Crashzam: Sound-based Car Crash Detection

Matteo Sammarco

and Marcin Detyniecki

1,2,3

AXA Data Innovation Lab, 48 rue Carnot, 92150, Suresnes, France

Sorbonne Universit

es, UPMC Univ Paris 06, CNRS, LIP6 UMR 7606, 4 place Jussieu, 75005 Paris, France

Polish Academy of Sciences, IBS PAN Newelska, 6, Warsaw 01-447, Poland

Keywords:

Real-time Incident Detection, Road Safety, Smart Vehicle, Audio Signal Processing.

Abstract:

Connected vehicles, combined with embedded smart computation capabilities, will certainly lead to a new

generation of services and opportunities for drivers, car manufacturers, insurance and service companies. One

of the main challenges remaining in this ﬁeld is how to detect key triggering events. One of these crucial

moments is a car accident, for which not only smart connected vehicles can improve drivers’ safety as car

accidents are still one of the main causes of fatalities worldwide, but also help them during minor, but very

stressful moments. In this paper, we present Crashzam which is an innovative way to detect any type car

accidents based on sound produced by car impact, while, so far, crash detection is only a prerogative of

accelerometer sensor time series analysis, or its proxy: activation of the airbag. We describe the system

design, the sound detection model, and the results based on a dataset with in-car cabin sounds of real crashes.

We have beforehand built such dataset with real car accident sounds. Classiﬁcation is built upon features

extracted from the time and frequency domain of the audio signal and from its spectrogram image. Results

show that the proposed model is able to easily identify crash sounds from other sounds reproduced in-car

cabins. Moreover, considering that Crashzam can run on smartphones, it is a low cost and energy solution,

contributing to the spreading of such a car safety feature and reducing delays in providing assistance when an

accident occurs.

1 INTRODUCTION

Although the effort in launching road safety pro-

grams in many countries, road trafﬁc death ﬁgures

remain stable worldwide at nearly 1.2 million since

2004 (World Health Organization, 2015). Causes of

such a plague are diverse: speed, drink-driving, drug-

driving, unused safety belt, bad weather and road con-

ditions, and bad car break and wheel conditions. It

results essential to notify a crash as fast as possible

for ﬁrst aid as a correlation between delaying emer-

gency medical care and mortality rate has been pro-

ved (Evanco, 1996).

Some car manufacturers offer for their high-end

products an automatic collision notiﬁcation which

mainly monitors the airbag deployment to detect a

severe collision and calling assistance with the em-

bedded cellular radios. The BMW’s Automatic Crash

Notiﬁcation System and the GM’s OnStar are just two

examples. These products remain mostly restricted as

option to luxury market sectors and a large part of the

circulating vehicles do not embed an OEM automatic

accident detection and notiﬁcation system.

Relatively cheaper third party solutions foreseen

the installation of boxes under the hood, wind-screen

boxes or OBDII dongles which embed an acceleration

sensor as along as a proprietary algorithm to detect

shocks. In fact, state of the art solutions employ acce-

lerometer data to detect more or less severe impacts

triggered by the sudden variation of acceleration on

one or more axes (White et al., 2011; Thompson et al.,

2010; Zaldivar et al., 2011; Punetha et al., 2012; Lahn

et al., 2015; Aloul et al., 2015). Although the use of

accelerometer sensors leads to a precise impact iden-

tiﬁcation both for angle and severity, dedicated har-

dware must be professionally installed to achieve the

maximum accuracy.

The most cost effective and practical solution,

instead, relies on acceleration time series recorded

on drivers’ smartphones. On these bases, some mo-

bile applications are already available on the mar-

ket (Zendrive, 2017; Sosmartapp, 2017; TrueMotion,

2017). Nevertheless, smartphone data is hard to ana-

lyze due to calibration, noise and rotation issues. In

addition, it is not clear the optimal frequency for acce-

leration samples and the time window width to record.

Sammarco, M. and Detyniecki, M.

Crashzam: Sound-based Car Crash Detection.

DOI: 10.5220/0006629200270035

In Proceedings of the 4th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2018), pages 27-35

ISBN: 978-989-758-293-6

Also, in many situations, relying only on acceleration

data may lead to false predictions: street bumps, holes

and bad street conditions trigger false positives, whe-

reas collisions coming from the back while standing

still may be classiﬁed as normal accelerations.

