fect the estimation performance.

4.3.3 Prototype 2
Calibrating prototype 2 in the anechoic chamber made the model significantly
more accurate, even if the uneven behavior of the microphones leaves a lot to
be desired. With the best model order of 3 and a sound source 1000m away it
would mean a standard deviation of 133m and a worst result of 263m. If this
kind of performance can be replicated in real world conditions it would likely
be sufficient in the purpose of finding a poacher on the savannah. Unfortunately,
when estimating the direction to real gunshots the performance was not as good.
With the best model order for these sounds of 5 and a gunshot 1000m away the
standard deviation would be 1101 a worst result of 1459m. This result would
not give sufficient accuracy to merit implementation on an anti-poaching device.
However, since a doa estimation in it self does not give a range estimate this
inaccuracy may be less important if several estimates are fused. While testing
50 4 DOA Estimation

indicates that the method does indeed work to give estimates in the general direc-
tion of a gunshot, it might be just as accurate as guessing a direction solely based
on what microphone had the highest received power. Gauging the direction to a
gunshot with echoes and other noises that interfere with the signal is inherently
difficult. It is also possible that, while the device was not disassembled between
the gathering of calibration data and the gathering of gunshot data, some of the
microphones may have shifted in their position, thus making the estimate worse.
It should be noted that the gunshot data was collected few days before the cali-
bration, making microphone performance drift an unlikely factor. The true angle
to the shooter is also much more difficult to gauge with only a protractor with a
radius of about 20 cm, something that may have lead to inaccurate angles. The
spectrogram of the gunshot shows that the sound impulse is very short and has a
very wide spectrum, with the highest intensity at low frequencies. Lower frequen-
cies also travel further in the atmosphere, which makes the prospect of making
estimations at longer distances good. However it is also at lower frequencies that
the directional sensitivity is the least prominent.

There is a potential simple explanation to the poor performance however. When

the shot was fired that that close to the device the sound is too loud for the micro-
phones to measure, resulting in a "clipped" signal. This also means that the power
is calculated incorrectly, resulting in the bad measurements. Of course, it is diffi-
cult to ascertain if this was indeed what caused the estimate to be inaccurate and
to what degree.

4.3.4 Solver error

Exactly why the solver fails when the model order is too high for the two proto-
types is unknown. However, a likely culprit may be because of the larger spacing
between the calibration points that causes the solver to find some otherwise im-
possible Fourier series that nonetheless give a lower loss function value than a
more reasonable and smooth model. What makes this more likely is that a higher
increment in angle between the calibration points caused the calibration to fail
earlier.
 Conclusion
5
The estimation method can, in a controlled environment and with a good array,
give excellent angle estimates. The step to translating this for use in practise
where microphones need to be cheap and outside conditions cannot be controlled
will need further work.

Shots can be differentiated from other sound. We found many models that man-
aged to do so all better than a coin flip, and as such if more shots are fired there
is a very good chance that at least one of the shots could be detected. Since the
models were better than a coin flip the features managed to describe the data but
some tuning of the features is necessary for good results. The largest problem
with the solutions is the false positive rate and is what really hinders deployment
of any of the detection models.

The microphones were limited to an audible range of about 1.5km when used in
Swedish forests, even though the human ear could clearly hear gunshots from
that range. An open plain would have the sound travel further.

The device was solid enough to withstand one night of heavy rainfall. However
the microphones are a weak point where dust and water could potentially leak
in if left outside for a long time. A method to shield these from the elements
without compromising on detection and estimation performance is desirable.

5.1 Future work

A lot of the time dedicated to this project has be collecting data of guns being
fired, this does not only include the physical gathering but an equally time con-

51
52 5 Conclusion

suming part has been talking to potential people that could allow us to be present
while they shoot. We have found that military personnel has by far been the most
helpful with this and if more data is to be collected these are the people to turn
to first.

