ASIMS: Acceleration Spectrograms Based Intelligent Mobility System

for Vehicle Damage Detection

Sara Khan

, Mehmed Y

uksel

and Andre Ferreira

Robert Bosch GmbH, Renningen, Germany

DFKI GmbH Robotics Innovation Center, Bremen, Germany

Robert Bosch Car Multimedia S.A., Braga, Portugal

Keywords:

Automobile, Machine Learning, Damage detection, Cosmetic Damages, Inertial Sensors, Autoencoders.

Abstract:

Every vehicle is susceptible to several types of small physical damage such as dents and scratches. These

damages can be seen as cosmetic damages as they impact the vehicle’s visual and value but do not alter its

main functions. Vehicle owners, insurance companies, and the car-rental/taxi-service companies are especially

keen to detect the events that generate these kinds of damages. The ability to detect impact events is valuable

to monitor the occurrence of possible damages to the vehicles. In this paper, we present a novel acceleration

spectrogram-based Machine Learning (ML) approach for dynamic (real-time) small vehicle damage detection

using inertial sensors. Inertial sensors are low-resource consumption sensors, which makes the proposed

solution economical. Conventionally, inertial sensors are used in the airbag control system but they are not

developed to detect impacts that generate minor damages. Most of the previous work on small impact detection

either uses smartphone inertial data which is not accurate or focuses on static damage detection based on

image sensory inputs. Our intelligent impact and damage detection ML-based system uses autoencoders

as an automatic feature extractor using acceleration spectrograms and classiﬁes the sensory encoded feature

representation into damage or non-damage. It can achieve an accuracy of 0.8. This approach sets the stage for

various potential research directions in damage detection.

1 INTRODUCTION

The common interest of car manufacturers, and car

owners - from individual persons to car-sharing com-

panies, passengers, or other players (insurance, in-

spection and service companies, etc.) is to moni-

tor the vehicle status in various aspects.This vehi-

cle status information is imperative to detect risks

and anomalies or accurately predict malfunctions that

may occur in non-autonomous or autonomous driving

functions to avoid unexpected error cascades.

Traditionally, inertial sensors have been used in

automotive systems for airbag control (Shi et al.,

2008) and can help in detecting big damages. But

what if we want to detect small damages like minor

dents or scratches? Small damage detection infor-

mation has a variety of applications in new mobil-

ity services like car-sharing and ride-sharing applica-

tions where vehicle damage information is very cru-

cial to know. For example, a ride-sharing application

involves a user picking up a vehicle from one point,

using it for personal riding, and then dropping the

vehicle at the vehicle drop point. If damage occurs

during the user’s ride, then the damage event detec-

tion system provides the damage information to the

stakeholders even if they are imperceptible to take

the corresponding actions for the occurred damage.

Also, this damage prediction information can be used

in identifying the damages easily and in turn reduces

the tedious process of vehicle inspection by insurance

companies while claiming insurance in case of vehi-

cle damages (Li et al., 2021).

Thus, in this paper, we propose a small damage

detection system where the goal is to detect events

resulting in damages to the vehicle and classify the

damages as cosmetic damages and non-damages. A

damaging event can be deﬁned as an event that has

a physically negative impact on the vehicle structure

but may not necessarily result in the vehicle mal-

functioning. Previous research works in the ﬁeld of

damage detection are static large damage detection

and use camera sensors data as the main input. On

Khan, S., Yüksel, M. and Ferreira, A.

ASIMS: Acceleration Spectrograms Based Intelligent Mobility System for Vehicle Damage Detection.

DOI: 10.5220/0011763200003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 179-186

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

179

the contrary, our work focuses solely on small dam-

age detection with inertial sensors which is a light

weighted and efﬁcient solution. Also, we build our

model using real-life data, which has been not used

before in general. In this paper, we propose a method

for small damage detection using a machine learn-

ing approach termed as ASIMS. Traditional machine

learning involves computing features to train machine

learning models. But this process has high complex-

ity and requires domain expertise (Vejdannik et al.,

2018). Thus, we compute feature representation us-

ing the deep learning-based automatic feature extrac-

tor method approach. And, use this compressed rep-

resentation for binary damage detection and classiﬁ-

cation.