As a trendy solution, social networks and micro-

blogs provide a global source where to share what we

experience and see around us. Car accidents usually

attract people curiosity, who might post tweets and

photos of the event (Schulz et al., 2013). Although

this provides a zero-cost solution, many cons affect

it: accidents could not be immediately advertised, pe-

ople might provide confused or misleading informa-

tion, necessary third party people identiﬁcation comes

with privacy issues. Finally, the use of smartphones

during driving is a source of accident itself.

As electric lamps were not invented by impro-

ving candles, the goal of this work is not to im-

prove accelerometer-based car crashing detection al-

gorithms, but to detect crashes by sound.

Crashzam is an innovative way to detect car cras-

hes based on sound recognition techniques. It does

not suffer from neither calibration and sensor conﬁ-

guration problems nor the delay and subjectiveness

of human notiﬁcations. Being in continuous listening

speciﬁcally for crash sounds, our solution is able to

detect whatever a car impact occurred. It can take ad-

vantage of the presence of microphones inside car ca-

bins: hands free car kits, Bluetooth kits, car audio sy-

stems with voice command, wind-screen SOS boxes,

dash cameras and smartphones are some hardware

equipments embedding microphones, which are usu-

ally present in a car.

It does not matter where and with which angle re-

spect to the horizon these devices are placed in the car

cabin, since the received sound will not be affected by

calibration issues. The only assumption is that a crash

produces a sound.

Despite some works based on crash sound de-

tection which are mainly designed for road surveil-

lance purposes (Rabaoui et al., 2008; Carletti et al.,

2013; Foggia et al., 2016a; Clavel and Ehrette, 2008;

Valenzise et al., 2007), this work focuses on drivers’

safety by detecting crash sound from inside the car

cabin.

But in order to perform, Crashzam must take into

account the presence of environmental noise and any

other overlapping sound which could happen in a car

cabin (engine rotation, car horn, radio music, etc.).

The advantage for drivers is to have a low cost in-

stantaneous crash detection rid of accelerometer sen-

sor discrepancies.Other actors involved in road safety,

like medical assistance services and insurance compa-

nies, can take advantage of this solution for ﬁrst noti-

: acceleration

event

: start

recording

PCM buffer

x+Tb/2

: data compression

and transmission

Crash

detected?

Time and

frequency analysis

Spectrogram

image analysis

Alert driver

and assistance

Yes

Figure 1: Crashzam high level system design: while dri-

ving, the smartphone microphone records a sliding buffer

of PCM samples which are analyzed if an acceleration event

occurs (e.g., acceleration norm is over a certain threshold).

Accelerometer is used as a mere trigger. If Crashzam de-

tects a crash, the driver is solicited and assistance is alerted

in case of no answer from the driver.

ﬁcation of loss (FNOL) and fast ﬁrst aid.

The main contribution of this paper is the design

of a low cost, smartphone-based car crash detection

system through a new crash sound detection model. It

can detect crash sounds and distinguish other sounds

generated inside the car cabin. To this aim, we col-

lect a novel dataset of crash sounds recorded from

inside the car cabin. Then, we create and select a

set of features computed from the time and frequency

sound signal domain and from the sound spectrogram

image. Finally we propose a combination of ma-

chine learning models for the automatic classiﬁcation

of sounds reproduced inside vehicles.

This manuscript is structured as follows: in

Section 2 we provide a top-down description of

Crashzam, while in Sections 3 and 4 we explore in

detail the model to detect crashes from audio clips.

In Section 5 we present the dataset that we have built

and that we have used to train and test our model. In

Sections 6 and 7 we discuss about the obtained re-

sults and the performances in according to the model

parameters. Finally, we give our conclusions and per-

spectives in Section 8.

2 SYSTEM DESIGN

Considering Crashzam running on a smartphone ap-

plication, we show in Figure 1 the high level sy-

stem design. Android and iOS operating systems pro-

vide an activity recognition mechanism which starts

at time t

the recording of a PCM samples buffer si-

multaneously to the detection of driving activity. The

sliding buffer has a duration of T

seconds. In our

proof of concept we set T

= 5 seconds, with PCM

sampled at 16kHz and quantized at 16 bits at least. At

, also the location service (GPS) and the accelero-

meter sensor are activated. We use the accelerometer

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

as a mere trigger: if at time t

the norm of the three

acceleration axes exceeds 2.5g, then we continue re-

cording for T

/2 seconds before analyzing data. In

this way we use half of the buffer to record what hap-

pened before a probable impact event and the other

half to record what happened after it. In fact, a typi-

cal crash event includes tire skidding or horn sounds

before the crash and human screams before or after it,

although the crash itself will present the maximum

signal amplitude. Thus, in the following we train

our models to distinguish between car impact sounds

which are high energy, percussive sounds and other

sounds which are likely to be produced during a crash

but not necessarily like car horn or harsh deceleration

sounds. It is worth to note that Crashzam can also

work in combination with a traditional accelerometer-

based car crash detection.