5.1.1 Detection
To improve the detection in the future the first step is always to improve the data
available to the models. Specifically more positive data would be required. There
is plenty of negative data so initially more positive data would be required but
different negative data would also serve the project. The best kind of data to
gather would be sounds of shot gathered in Kenya specifically on the savannah,
as it stands all the data is gathered in Swedish forests or Swedish hilly forest. One
intermediate step is to gather gun shot data in Sweden but that at least is on a
field or at the very least where there is a line of sight to the shooter. Another way
to gather different gun shots would be to gather data where the shock wave from
the bullet is present. To see if it effects the detection. If it is not possible to do any
of these more data from Swedish forests would still improve the performance.

There is a need to verify that shots sounds traveling through forest do not signifi-
cantly alter the characteristic of a shot. If this is the case, all the positive data in
the data set would need to be replaced. The easiest way to test is to gather some
shot data from a field in Sweden or better yet on the savannah in Kenya, then send
this data to a model trained on the existing data set and analyze the performance.
If the performance is significantly worse some it could indicate that some impor-
tant characteristic of the sound is attenuated when going through forest. Another
approach would be to do some kind of principal component analysis and check
if the different shot files are similar.

Even though this thesis finds that the neural net solution performed worse than
feature based methods we still believe it has the most potential. A neural net
solution is easier to scale with more layers, and then do residual layers or batch
normalization. The field of neural net image classification is wast and constantly
growing providing room to improve the model, especially if more processing
power is available. Feature based solutions require the construction of better fea-
tures to be improved. Identifying and implementing better features is time con-
suming and require a lot of knowledge about the specific problem to be solved.
If this route is taken it is expected that naive Bayes would not remain the best
model, but that it would be overtaken by svm. Although testing all approaches
to feature based methods is probably wise.

If for what ever reason a neural network solution is not desired the current mod-
els could be improved with some parameter tuning. Very little tuning has been
performed during this project as having better data and better features will im-
prove the models a lot more and therefore time was placed there. The models
performance could probably be improved if the features were tuned or if hyper
5.1 Future work 53

parameter optimization was performed. These things would not improve the
model greatly but some improvement could probably be gained.

5.1.2 Estimation
While the training phase of the calibration is fairly quick, collecting data takes a
long time. A standardized testing protocol would speed up the calibration data
collection process and allow for more precise models and perhaps models tuned
to different frequency spectrums. Such a protocol should contain information
on how the device should be mounted, how the angles are accurately measured
to the sound source, what sounds should be played and how the recordings are
saved.

In future implementations the microphone array should be constructed so that

the microphones do not need to be moved from their mounting position other
that to replace them. This is because the calibration relies on the properties of
the microphones being constant, and even slight movements to them in relation
to each other will have an adverse effect on the doa estimate.

With how the code works at the moment, it only gives a doa estimate and no
other information. Calculating several estimates gives a general idea of how ac-
curate the model is in general, but it would be desirable to have the doa estimate
to be given along with a confidence interval, i.e. how sure the model is of the es-
timate. For example, let’s say a sound is coming from a true angle of 40° and the
model gives an estimate of 42°. While this is a good estimate from what is known
about the true angle, this gives no information on how confident the model is of
the estimate. A confidence interval of 3° in either direction let the user know that
the estimate is likely to be accurate, whereas a confidence interval of 30° may
mean that the estimate isn’t as useful. Even if, as in this example, the ground
truth angle was not far from the estimate. What the confidence interval is for any
given sound is likely to vary depending on the true angle from which the sound
came and the strength of the sound.

Currently the usage of higher model orders is rather inefficient. As show in the
results in Section 4 higher model order is not necessary for representing lower
frequencies. A dynamic way to assign lower model order to simpler frequency
bands, and higher model order for more complex ones is an excellent way to save
on resources, both for optimisation time and number of parameters. For this bic
could be used, i.e. each band is tested individually with Equation 4.9 for the best
model order and then saves this to a new vector.

5.1.3 Localization
Localization by fusing multiple doa estimates can give the exact location of a
fired shot. It is also theoretically possible to use the received signal strength (rss)
to get a distance estimate together with the doa estimate which would allow for
54 5 Conclusion

location estimation with only a single detection. This future expansion would
make the finding of the poacher much easier and the theory for doing this.