2 RELATED WORK

Vehicle Damage Detection is an established domain

in the literature. The damages to automobiles can

either be signiﬁcant or small. Figure 1 gives an

overview of cosmetic (or small) damages and sig-

niﬁcant (or large) damages that are present on ve-

hicles. Cosmetic damage image contains a slight

scratch on the vehicle and negatively affects vehicle’s

appearance. Alternatively, signiﬁcant damage image

contains one or more damages to the vehicle’s body

which may impact its functionality too. In the auto-

motive industry, there has been a substantial amount

of prior research on damage detection. Most of the

previous work concentrates on static damage detec-

tion using camera sensors. In the below sections, we

summarise the current work on damage detection and

also describe how our work is different from previous

work and points in a different direction.

Cosmetic Damage Signiﬁcant Damage

Figure 1: Kinds of damages.

2.1 Damage Detection Using Cameras

Researchers in (Li et al., 2021) aim to automatically

detect and classify vehicle damages, to fasten the pro-

cess of insurance claims. They have used a com-

bination of deep learning and transfer learning for

high performance and overcome training data limi-

tations. Unlike other works, the authors concentrate

on more relevant inputs by reducing the noise and

irrelevant items. For the dataset, the authors have

used web scraping to extract damages from Google

images. Furthermore, they are manually labeled into

different damage categories (e.g., bumper damage, no

damage, glass damage, etc.). They ﬁrst use Mask

Regional-Convolutional Neural Network (Mask R-

CNN) to crop the damaged car and remove other irrel-

evant items. And then (O’Shea and Nash, 2015) pre-

trained on ImageNet (Deng et al., 2009) is used for

classiﬁcation. Using transfer learning, eight different

architectures are evaluated. Among them, (Simonyan

and Zisserman, 2015) performs the best with 87.5 ac-

curacy. Along similar lines, the work in (Dwivedi

et al., 2020) uses CNN and pre-trained models for

classiﬁcation and detection into 7 different classes. In

(Kyu and Woraratpanya, 2020), the authors created

their own dataset consisting of approximately 1200

images. They categorized damage severity into mi-

nor, moderate, and severe using a predeﬁned VGG-16

model with L2 regularisation. In another work (Singh

et al., 2019), the authors proposed an end-to-end sys-

tem for claiming insurance for car damages. To do

so, the authors propose a model that takes images as

input and classiﬁes them into non-damage or damage.

It further localizes damage and classiﬁes its severity

into mild and severe. It also provides additional de-

cisions on whether the damaged part needs to be re-

placed or not. For purpose of modeling, they have

used instance segmentation models like PANet(Liu

et al., 2018), Mask R-CNN, and ensemble version of

both using transfer learning based on VGG16.

The drawback of camera sensor-based damage de-

tection is it cannot be used for real-time detection.

The cameras present in autonomous driving systems

are used to capture surroundings rather than the car

itself (Ess et al., 2010). The data from these cam-

era sensors is huge and requires a lot of computation.

Also, it is not economical to have such an expensive

sensor setup for a ﬂeet of shared vehicles.

2.2 Damage Detection Using Audio

On the contrary, not all of the previous work relies on

camera sensors for damage detection. The authors in

(Sammarco and Detyniecki, 2018) use audio signals

from car impact to detect accidents using their created

dataset (Sammarco and Detyniecki, 2019). They ex-

tract time and frequency domain features of input sig-

nals and use them for classiﬁcation. Their proposed

model is able to differentiate between crash sounds

and other in-vehicle sounds. In (Choi et al., 2021),

the authors use multi-modal data including both audio

and video signals for crash detection. A Gated Recur-

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

180

rent Unit (GRU) and Convolutional Neural Network

(CNN) based classiﬁers are used for the input. The

results show that ensemble classiﬁers perform better

than a single classiﬁer. In (Hashimoto et al., 2019),

the authors proposed a method to detect abnormal

vibrations in cars. They used piezoelectric sensors

waveform as an input and feature processing using

Mel-frequency coefﬁcients, and CNN for modeling

provided optimal results. There are some other ap-

plications of damage detection in road damage and

structural health monitoring using audio data. The

work present in (Gontscharov et al., 2014) presents an

approach for the automatic detection of minor vehicle

body damages. The author uses a sensor network in-

tegrated vehicle body and uses acoustic vehicle noise

level for structural vehicle damage detection.