At time t

+ T

/2, the array of PCM is compres-

sed and transmitted to the server as along as the last

GPS position, as a JSON object to a server dedicated

to analyze the recorded sound. Data is independently

analyzed by two models detailed in Sections 3 and 4.

Finally, both results are combined by a weighted vo-

ting classiﬁer. The ﬁrst model focuses on time and

frequency aspects of the sound. It mainly detects high

energy, abrupt changing sounds. But since car horn or

engine starting are such kind of sound too, we include

a second model based on spectrogram image analysis

calibrated on the detection of percussive, high energy,

hollow sounds, which correspond to crashes.

A side effect feature of having two independent

models, is that the system is modular and the ﬁrst

model can be reused in other application ﬁelds (e.g.,

glass breaking detection).

The result of the classiﬁcation

y is based on the

probability output of the two models j ∈ 1, 2 in accor-

ding to the relation:

y = argmax

∑

j=1

i j

, (1)

where i ∈ 0,1 represents the sound class (“Other” or

“Crash” sound).

At the end of the classiﬁcation pipeline, if a crash

is detected, the driver is solicited for an interaction

(e.g., pushing a safe button or answering to a call),

otherwise assistance is alerted.

We have monitored the battery usage on a Nexus

5 smartphone when activating motion sensors (acce-

lerometer, gyroscope, and magnetometer), and recor-

ding audio signals. Nexus 5 is equipped with a 2300

mAh, 3.8 V battery and an Android Nougat OS. Fi-

gure 2 shows that the natural discharging from 100%

to 5% lasts four hours when all not vital service but

the screen are switched off. Activating motion sen-

Figure 2: Nexus 5 battery discharging when activating or

not motion sensors and microphone.

sors only, battery life is 30 minutes shorter, while ad-

ding audio recording and motion sensors it becomes

50 minutes shorter. Knowing that the battery used for

the test can provide at most 2300 mA in an hour, we

conclude that the microphone consumed about 80 mA

per hour which corresponds to 3.5% of the battery ca-

pacity.

3 FIRST MODEL: TIME AND

FREQUENCY ANALYSIS

A large set of metrics and features can be extracted

from audio signals (Peeters and Rodet, 2004). For

the ﬁrst model, we have selected here some time and

frequency-based features largely adopted in the litera-

ture for tasks of audio event detection or music genre

classiﬁcation.

3.1 Time Domain Features

Let us consider x(t) a discrete audio signal of N sam-

ples.

• Zero-Crossing Rate (ZCR). The rate a dis-

crete signal x(t) changes sign during its dura-

tion N is a key feature to recognize percussive

sounds (Gouyon et al., 2000b; Gouyon et al.,

2000a).

ZCR

N−2

∑

t=0

|sgn(x(t + 1)) −sgn(x(t))|, (2)

where sgn(x(t)) =

(

1, if x(t) ≥ 0,

−1, if x(t) < 0.

• Signal Power. It is the sum of squares of the sig-

nal values normalized by the signal length.

N−1

∑

t=0

|x(t)|

(3)

We expect crash sounds to show high power.

Crashzam: Sound-based Car Crash Detection

• Entropy. Entropy of a discrete random variable

X with possible values x

,...,x

and probability

function P(X) is usually deﬁned as:

H(x) = −

∑

i=1

P(x

)log

P(x

). (4)

We consider P(x

) = e

f rame

, where E

∑

N−1

t=0

|x(t)|

is the signal energy and E

f rame

is the

energy of the j

of K ﬁx-sized sub-frames the sig-

nal is split into. Thus, our entropy becomes:

H(x) = −

∑

j=1

log

). (5)

Entropy is usually interpreted as a measure of

abrupt changes in energy (Pikrakis et al., 2008;

Giannakopoulos et al., 2007). We expect crash

sounds having high entropy.

3.2 Frequency Domain Features

It is often useful to analyze discrete signals in the fre-

quency domain through a Discrete Fourier Transform

(DFT). The original signal is split in ﬁxed-size smal-

ler frames, and the DFT is applied on each frame re-

turning an array of coefﬁcients having the same length

of the number of samples in the frame. Let us consi-

der X

(k), k = 1, ...,M, the magnitude of DFT coefﬁ-

cients of the i

frame. We compute average and stan-

dard deviation over all frames.

