VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)

ISSN(Online): 2581-7280 Article No. X

PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

Survey of accident detection systems

Raj Shah1, Heena Shaikh2 , Devashish Shetty3, Dr. Tatwadarshi

Nagarhalli4
1(Department of Computer Engineering, Mumbai University,
India) Email: [email protected]
2(Department of Computer Engineering, Mumbai University,
India) Email: [email protected]
3(Department of Computer Engineering, Mumbai University, India)
Email:[email protected]
4(Department of Computer Engineering, Mumbai University,
India) Email: [email protected]

Abstract: Vehicle accidents are by all accounts appalling and frightening occasions occurring which
cause various deaths. As the number of accidents per year is increasing tremendously and so the lives
affected by accidents. There are traditional ways to help the needy or the victim that is informing the
right authority but needs assistance or help from others, but this tends to take ample of time and due to it
could cost lives. So there is a need to develop an accident detection system that would detect and alert the
proper authorities about the accident. The sudden assistance to the alert would in return lead to saving
as many lives as possible. Many researchers have analyzed this technique using Convolutional neural
network, HDNN, RCNN, etc. This paper will give us an overview of various techniques or methods that
are used to detect accidents.
Keywords- Accident detection system, Convolutional neural network, HDNN, RCNN.

I. INTRODUCTION
In the report of NCBI [1], it is clearly stated that there have been 464,910 accidents claimed and also
it had been stated that there are daily 405 deaths and 1290 injuries thanks to road accidents. Also quite 30% of
those deaths are thanks to late medication or late awareness of the accident. The worldwide economic loss
caused by road traffic accidents annually is estimated at 518 billion US dollars [2]. With the increase in the
vehicles, the accidents due to vehicles have also been increased. As since the accident is detected through
traditional information and communication technologies, Ad hoc Networks and random forest Classifier [11].
But now with the help of machine learning and deep learning-based techniques, various models are being
implemented for the detection of vehicle accidents and their severity with the help of a considerable amount of
footage and images. Since accident rates are increasing rapidly, there is an urgent need for researchers to
develop a system that would detect such accidents through videos. This paper focuses on various techniques like
the deep learning-based RCNN model [6][7][8][12], CNN model [9][13]. Also, many different approaches
such as considering the manipulation in the edge detection [5], GPS module [14].

II. L ITERATURE S URVEY

Md. Farhan Labib, et.al. [3], Analysis has been done, by using Decision Tree, K-Nearest Neighbors
(KNN), Naïve Bayes, and AdaBoost these four supervised learning techniques, the severity of
accidents into Fatal, Grievous, easy Injury and Motor Collision these four classes. Eventually, the
least complex execution is accomplished by AdaBoost. In this paper, the author characterized the
force of accidents into four categories Fatal, Grievous, Simple Injury, and Motor Collision. To sort
1
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
the seriousness of accidents in these four classes, eleven primary factors that influence the most
extreme number of accidents in Bangladesh are chosen as the features. Previously occurred, virtually
forty-three thousand traffic accidents knowledge in People's Republic of Bangladesh from 2001 to
2015 area unit used as our learning materials. Among these four techniques best performance is
achieved by AdaBoost and its accuracy was 80 percent. The machine learning concept remains
within the formative stage here. In the paper, the authors proposed a way to detect the likelihood of
accidents on the road by using vision-based techniques.

Nikhil Sharma, Rajat Rathi, et.al.[4], Loop detectors and automatic Vehicle Location sensors are
tools that may sensor real-life traffic but the installation of those tools on urban roads may be a myth.
As large-scale expertise in traffic, modelling isn't required therefore data driven techniques are
approximately faster to create. With a really large dataset of images, at some point it's
difficult to realize high accuracy and hence Researchers proposed to make some new layers of
net then integrate them with the old layer’s architecture net. this HDNN hybrid model of
CNN and SVM model. HDNN was applied on two datasets one among images of traffic
conditions and second of road accidents in India from different areas. The information set of
traffic pictures consists of categorical data, distributed in four categories: Dense Traffic, distributed
Traffic, Accidents and fireplace. On the dataset of traffic images, they applied convolutional layers
with activation function, the main purpose of convolution layers is to seek out features in images
using feature detectors. After every convolutional layer applied a pooling layer to preserve features
from any quiet distortions and Flatten layer which reduces all dimensions of input to at least one
dimension. Then the flattened matrix goes through a totally connected layer for classification of
traffic flow.