2.3 Damage Detection Using Inertial

Sensors

Conventional solutions use accelerometer data to de-

tect more or fewer spikes caused by sudden changes

in acceleration in one or more axes (Punetha et al.,

2012). The abrupt changes in G-force trigger the

air bag control for vehicle safety. But it is not able

to detect impacts that are minor or cosmetic. Many

previous works have used inertial sensors in smart-

phones for accident or incident detection in vehicles.

The work in (Zaldivar et al., 2011) combines the ac-

celeration sensors present in smartphones with vehi-

cle’s onboard diagnostics. With their onboard inter-

face, they are able to detect accidents. Another work

(White et al., 2011) presents a smartphone-based ac-

cident detection system named WreckWarch which

records forces through accelerometers experienced in

collisions. But none of the smartphone-based systems

are accurate as smartphone data is hard to analyze due

to calibration, noise, and rotation issues.Also, relying

solely on acceleration data can lead to incorrect pre-

dictions in many situations. Bumps, potholes, and

poor road conditions cause false alarms, while sta-

tionary rear-end collisions can be classiﬁed as nor-

mal acceleration (Sammarco and Detyniecki, 2018).

Works in (Gontscharov et al., 2014), (Sammarco and

Detyniecki, 2018) and (Punetha et al., 2012) moti-

vates us to build a data-driven intelligent system for

detecting small damages in real-time.

The main contributions of our work are:

1. Novel Machine Learning(ML) based approach us-

ing acceleration signal spectrogram for small ve-

hicle damage detection.

2. Use of damage data acquired through ﬁeld tests

for empirical evaluation.

Table 1: Examples of different damage and non-damage

events.

Damage Event Non-Damage Event

Passive vehicle collision Sun visor

Active vehicle collision Trunk opening /closing

3. Overview of state-of-the art in automotive damage

detection.

3 SMALL DAMAGE DETECTION

Small Damage Detection is a data-driven damage de-

tection system that detects damages using accelerom-

eter information. It is an event-based mechanism.

Events are triggered while a person is riding the ve-

hicle. Events are categorized into damage and non-

damage events (Gontscharov et al., 2014). Table 1

shows some examples of damage and non-damage

events. For example, a passive vehicle collision is a

damage event that occurs when another vehicle or ob-

ject hits our vehicle.

The damage detection system work as shown in Fig-

ure 2. When the sensory data hits a certain thresh-

old, the event is triggered and the information within

a certain window is collected. This raw information

is passed through pre-processing step and fed into a

trained machine-learning model. The prediction of

the model is either damage or non-damage(or back-

ground). If damage occurs, the information is sent

to the cloud service for further steps. The prediction

system follows the CRISP-DM framework (Wirth and

Hipp, 2000), which includes data understanding, and

data preparation as part of data analysis, modeling,

and evaluation.

Figure 2: Small Damage Detection System.

3.1 Dataset and Pre-Processing

The primary source of data is real-ﬁeld test data col-

lected by creating the damage scenarios on vehicles.

A small piece of hardware containing inertial sensors

is mounted on the windshield of the car. Some of

the cars used in data collection are BMW 5 series,

BMW i3 and Mercedes GLA 180. The roads that

were tested on to collect the data were categorised as

asphalt, stone road, mud, dirt, snow and gravel. There

were 10 different drivers who performed the maneu-

ASIMS: Acceleration Spectrograms Based Intelligent Mobility System for Vehicle Damage Detection

181

vers with the speed range from 10 km/hr to 100km/hr

under different weather conditions like sunny, rain-

ing and cloudy. The database includes accelerometer

and gyroscope sensory information along with times-

tamps. A continuous stream of data is collected keep-

ing possible damage and non-damage events. Some

examples of such events are mentioned in table 1. We

gather those raw signals that cross a predetermined

threshold as damaging occurrences have high values

of inertial sensors. Then, a window size of one second

is collected where it captures information of 100ms

before the threshold and 900 ms after the threshold.