• Spectral Centroid (SC). For the i

frame, SC is

the average of frequencies present in the signal,

weighted by their amplitudes:

∑

k=1

(k)

∑

k=1

(k)

, (6)

The SC represents the barycenter of the spectrum

and higher values correspond to brighter

sounds (Grey and Gordon, 1978). Crash sounds

have a low SC.

• Spectral Spread (SS). It represents the deviation

from the SC:

∑

k=1

(k −SC

)

(k)

∑

k=1

(k)

(7)

Low values of SS correspond to signals whose

spectrum is concentrated around the spectral cen-

troid. Usually crash sounds present a high SS va-

lue.

• Spectral Flux (SF). SF represents the spectral

change by comparing the power spectrum of two

consecutive frames:

i,i−1

∑

k=2

(

(k)

∑

l=1

(l)

−

i−1

(k)

∑

l=1

i−1

(l)

)

(8)

SF is mainly used for onset detection, thus appli-

cable to crash detection too (Dixon, 2006).

• Spectral Rolloff (SR). The S R is the frequency

below which 90% of the magnitude distribution of

the spectrum is concentrated. SR is the frequency

which satisﬁes the following relation:

∑

k=1

(k) = 0.9

∑

k=1

(k). (9)

It is useful to discriminate sounds like human

voice signals whose energy is concentrated under

4 kHz and music.

• Spectral Entropy (SE). Similarly to the entropy

in the time domain, let us consider the spectrum

divided in k ﬁxed-size frequency sub-bands, SE

is:

SE = −

∑

i=1

log

), (10)

where P

∑

i=1

and E

is the energy in the i

sub-band. In crash sounds, SE should have low

values as the energy is spread on all the sub-bands.

• Chroma Vector (CV). With the Chroma Vector

we group all the DFT coefﬁcients into 12 bins cor-

responding to the 12 pitches of an equal tempered

scale. Each element of the vector is the mean of

DFT coefﬁcients:

∑

f ∈F

( f )

, k ∈ [1 − 12], (11)

where F

is the set of frequencies included in

the same bin. CV is widely used for audio ma-

tching (Kurth and M

uller, 2008; Mueller et al.,

2005).

• Mel-Frequency Cepstral Coefﬁcients

(MFCCs). MFCCs are 13 coefﬁcients forming

a cepstral representation where the frequencies

are distributed according to the mel scale. The

mapping between the mel and frequency scale

using k triangular overlapping windows is the

following:

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

Mel( f ) = 2595 log

(1 +

100

). (12)

Then, MFCCs are computed as the discrete cosine

transform (DCT) of cepstrum powers at each mel

frequency:

c(k) = DCFlog |DFT m(n)|. (13)

MFCC are useful to analyze abrupt changes in

the spectrum and widely employed in human

speech domain where they are particularly ef-

fective (Gonzalez, 2013; Sengupta et al., 2016;

Ganchev et al., 2005; M

uller, 2007).

4 SECOND MODEL:

SPECTROGRAM IMAGE

ANALYSIS

The second model is based on spectrogram image fea-

tures and it is speciﬁcally designed to discern between

percussive and sustained sounds. The former have

high amplitude values distributed on all the frequen-

cies at a certain time, while the latter present high va-

lues of amplitude at certain frequencies for long time.

The intuition behind this model is that a crash is a per-

cussive sound, while many other sounds reproduced

in car like car horn, harsh acceleration or tire skidding

are sustained sounds.

4.1 Speciﬁc Spectrogram Image

Features

Starting from a spectrogram images like the ones in

Figure 3, we extract the amplitude matrix and we se-

lect a constellation of peaks, which are points, located

in time and frequency, exceeding a certain amplitude

threshold A

. Such peaks are local maxima, meaning

that in a region with radius R = 2 ∗ D + 1, where D

is the maximum Manhattan distance from the center

of the region, they show the maximum value. If more

points in a region are candidate to be a peak (i.e., they

have the same amplitude), all of them are selected.

Thus, we convert the amplitude matrix to a matrix P

which localizes a constellation of peaks p

f t

and we

analyze the patterns created by such peaks. P is a

F × T binary, sparse matrix, where f ∈ F denotes a

frequency in the range [0,

sampling rate

] and t ∈ T de-

notes a small time bin the original signal is split into

to compute the DFT.

A considerable advantage considering peaks ex-

tracted from the spectrogram is that we ﬁlter out noise

and background sounds.

• Peaks Vertical Alignment. Given the matrix of

peaks, we get the mean µ

and standard deviation

from the distribution of frequency gaps among

peaks in the same time bin.