Elijah Blessing Rajsingh, Salaja Silas, et.al.[5] ,Accident images are the significant data for
analysing the severity of collision in road accidents.Edge detection algorithms locate the
significant change of pixel intensity so as to spot the objects in a picture. the sides are
typically the border line among different regions in a picture. The challenge of an edge
detection algorithm is to stipulate the important edges appropriately and have less possibility
of erroneously considering the non- edges as a foothold. Filtering eliminates the unwanted
noise from a picture and scale image coordinates. Image en hancement provides better contrast
and more detailed information about the image than the original image.Image enhancement is
performed by computing gradient magnitude. The desirable edge detection algorithms satisfy the
subsequent criteria Good detection: the probability of neglecting real edges must be minimal. Good
localization: the sting detected by the sting detection algorithm must be on the brink of the important
edges. Minimal response: it must return just one response for every true edge. Therefor e the
performance of those algorithms were evaluated in terms of PSNR (Peak signal to noise ratio) and
MSE(Mean square error). Higher the PSNR value and lower the MSE value indicate the specified
edge detection algorithm. The experimental results show that the Sobel edge detection algorithm is
acceptable for edge detection in the accident images.

Kyu Beom Lee, Hyu Soung Shin [6] Object Detection and Tracking System (ODTS) along with a
notable deep learning organization, Faster Regional Convolution Neural O rganization (Faster R-
CNN), for detection and chase algorithmic rule square measure attending to be introduced and
applied for automatic detection and watching of sudden events on CCTVs in tunnels, that square
measure possible to (1) Wrong-Way Driving (WWD), (2) Stop, (3) Person out of an automotive
vehicle in tunnel (4) hearth. ODTS accepts a video frame time as associate degree input to urge
Bounding Box (BBox) results by Object Detection and compares the BBoxs of this and former video
frames to assign a unique identification number to every moving and object to be detected. This
framework makes it conceivable to follow a moving object in time, which isn't normal to be
accomplished in conventional object detection frameworks. Deep learning model in ODTS wa s
prepared with a dataset of event images in tunnels to Average Precision (AP) estimations of 0.8479,
0.7161, and 0.9085 for target objects: Car, Person, and Fire, etc At that point, supported trained deep
learning model, the ODTS based tunnels surveillance accidents Identification framework was tried

2
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
utilizing four accident recordings which including every accident. This object following technology
has been with success utilized for the tracing of targeted pedestrian and also the moving vehicle,
accident observation in traffic camera, criminal and security observation within the sure native space
of concern

Murugan. V, Vijaykumar V.R, and Nidhila. A. [7] , The object detection technique is Gaussian
Mixture Model (GMM) The outcome of the detection is the i ntended region of interest, i.e. moving
vehicles are the proposed region of interest, which were identified using background subtraction
framework. Identification of these regions of interest from the input frames defines the region
proposals in the region based CNN(Convolutional neural network). and the detected vehicles are
recognized as variant classes using an SVM classifier. The efficiency of this algorithm is evaluated
by determining the deviation in pixel values concerning ground truth pixels. Backgr ound subtraction
is one of the most promising algorithms, which is then followed by kalman tracking and
recognition by using the ANFIS classifier. Deep learning framework RCNN is used for the
recognition of various classes of vehicles. RCNN architecture includes three major phases i.
Capturing the region proposal. ii. CNN layers for feature extraction iii. Classifying the regions
to recognize the variant classes. Performance metric values acquired show the proficiency of the
RCNN with the state of art techniques.

Vipul Gaurav, Sanyam Kumar Singh, et.al.[8],the system provides a whole resolution to a distinctive,
police investigation and mitigating concerning accidents and additionally provides the feasibility study to
spot accident-prone areas across the Republic of India. The regression model outperformed its counterparts
to predict the number of potential accidents that will occur in ANy given state across the Republic of
India with an accuracy of 80- ninetieth. The image segmentation of the accidents is handled by the cloaked
R-CNN model that gives correct instance segmentation higher than the quick R-CNN model which will
provide solely linguistic segmentation and can't differentiate between the categories.The author
experimented with numerous image augmentation techniques and located watershed segmentation and
smart edge detection to be appropriate for distinctive the region of interest of any automotive accident
pictures or footage. They additionally created the use of electric shock sensors in extremely prone areas
wherever the frequency of accidents is extraordinarily high like Mumbai-Pune main road to observe the
shock generated throughout a collision and this acts as an additional validation step in predicting the
prevalence of road accidents with high severity.The alert system is predicated on the outputs of the image
segmentation, electricity sensors, and also the accident severity classification model that is integrable
with any monitor. This method offers America an inspiration that whereas detection of car Rcnn
incorporates a higher edge over quick Rcnn the accuracy or rcnn detects well as compared to quick Rcnn.