The reason to use 100 ms before threshold is to give

a small context before the event. This event informa-

tion contains a damaging or non-damaging event. The

inertial sensors are sampled at a frequency of 1600

Hz. Thus, for each event (of window size 1), it con-

tains the x,y, and z-axis of acceleration values along

with car-speciﬁc information and corresponding dam-

age label. After this, the signals are passed through a

low pass ﬁlter of 220Hz.

Figure 3: Acceleration x-axis, y-axis and z-axis for a dam-

aged event.

Figure 3 and 4 depict the measurements from ac-

celerometer and gyroscope in all three directions- X,

Y, and Z axis. Acceleration and Gyroscope are both

digital signals. Acceleration measures linear acceler-

ation while the gyroscope provides angular velocity.

In the above-discussed ﬁgures, we can see the rate of

change of acceleration differs from the rate of change

of gyroscope concerning time duration. Both of these

ﬁgures are examples of a damage event in which a ve-

hicle hits the left door and a dent is formed. The high

spikes between intervals 0.1 and 0.4 clearly show the

occurrence of a damage event that results in a dent.

This work uses acceleration information only.

Figure 4: Gyroscope x-axis, y-axis and z-axis for a dam-

aged event.

Figure 5: Modelling framework.

3.2 Methodology: ASIMS

The modelling approach of ASIMS system is divided

into two sub-sections. The ﬁrst sub-section explains

the usage of deep learning as an automatic feature

extractor to encode feature representations, which is

trained using unsupervised learning. The second sub-

section uses this encoded representation for a binary

classiﬁcation task, which is trained in a supervised

technique using damage labels. Figure 5 depicts the

modeling framework at the train and test level. It will

be explained in the below sections.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

182

3.2.1 Autoencoders as Feature Extractor

Autoencoder is an unsupervised learning technique

that is used to reduce dimensionality and help reduce

noise (Liu et al., 2016). It seeks an input image, con-

verts into a latent or compressed representation, and

reconstructs it as an output image. In recent years,

many previous works have suggested autoencoder as

an automatic feature extractor. Although it’s a black

box, it is fast and doesn’t require domain expertise.

As a result, we will employ it as a tool for feature ex-

traction in this study.

To use autoencoders, we ﬁrst convert the gener-

ated event signal information of window size of

one second into a spectrogram image using Short-

Time Fourier Transform (STFT) (Mateo and Talavera,

2017) at a sampling frequency of 1600 Hz. A spectro-

gram is a visual representation of a signal (French and

Handy, 2007). This results in three spectrograms: x,

y, and z for each window sample. Figure 6 depicts the

Figure 6: An acceleration y-axis signal into acceleration

spectrogram through STFT.

transformation of the y-axis of the acceleration signal

into a spectrogram. These generated spectrograms are

trained using autoencoders as shown in the training

section of Figure 5. The spectrograms are fed into an

encoder which creates a compressed encoded repre-

sentation. This encoded representation captures the

characteristics of input data. Then, a decoder recon-

structs the output based on latent representation. For

our study, we discard the decoder and use latent rep-

resentation to train our damage classiﬁcation model.

3.2.2 Damage Prediction

The goal of damage detection is to build a Gradi-

ent Boosting (GB) based binary classiﬁcation model,

which takes the input as ﬂattened encoded representa-

tion of a spectrogram obtained from autoencoder and

outputs the damage predictions. This ﬂattened en-

coded representation of an input image and the corre-

sponding damage label are used as input-output pairs

for supervised training. The trained model is then

used for prediction as shown in the test section of Fig-

ure 5.

3.3 Implementation

This work was implemented in Python language. The

modeling was done in two parts. The ﬁrst sub-part

involves converting signals into spectrograms. This

was done using Matplotlib library (Hunter, 2007).