∑

t∈T

|I|−1

∑

j=1

j+1

− I

), (14)

where I = {i ∈ F|p

6= 0 and

∑

> 1} ∀t,

and N

∑

t∈T

[(

∑

f ∈F

f t

) − 1].

∑

t∈T

|I|−1

∑

j=1

[(I

j+1

− I

) − µ

]

. (15)

Being a percussive sound, a crash will show peaks

stacked on the same time bin along all the fre-

quency range as shown in Figure 3(a). Thus, we

expect low average and standard deviation values.

• Peaks Horizontal Alignment. Given the matrix

of peaks, we get the mean µ

and standard de-

viation σ

from the distribution of time delays

among peaks at the same frequency.

∑

f ∈F

|J|−1

∑

i=1

i+1

− J

), (16)

where J = { j ∈ T |p

f j

6= 0 and

∑

f j

1} ∀ f , and N

∑

f ∈F

[(

∑

t∈T

f t

) − 1].

∑

f ∈F

|J|−1

∑

i=1

[(J

i+1

− J

) − µ

]

. (17)

Sounds such as car horn, tire skidding, and harsh

acceleration are sustained sounds and thus they

will present horizontal stripes of peaks on the

spectrogram as shown in Figures 3(c) and 3(e).

• Peak Entropy. Entropy of the peak constellation

is computed similarly to the entropy in the time

and the frequency domain. We split the spectro-

gram in k time bins and for the i

bin, we compute

the ratio between peaks included in that bin and all

the peaks in the constellation: e

∑

peaks

∑

peaks

PE = −

∑

i=1

log

). (18)

5 DATASET

In Spring 2016, AXA Winterthur has set the annual

crash test campaign in Switzerland. During this oc-

casion, the AXA Data Innovation Lab collected about

Crashzam: Sound-based Car Crash Detection

(a) Crash

(b) Car alarm (c) Tire skidding (d) Engine starting

(e) Car horn (f) Radio Music (g) Harsh acceleration

Figure 3: Example of spectrograms with constellation of peaks.

6.2 GB of data, including audio ﬁles recording cras-

hes from inside the car cabins. This original dataset

is composed by 46 signals of crash and, to the best of

our knowledge, it is the only dataset collecting crash

sounds from inside the car cabin

. Moreover, each

test was controlled in impact speed and angle, thus

such dataset has been essential to study crash dyna-

mics.

Therefore, to extend the set of controlled positive

samples, we extracted sounds recorded by dash ca-

meras published on the “Car Crash Time” Youtube

channel which provides several hours of crash recor-

dings (Car Crash Time, 2017). With respect to the

controlled dataset, such sounds are very realistic and

genuine, including background noises like rain, hail

or screams before the impacts. Overall, we collected

410 crash sounds recorded inside the cabin of cars in-

volved in the shock. Most of them (87%) come from

“Car Crash Time” and the rest from the AXA Wintert-

hur crash test campaign. We kept 100 samples aside

as test set.

As negative samples, we choose to include in

the dataset any sound which might be generated or

listened to inside a car cabin like radio music, pe-

ople talking, engine starting, car door opening or clo-

sing. We also included sounds that are often lin-

ked to car shocks like car horn, harsh decelerations,

and tire skidding to control the false positive rate.

Such sounds have been mostly imported from Free-

Sound (FreeSound, 2017) and Urban Sound Data-

set (Urban Sound Dataset, 2017). Although gathering

There exist a dataset of sounds provided by the MI-

VIA Computer Science department of Universit

a di Salerno

(Italy) for research purposes which includes crash sounds

recorded only from outside the car cabin (Foggia et al.,

2016b). Thus, such sounds were not eligible to include in

our dataset. Nevertheless it also contains tire skidding and

scream sounds.

Table 1: Dataset distribution in classes and sub-classes.

Class Sound type %

Crash

AXA Winterthur crash campaign 13

Car Crash Time 87

Other

Harsh acceleration or deceleration 10

Car horn 10

Car door opening and closing 8

Radio music 11

People talking 14

Tire skidding 10

Car alarm 5

Rain, hail, strong wind 10

Engine during driving 22

Figure 4: PDF for samples duration in the dataset.

negative sounds is a simpler task compared to the po-

sitive samples, we collected an equal number of sam-

ples in order to have a balanced dataset. Table 1 sum-

marize the dataset distribution into categories “Crash”

and “Other” and sub-categories.

All audio clips are sampled at 16 KHz, quantized

at 16 bits and all amplitudes are normalized. They

represent a mix of collision types: vehicle to barrier,

vehicle to vehicle, frontal impact, side impact, and at

different speeds.