Triangulation

If the devices are close enough, so that multiple devices can hear the sounds, sev-
eral doa estimations can be combined to locate the source. This can be done with
triangulation, an illustration which can be seen in Figure 5.1. Methods for fusing
doa estimates can be found in books such as Statistical Sensor Fusion by Fredrik
Gustafsson [11].

Figure 5.1: How the stations can work together to find the location of the
shooter.

Distance estimation

If no more than one detection is made then triangulation cannot be performed. In

this case, a distance estimate could be made with an rss method. Since all sounds
have a predictable falloff in intensity and together with a rough estimate of how
loud the gun was, received signal strength can be calculated. That together with
the doa estimate will provide a coarse estimate to the position of the gunman.
An illustration can be seen in Figure 5.2.

Some models such as ISO 9613-2 have been created to predict loudness of a sound
a from a distance based upon strength at emission, atmospheric conditions and
5.1 Future work 55

obstacles. This could then used to create a loss function to estimate the distance.
However, doing so would require good knowledge of the current conditions as
well as the type of rifle being discharged. The former can be solved with weather
predictions or weather stations feeding data to the model, while the latter will
either have to be a complete guess or attempting to classify the type of weapon
used during the detection step.

Figure 5.2: How a single station can estimate the position of the shooter
 Bibliography

[1] Africa’s poaching crisis. URL https://campaign.awf.org/

poaching-infographic/.

[2] Poaching numbers | conservation | save the rhino interna-

tional. URL https://www.savetherhino.org/rhino-info/
poaching-stats/.

[3] Safereaction. URL https://safereaction.com/.

[4] Hervé Abdi and Lynne J. Williams. Principal component analysis. WIREs
Computational Statistics, 2(4):433–459, 2010. doi: https://doi.org/10.
1002/wics.101. URL https://wires.onlinelibrary.wiley.com/
doi/abs/10.1002/wics.101.

[5] L. Breiman. Random forests. Mach Learning, pages 5–32, 2001.

[6] R. Emslie. Diceros bicornis, Black Rhino. The IUCN Red List of Threat-
ened Species, 2020. doi: https://dx.doi.org/10.2305/IUCN.UK.2020-1.
RLTS.T6557A152728945.en.

[7] R. Emslie. Ceratotherium simum, White Rhino. The IUCN Red List
of Threatened Species, 2020. doi: https://dx.doi.org/10.2305/IUCN.UK.
2020-1.RLTS.T4185A45813880.en.

[8] Julian Blanc Iain Douglas-Hamilton Patrick Omondif Kenneth P. Burnham

George Wittemyer, Joseph M. Northrupa. Illegal killing for ivory drives
global decline in african elephants. Proceedings of the National Academy of
Sciences, 2014. doi: 10.1073/pnas.1403984111.

[9] Edwards C.T.T Maisels F. Wittemyer G. Balfour-D. Taylor R.D. Gobush,

K.S. Loxodonta cyclotis, African Forest Elephant. The IUCN Red List
of Threatened Species, 2021. doi: https://dx.doi.org/10.2305/IUCN.UK.
2021-1.RLTS.T181007989A204404464.en.

[10] Edwards C.T.T Maisels F. Wittemyer G. Balfour-D. Taylor R.D. Gobush,

K.S. Loxodonta africana, African Savanna Elephant. The IUCN Red List

57
58 Bibliography

of Threatened Species, 2021. doi: https://dx.doi.org/10.2305/IUCN.UK.