The second subpart involves creating an encoded

feature representation using auto-encoders. This

was done using TensorFlow (Abadi et al., 2015) li-

brary. Grayscale spectrograms of X, Y, and Z were

fed channel-wise into the autoencoder. Grayscale

spectrograms contain the same information as color-

mapped spectrograms (French and Handy, 2007). The

colors only depict the aesthetic aspect of a spectro-

gram. The autoencoder architecture was a three-

convolution layer architecture with batch normaliza-

tion on each layer where the encoded representation

contains 1024 features. The neurons in each lay-

ers are 200,100, 1 respectively and a 3*3 convolu-

tion ﬁlter on each layer. Figure 7 shows the train-

ing validation curve for a hundred epochs and a learn-

ing rate of 0.0001. The training was performed us-

ing Tesla V100-SXM2-32GB GPU containing 32 GB

RAM. The data contained around unlabelled 50K im-

age samples for feature representation. It was split

into 80 percent training data and 20 percent valida-

tion data. The ratio between number of non-damaged

samples to damaged samples is 99:1. This high imbal-

ance data is a reﬂection of real-world scenario where

damage cases are pretty low. The second part involves

Figure 7: Training and validation curve for autoencoder ar-

chitecture.

ﬁtting the whole training data with encoded represen-

tation and corresponding binary labels in a supervised

learning fashion. Our modeling framework uses Gra-

dient Boosting (GB) as a classiﬁcation model. Our

proposed model was compared against Random for-

est (RF) and Gradient Boosting (GB) classical ML

models without encoded representation. All ML algo-

rithms are implemented using Scikit-learn (Pedregosa

ASIMS: Acceleration Spectrograms Based Intelligent Mobility System for Vehicle Damage Detection

183

et al., 2011) library with their default settings. The re-

sults were evaluated using labeled training data con-

taining approximately 2500 spectrograms for training

and testing. The training and the testing split is 4:1.

Figure 8: t-SNE representation using k-means. A value of

1 represents damage and 0 represents non-damage.

Figure 8 shows the data distribution of encoded

training data using t-SNE (van der Maaten and Hin-

ton, 2008) and clustered using the k-means algorithm

into two clusters. The data contains approximately

150 equally balanced damaged and non-damaged

samples. The orange color and turquoise color rep-

resent damage and non-damage samples respectively.

As we can see that individual representations of each

axis are non-separable. But when we consider the

encoded representation altogether, two clusters are

formed. Does this convey that clustering is enough

for our problem? So, we increased the number of

samples and performed clustering. Figure 9 shows

the training data distribution with an increased num-

ber of samples. The data becomes more unbalanced

as the number of samples rises. Thus, the clustering

becomes intractable with increase in number of sam-

ples. This demonstrates that a simple unsupervised

technique like clustering is not suitable for this prob-

lem. Therefore, this enabled us to design supervised

learning based classiﬁcation.

Figure 9: Data imbalance ratio increase with increase num-

ber of samples

4 RESULTS AND DISCUSSIONS

We use a test dataset containing 50K unlabeled sam-

ples for autoencoder results. The training part is eval-

uated using Linear Algebra (LA) norm. LA norm is a

mean squared metric that measures the squared differ-

ence between the input and the reconstructed image.

Table 2 displays LA norm results for the test dataset

Table 2: Test results using autoenoder. L-R columns: num-

ber of test samples, minimum , maximum , mean and stan-

dard deviation of errors.

No. Min. Max. Mean SD.

50K 0.232 0.545 0.121 0.5342

containing around 50K samples. The second column

and third columns depict the minimum and maximum

error between any image and its prediction of all the

test samples. The low values explain that the feature

extractor model has been able to generalize well. Fig-

ure 10 shows an example of an x-axis grayscale spec-

trogram with its test image and its reconstruction.

Test Input Predicted Output

Figure 10: An input with dimensions 256*256 containing x-

axis acceleration spectrogram image and its corresponding

reconstruction with 0.079 error.

The latent representation of the trained model is

then used for classiﬁcation and compared to non-

encoded representations. For the damage detection

classiﬁcation, we use metrics calculated using a con-

fusion matrix for evaluation. Figure 11 summarises

the results of the confusion matrix for our model that

is trained and tested on approximately 500 spectro-

grams each. Every sample containing three spectro-

grams images per sample. A proportion of 0.14 and

0.05 are False Negatives and False Positives respec-

tively. Door closing is a non-damage event that the

model frequently interprets incorrectly and predicts

as damage. By including better data in the model,

this can be improved. Currently, STFT is used to

construct spectrograms. Instead, Continuous Wavelet

Transform (CWT) (Yunhui and Qiuqi, 2004) can be

applied. As it provides ﬁner details by creating bins

of dynamic sizes. Table 3 summarises evaluation re-

sults. The training data contains around 6000 spec-

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

184

trogram images (three for each sample) and the test

dataset contained approximately 1200 images. Ac-

curacy, Precision, and F1 score are the metrics com-

puted from the confusion matrix for evaluation. Our

proposed model ASIMS is compared against Gradient

Boosting (GB) present in the second row and Ran-

dom Forest (RF) respectively. Both of these classiﬁ-

cation models are trained without encoded represen-

tation. The high values of accuracy and F1-score con-

ﬁrm that our approach outperforms the traditional ML

models. Accuracy can be misleading when classes are

imbalanced. Therefore, we measure F1-score as well.