A crash can occur between two or more cars, thus

it can last more or less time. Figure 4 shows the dis-

tribution of sample durations. The vast majority of

crashes involve only two cars and last ﬁve seconds,

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

while the longest crash can last also the double. We

selected negative samples in order to follow the same

distribution as crash sounds.

6 EVALUATION

Both time-frequency domain and spectrogram image

classiﬁers are random forest classiﬁers trained with

cross validation using the Python scikit-learn library.

Number of trees are 200 and 100 respectively while

all the other parameters are left unchanged.

Figures 5 shows ROC and precision-recall cur-

ves for classiﬁcation done by the two models in-

dependently and by combining their result proba-

bilities together. The time and frequency based

classiﬁer (Model 1) performs in line with other

models presented in the literature on sound de-

tection like gun shots or human screams for surveil-

lance purposes (Crocco et al., 2016). Its accuracy

(

∑

(True Positive)+

∑

(True Negative)

∑

Samples

) is equal to 0.875.

Nevertheless, it is wrong for some speciﬁc sub-types

of sounds with abrupt changes and high energy like

car horn or tire skidding.

As the spectrogram based classiﬁer (Model 2) is

speciﬁcally designed to discern between sustained

and percussive sounds, it is able to help the ﬁrst mo-

dels in such ambiguous situations. Although in ab-

solute terms Model 2 performs worse than Model 1,

the combination of the two (Model 3) in according to

Equation 1 brings the overall accuracy to 0.9. Models

weights in Equation 1 are set to w

= 0.6 and w

= 0.4

respectively. They come from an exhaustive research

to obtain the best results.

Most of the ﬁnal misclassiﬁcations correspond to

slammed door sounds.

7 DISCUSSION

Creating the constellation of peaks is highly depen-

dent from the spectrogram image in background.

Spectrogram is created spitting the time series of

PCM in small windows having a certain overlap and

computing the Fourier transform in each window.

Being samples also quantized at 16 bits, the constel-

lation of peaks will be depended by the number of

samples per window (NFFT) and by the overlap be-

tween consecutive windows. A short window will

tend to produce many peaks at the same frequency

since the same sound amplitude will be replicated for

many windows. Having a large overlap will have the

same impact. On the other hand, a large window will

(a) ROC curve.

(b) Precision-recall curve.

Figure 5: Comparison of accuracy, precision and recall

among singular models (Model 1 and 2) and the combina-

tion of such models in according to Equation 1 (Model 3).

Figure 6: Effects on classiﬁcation tuning the number of data

points used in each block for the Fast Fourier Transform

(NFFT) to create the spectrogram and the number of over-

lapped points.

increment horizontal peak gaps, but will tend to re-

duce vertical frequency gaps among peaks. Figure 6

shows how the accuracy on the testing dataset chan-

ges, varying both NFFT and the overlap, where the

overlap is at most equal to NFFT. The impact of such

parameters is quite important since the accuracy scale

is in the range [0.65-0.82]. Intermediate values for

NFFT, 256 or 512, introduce the least offset and they

are more independent to the overlap, thus we choose

NFFT = 512 and overlap = 354. Once the NFFT

and overlap set, the accuracy varies a little changing

the threshold and the distance to ﬁnd the local peaks.

We have set a threshold A

= 50 dB and a minimum

distance D = 3.

Crashzam: Sound-based Car Crash Detection

8 CONCLUSION AND

PERSPECTIVES

This manuscript introduces Crashzam, an innovative

way to detect car accidents with sound recorded by

smartphones or any other microphone-equipped smart

device installed in cars. The goal is to enhance dri-

vers’ safety with a low cost solution, capable to pro-

pose domain speciﬁc services like medical assistance

calling, ﬁrst notiﬁcation of loss (FNOL) or advice and

coordination during a minor event.

Sounds are analyzed by two models and the ﬁ-

nal detection result reﬂects a combination of both of

them. The ﬁrst model analyses the signal in time

and frequency, computing well-known features in the

sound recognition literature. The second one is based

on the analysis of spectrogram images and the disco-

very of patterns among a constellation of amplitude

peaks.

As no dataset of crash sounds recorded inside the

car cabin existed, we built one. Crash samples come

from both controlled experiments and real, genuine

conditions.

Detection results show a pretty accurate classiﬁ-

cation between crash sounds and other sounds likely

to be reproduced in car. Also, combining models ana-

lyzing speciﬁc aspects of sound signals (time series,

frequency, spectrogram) helps the system to be more

accurate and reliable when the environment is noisy.