2022-2.RLTS.T181008073A223031019.en.
[11] Fredrik Gustafsson. Statistical Sensor Fusion. Studentlitteratur AB, third
edition. ISBN 978-91-44-12724-8.
[12] Fredrik Gustafsson, Gustaf Hendeby, David Lindgren, George Mathai, and
Hans Habberstad. Direction of arrival estimation in sensor arrays using
local series expansion of the received signal. In 2015 18th International
Conference on Information Fusion (Fusion), pages 761–766, 2015.
[13] Michael G. Haag and Lucien C. Haag. Chapter 17 - the sound levels of
gunshots, supersonic bullets, and other impulse sounds. In Michael G.
Haag and Lucien C. Haag, editors, Shooting Incident Reconstruction (Third
Edition), pages 407–444. Academic Press, San Diego, third edition edi-
tion, 2021. ISBN 978-0-12-819397-6. doi: https://doi.org/10.1016/
B978-0-12-819397-6.00017-9. URL https://www.sciencedirect.
com/science/article/pii/B9780128193976000179.
[14] P Hart. The condensed nearest neighbour rule. IEEE Transactions on Infor-
mation Theory, 14:515–516, 1968.
[15] Huan Liu Hiroshi Motoda. Feature selection extraction and construction.
Towards the Foundation of Data Mining Workshop, Sixth Pacific-Asia Con-
ference on Knowledge Discovery and Data Mining (PAKDD 2002), Taipei,
Taiwan, pp. 67–72, 2002.
[16] S. B. Kotsiantis. Supervised machine learning: A review of classification
techniques. Informatica 31 (2007) 249-268, 2007.
[17] Lee Sangseok Lee Dong Kyu, In Junyong. Standard deviation and standard
error of the mean. kja, 68(3):220–223, 2015. doi: 10.4097/kjae.2015.68.3.
220. URL http://www.e-sciencecentral.org/articles/?scid=
1156109.
[18] David Lindgren, Olof Wilsson, Fredrik Gustafsson, and Hans Habberstad.
Shooter localization in wireless microphone networks. EURASIP Journal
on Advances in Signal Processing, 2010(1):1 – 11, 2010. ISSN 1687-
6180. URL https://login.e.bibl.liu.se/login?url=https:
//search.ebscohost.com/login.aspx?direct=true&AuthType=
ip,uid&db=edssjs&AN=edssjs.5FC20B75&lang=sv&site=
eds-live&scope=site.
[19] Lennart Ljung, Torkel Glad, and Anders Hansson. Modeling and Identi-
fication of Dynamic Systems. Studentlitteratur AB, second edition. ISBN
978-91-44-15345-2.
[20] T.Deepa L.Ladha. Feature selection methods and algorithms. International
Journal on Computer Science and Engineering (IJCSE), vol.3(5), pp. 1787-
1797, 2011.
Bibliography 59

[21] J. Löfberg. Yalmip : A toolbox for modeling and optimization in matlab. In

In Proceedings of the CACSD Conference, Taipei, Taiwan, 2004.

[22] Beth Logan. Mel frequency cepstral coefficients for music modeling. In
Proceedings of International Symposium on Music Information Retrieval
(ISMIR), 2000.
[23] Abiodun Esther Omolara Kemi Victoria Dada Nachaat AbdElatif Mohamed
Humaira Arshad Oludare Isaac Abiodun, Aman Jantan. State-of-the-art in
artificial neural network applications: A survey. Heliyon, volume 4, issue
11, 2018.
[24] Radu Mariescu-Istodor Pasi Fränti. Soft precision and recall. Elsevier B.V,
2023.

[25] Julian Blanc Carsten F. Dormann Colin M. Beale Severin Hauenstein,

Mrigesh Kshatriya. African elephant poaching rates correlate with local
poverty, national corruption and global ivory price. Nature Communica-
tions 10 (1), 2019. doi: 10.1038/s41467-019-09993-2.
[26] Small Arms Survey. Small Arms Survey 2015: Weapons and the World. Cam-
bridge University Press, 2015. ISBN 9781107323636.
[27] K.M. Ting. Confusion matrix. Encyclopedia of Machine Learning, 2011.
[28] Lennart Ljung Torkel Glad. Reglerteori - Flervariabla och olinjära metoder.
Studentlitteratur AB, second edition. ISBN 978-91-44-03003-6.

[29] R.W. Wall. Simple methods for detecting zero crossing. 2003.
[30] Jan Jongboom Zach Shelby. Edge impulse, 2019. URL https://docs.
edgeimpulse.comhttps://docs.edgeimpulse.com/docs/.
[31] Gustav Zetterqvist, Fredrik Gustafsson, and Gustaf Hendeby. Using
received power in microphone arrays to estimate direction of arrival.
In ICASSP 2023 - 2023 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pages 1–5, 2023. doi: 10.1109/
ICASSP49357.2023.10097197.