Thus, encoded representation has a positive effect on

performance. Also, the number of features used in

our model is one-third of the feature used by other

models.

Figure 11: Confusion matrix for test data where 1 represents

damage and 0 represents non-damage.

Table 3: Test results for damage classiﬁcation.

Model Features Accuracy Precision F1

ASIMS 1024 0.82 0.84 0.89

GB 3072 0.70 1.00 0.09

RF 3072 0.20 0.82 0.82

5 CONCLUSIONS

This research introduces a semi-supervised machine

learning approach for small damage detection based

on acceleration spectrograms. Autoencoders are used

to construct the encoded feature representation in an

unsupervised manner, and this encoded representa-

tion is then utilised to train classiﬁer with labels in

a supervised manner. The majority of the earlier re-

search focused on camera sensor-based static damage

detection. Conventionally, inertial sensors are used in

air bag control system to detect signiﬁcant damages.

However, they are unable to identify minor damages.

We collected and used real ﬁeld data with custom

hardware solution as opposed to other research that

work on synthetic or smartphone based data. As a re-

sult, our outcome is closer to a the real-world closer.

We show how to anticipate damage occurrences us-

ing feature-encoded methods, and we assess our per-

formance using well-known metrics like accuracy and

F1 score. Gradient Boosting (GB) and Random For-

est (RF) without encoded representations were used

to compare our results, and we discovered that our

model outperforms them both. As a result, various

research directions can be started using this baseline

results.

6 OUTLOOK

As a further extension of this work, ﬁrst, we can in-

clude gyroscopic data. It may boost the performance

of detection by reducing false positives and false neg-

atives. Second, the dataset is imbalanced. And we can

use techniques of class weights and over-sampling for

improving results. Third, the current algorithm de-

tects and classiﬁes into damages and non-damages. It

would be more helpful if we can further categorize

into damage type and damage severity. Fourth, the

dataset is highly asymmetric as undamaged samples

outperform damaged samples which makes training

the model quite difﬁcult. Thus, we would like to ex-

plore advanced deep-learning models. Fifth, it will

be helpful to compare the results of the autoencoder

present in this study to traditional feature engineering

methods. Lastly, the current framework includes in-

ertial sensor information. But from different damage

studies present in section 2, audio has helped in im-

proving damage detection. Thus, we would consider

using audio information as another modality, which

may in turn help in improving the prediction perfor-

mance.

REFERENCES

Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z.,

Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin,

M., Ghemawat, S., Goodfellow, I., Harp, A., Irving,

G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kud-

lur, M., Levenberg, J., Man

e, D., Monga, R., Moore,

S., Murray, D., Olah, C., Schuster, M., Shlens, J.,

Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Van-

houcke, V., Vasudevan, V., Vi

egas, F., Vinyals, O.,

Warden, P., Wattenberg, M., Wicke, M., Yu, Y., and

Zheng, X. (2015). TensorFlow: Large-scale machine

learning on heterogeneous systems. Software avail-

able from tensorﬂow.org.

Choi, J., Kong, C., Kim, G., and Lim, S. (2021). Car crash

detection using ensemble deep learning and multi-

modal data from dashboard cameras. Expert Systems

with Applications, 183:115400.

ASIMS: Acceleration Spectrograms Based Intelligent Mobility System for Vehicle Damage Detection

185

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). Imagenet: A large-scale hierarchical

image database. In 2009 IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 248–255.

Dwivedi, M., Malik, H. S., Omkar, S. N., Monis, E. B.,

Khanna, B., Samal, S. R., Tiwari, A., and Rathi, A. M.

(2020). Deep learning-based car damage classiﬁcation

and detection.