A wide range of perspectives are possible. We

proposed Crashzam in the context of driving safety

and connected car, but the same concept could be app-

lied to other domains such as the connected home and

the connected health. For instance, Minut, a Scan-

dinavian startup, is specialized on sound-based home

surveillance (Minut, 2017). Their devices are able to

detect alarm or glass breaking. Remaining in a the te-

lematics context, the detection of repetitive car horn

sounds gives an insight on the drivers’ driving style.

As regards the system design, the tendency is to

embed models into devices. For instance, one of

the most advertised novelty in the 2017 Google I/O

was the possibility to embed TensorFlow models in-

side smartphones (Android and iOS) (Google I/O ’17,

2017). In the safe driving context, having an auto-

nomous system that gets rid of server calling would

mean a full time detection availability.

Deep learning has gained more and more estima-

tion on image and audio classiﬁcation tasks. Spectro-

gram image processing may be tested with Convolu-

tional Neural Networks to discover patterns. Nevert-

heless, in some working domains like insurance, it is

essential to know how to explain the results.

ACKNOWLEDGEMENTS

The authors would like to thank AXA Winterthur for

inviting the AXA Data Innovation Lab to participate

to the Spring 2016 crash test campaign and AXA Sin-

gapore for helping with the data crawling and clea-

ning.

REFERENCES

Aloul, F. A., Zualkernan, I. A., Abu-Salma, R., Al-Ali, H.,

and Al-Merri, M. (2015). ibump: Smartphone appli-

cation to detect car accidents. Computers & Electrical

Engineering, 43:66–75.

Car Crash Time (2017). Youtube channel. https://www.

youtube.com/channel/UCwuZi C yFtHsfCicthqygw.

Accessed: 2017-05-03.

Carletti, V., Foggia, P., Percannella, G., Saggese, A., Stris-

ciuglio, N., and Vento, M. (2013). Audio surveillance

using a bag of aural words classiﬁer. 2013 10th IEEE

International Conference on Advanced Video and Sig-

nal Based Surveillance, AVSS 2013, (December):81–

86.

Clavel, C. and Ehrette, T. (2008). Fear-type emotion re-

cognition and abnormal events detection for an audio-

based surveillance system. WIT Transactions on In-

formation and Communication Technologies, 39:471–

479.

Crocco, M., Cristani, M., Trucco, A., and Murino, V.

(2016). Audio Surveillance: a Systematic Review.

ACM Computing Surveys, 48(4):1–44.

Dixon, S. (2006). Onset Detection Revisited. In Proc. of the

Int. Conf. on Digital Audio Effects (DAFx-06), pages

133–137, Montreal, Quebec, Canada.

Evanco, W. M. (1996). The impact of rapid incident de-

tection on freeway accident fatalities. Technical re-

port.

Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., and

Vento, M. (2016a). Audio surveillance of roads: A

system for detecting anomalous sounds. IEEE Tran-

sactions on Intelligent Transportation Systems.

Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., and

Vento, M. (2016b). Audio surveillance of roads:

A system for detecting anomalous sounds. IEEE

Transactions on Intelligent Transportation Systems,

17(1):279–288.

FreeSound (2017). On-line free sound samples.

www.freesound.org. Accessed: 2017-05-03.

Ganchev, T., Fakotakis, N., and Kokkinakis, G. (2005).

Comparative evaluation of various mfcc implementa-

tions on the speaker veriﬁcation task. In in Proc. of

the SPECOM-2005, pages 191–194.

Giannakopoulos, T., Pikrakis, A., and Theodoridis, S.

(2007). A multi-class audio classiﬁcation method with

respect to violent content in movies using bayesian

networks. In IEEE 9th Workshop on Multimedia Sig-

nal Processing, MMSP 2007, Chania, Crete, Greece,

October 1-3, 2007, pages 90–93.

VEHITS 2018 - 4th International Conference on Vehicle Technology and Intelligent Transport Systems

Gonzalez, R. (2013). Better than MFCC audio classiﬁcation

features. The Era of Interactive Media, pages 291–

301.

Google I/O ’17 (2017). Android meets Ten-

sorFlow: How to accelerate your app with AI.

https://www.youtube.com/watch?v=25ISTLhz0ys.

Accessed: 2017-05-03.

Gouyon, F., Delerue, O., and Pachet, F. (2000a). Classifying

percussive sounds: a matter of zero-crossing rate ? In

Proceedings of DAFX 00, Verona (It).

Gouyon, F., Pachet, F., and Delerue, O. (2000b). On the use

of Zero-Crossing rate for an application of classiﬁca-

tion of percussive sounds. International Conference

on Digital Audio Effects, (August 2002):3–8.

Grey, J. M. and Gordon, J. W. (1978). Perceptual effects of

spectral modiﬁcations on musical timbres. The Jour-

nal of the Acoustical Society of America, 63(5):1493–

1500.

Kurth, F. and M

uller, M. (2008). Efﬁcient index-based au-

dio matching. IEEE Trans. Audio, Speech & Lan-

guage Processing, 16(2):382–395.

Lahn, J., Peter, H., and Braun, P. (2015). Car crash detection

on smartphones. In Proceedings of the 2Nd Internati-

onal Workshop on Sensor-based Activity Recognition

and Interaction, WOAR ’15, pages 12:1–12:4, New

York, NY, USA. ACM.

Minut (2017). Home surveillance. https://minut.com/.

Accessed: 2017-05-03.

Mueller, M., Kurth, F., and Clausen, M. (2005). Audio Ma-

tching via Chroma-Based Statistical Features. pages

288–295, London, UK. University of London.

uller, M. (2007). Information Retrieval for Music and

Motion. Springer-Verlag New York, Inc., Secaucus,

NJ, USA.

Peeters, G. and Rodet, X. (2004). A large set of audio fe-

ature for sound description (similarity and classiﬁca-

tion) in the cuidado project. Technical report, Ircam,

Analysis/Synthesis Team, 1 pl. Igor Stravinsky, 75004

Paris, France.

Pikrakis, A., Giannakopoulos, T., and Theodoridis, S.

(2008). Gunshot detection in audio streams from mo-

vies by means of dynamic programming and bayesian

networks. In Proceedings of the IEEE International

Conference on Acoustics, Speech, and Signal Proces-

sing, ICASSP 2008, March 30 - April 4, 2008, Caesars

Palace, Las Vegas, Nevada, USA, pages 21–24.

Punetha, D., Kumar, D., and Mehta, V. (2012). Article: De-

sign and realization of the accelerometer based trans-

portation system (ats). International Journal of Com-

puter Applications, 49(15):17–20. Full text available.

Rabaoui, A., Davy, M., Rossignol, S., and Ellouze, N.

(2008). Using one-class SVMs and wavelets for au-

dio surveillance. IEEE Transactions on Information

Forensics and Security, 3(4):763–775.

Schulz, A., Ristoski, P., and Paulheim, H. (2013). I see a car

crash: Real-time detection of small scale incidents in

microblogs. In Extended Semantic Web Conference,

pages 22–33. Springer.

Sengupta, N., Sahidullah, M., and Saha, G. (2016). Lung

sound classiﬁcation using cepstral-based statistical fe-

atures. Computers in Biology and Medicine, 75:118–

129.

Sosmartapp (2017). Automatic car crash detection app.

http://www.sosmartapp.com. Accessed: 2017-05-03.

Thompson, C., White, J., Dougherty, B., Albright, A., and

Schmidt, D. C. (2010). Using Smartphones to Detect

Car Accidents and Provide Situational Awareness to

Emergency Responders, pages 29–42. Springer Berlin

Heidelberg, Berlin, Heidelberg.

TrueMotion (2017). Driving Intelligence.

https://gotruemotion.com/. Accessed: 2017-05-

03.

Urban Sound Dataset (2017). https://serv.cusp.nyu.edu/

projects/urbansounddataset/urbansound8k.html.

Accessed: 2017-09-03.

Valenzise, G., Gerosa, L., Tagliasacchi, M., Antonacci, F.,

and Sarti, A. (2007). Scream and gunshot detection

and localization for audio-surveillance systems. 2007

IEEE Conference on Advanced Video and Signal Ba-

sed Surveillance, AVSS 2007 Proceedings, pages 21–

26.

White, J., Thompson, C., Turner, H., Dougherty, B., and

Schmidt, D. C. (2011). Wreckwatch: Automatic traf-

ﬁc accident detection and notiﬁcation with smartpho-

nes. Mobile Networks and Applications, 16(3):285–

303.

World Health Organization (2015). Global status report on

road safety.

Zaldivar, J., Calafate, C. T., Cano, J. C., and Manzoni,

P. (2011). Providing accident detection in vehi-

cular networks through obd-ii devices and android-

based smartphones. In Local Computer Networks

(LCN), 2011 IEEE 36th Conference on, pages 813–

819. IEEE.

Zendrive (2017). Safer Drivers, Safer Roads. http://

www.zendrive.com/. Accessed: 2017-05-03.

Crashzam: Sound-based Car Crash Detection