Ess, A., Leibe, B., and Van Gool, L. (2010). Object de-

tection and tracking for autonomous navigation in dy-

namic environments. I. J. Robotic Res., 29:1707–

1725.

French, M. and Handy, R. (2007). Spectrograms: Turning

signals into pictures. Journal of Engineering Technol-

ogy, 24:32–35.

Gontscharov, S., Baumgaertel, K., Kneifel, A., and Krieger,

K.-L. (2014). Algorithm development for minor dam-

age identiﬁcation in vehicle bodies using adaptive sen-

sor data processing. Procedia Technology, 15.

Hashimoto, W., Hirota, M., Araki, T., Yamamoto, Y., Egi,

M., Hirate, M., Maura, M., and Ishikawa, H. (2019).

Detection of car abnormal vibration using machine

learning. In 2019 IEEE International Symposium on

Multimedia (ISM), pages 40–407.

Hunter, J. D. (2007). Matplotlib: A 2d graphics environ-

ment. Computing in Science & Engineering, 9(3):90–

95.

Kyu, P. M. and Woraratpanya, K. (2020). Car damage de-

tection and classiﬁcation. In Proceedings of the 11th

International Conference on Advances in Information

Technology, IAIT2020, New York, NY, USA. Associ-

ation for Computing Machinery.

Li, L., Ono, K., and Ngan, C.-K. (2021). A Deep Learning

and Transfer Learning Approach for Vehicle Damage

Detection. The International FLAIRS Conference Pro-

ceedings, 34(1).

Liu, S., Qi, L., Qin, H., Shi, J., and Jia, J. (2018). Path ag-

gregation network for instance segmentation. In 2018

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition, pages 8759–8768.

Liu, S., Zhang, C., and Ma, J. (2016). Stacked auto-

encoders for feature extraction with neural networks.

pages 377–384.

Mateo, C. and Talavera, J. (2017). Short-time fourier trans-

form with the window size ﬁxed in the frequency do-

main. Digital Signal Processing, 77.

O’Shea, K. and Nash, R. (2015). An introduction to convo-

lutional neural networks. ArXiv e-prints.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Punetha, D., Kumar, D., and Mehta, V. (2012). Design and

realization of the accelerometer based transportation

system (ats). International Journal of Computer Ap-

plications, 49:17–20.

Sammarco, M. and Detyniecki, M. (2018). Crashzam:

Sound-based car crash detection.

Sammarco, M. and Detyniecki, M. (2019). Car Accident

Detection and Reconstruction Through Sound Analy-

sis with Crashzam, pages 159–180.

Shi, G., Chan, C.-S., Zhang, G., Li, W., Leong, P., and Le-

ung, K.-S. (2008). Towards a mobile airbag system

using mems sensors and embedded intelligence. pages

634–639.

Simonyan, K. and Zisserman, A. (2015). Very deep con-

volutional networks for large-scale image recognition.

In International Conference on Learning Representa-

tions.

Singh, R., Ayyar, M. P., Sri Pavan, T. V., Gosain, S., and

Shah, R. R. (2019). Automating car insurance claims

using deep learning techniques. In 2019 IEEE Fifth

International Conference on Multimedia Big Data

(BigMM), pages 199–207.

van der Maaten, L. and Hinton, G. (2008). Viualizing data

using t-sne. Journal of Machine Learning Research,

9:2579–2605.

Vejdannik, M., Sadr, A., Albuquerque, V., and Tavares, J.

(2018). Signal Processing for NDE, pages 1–19.

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

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

trafﬁc accident detection and notiﬁcation with smart-

phones. Mob. Netw. Appl., 16(3):285–303.

Wirth, R. and Hipp, J. (2000). Crisp-dm: Towards a stan-

dard process model for data mining. Proceedings of

the 4th International Conference on the Practical Ap-

plications of Knowledge Discovery and Data Mining.

Yunhui, S. and Qiuqi, R. (2004). Continuous wavelet trans-

forms. volume 1, pages 207– 210 vol.1.

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

P. (2011). Providing accident detection in vehicular

networks through obd-ii devices and android-based

smartphones. 2011 IEEE 36th Conference on Local

Computer Networks, pages 813–819.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

186