Bulbula Kumeda, Zhang Fengli, et.al.[9] , the detection of accidents and potential algorithms for self
acting image processing CNN is used. Tasks like determining the objects in an image, image localization,
and other pattern acknowledging tasks.Grouping of images using CNN techniques and categorize images
into classes that are Accident, Traffic, Flame, and low traffic. To identify the incidence of the accident,
various classes like drunk and drive, month, and weather during the accident, human factors, and light
conditions, etc are considered. Task perform,i.e used a traffic accident image dataset called Traffic-Net.
The CNN layer is the main block of CNN. The layer consists of neurons that appear for particular features
which will later make the neurons more active.

These determine the characteristics of the given image parameters like filters and kernel are
associated.Relu layer is known as the rectified linear unit which is the next step to the convolution layer.
The fully connected layer which connects the entire series of all layers. In this way, CNN works for the
detection of accidents through a dataset. From the images, the training loss could be seen depicting the
converges of the cost function. Miao Chong, Ajith Abraham et.al.[10] , the purpose that industry has tried
to set up and build safer cars, however traffic accidents are inevitable. Thus for these issues several
approaches like a neural network, nesting logic formulation, log-linear model, fuzzy ART maps. Neural
networks trained by mistreatment the hybrid learning approaches, support vector machines, call trees, and
a synchronic hybrid model involving call trees and neural networks. In a hybrid learning approach, a
3
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
Multilayer Perceptron (MLP) may be a feed forward neural network with one or additional hidden layers.
The network consists of an associate degree input layer of supply neurons, a minimum of one or additional
hidden layer of machine neurons, associate degree an output layer of machine neurons.The input layer
accepts input signals and redistributes these signals to all neurons in the hidden layer. The output layer
accepts a stimulus pattern from the hidden layer and generates the output pattern of the entire network.
Decision tree algorithm used a Classification and Regression Trees (CART) model which consists of a
hierarchy of univariate binary decisions. Hybrid decision trees integrate different learning models and give
efficient performance than the individual learning or decision-making models by decreasing their
individual limitations and exploiting their different mechanisms.

Nejdet Dogru, Abdulhamit Subasi [11],Real-time traffic management system which employs IoT
devices and sensors to collect real-time traffic information. The collected traffic data are then used to
analyze traffic density .ANNs motive to duplicate the behavior of a real neural network which consists
of large number of interconnected neurons.Support vector machines (SVMs) are recently developed and
attracting attention from researchers because of their notable accuracy and ability to manage with large
and high dimensional data sets. Random forest (RF) is a data mining tool to solve classification and
regression related mathematical problems.The performance of RF algorithm, in terms of its accuracy,
was found superior to ANN and SVM algorithms.

Helen Rose Mampilayil, Rahamathullah K.[12] ,The paper proposes a system which automatically
detects the one-way traffic rule violation without the assistance of human beings. The proposed
system can detect the 3 wheeler automobiles which violates the one-way traffic rule. Modules of the
proposed system are capturing the input stream, pre-processing the video, 3 wheeler vehicle
detection and tracking and finally the direction calculation. The input video stream from the digital
camera that's constant on the roadside. It is needed to initialize the route of the traffic so one can
detect the incorrect route of vehicles. In the pre-processing step the input video is transformed into
frames for video processing. Deep learning approach can be used to detect 3 wheeler vehicles and to
identify the one-way traffic rule violation. The trajectory factors of the vehicle are used to decide the
direction to which the vehicle is moving. If the route is opposite to the one-way traffic direction,
then it may be taken into consideration as rule violation.

Monagi H. Alkinani, Wazir Zada Khan, et.al.[13] , This paper talks about the reasons and consequences
of any other human volatile conduct known as Aggressive Driving behaviour (ADB). Aggressive driving
Behaviour (ADB) is an extensive organization of risky and competitive using patterns that result in
severe accidents. Human abnormal driving behaviours are affected by various factors including drivers
experience/inexperience of driving, age, and gender or illness. The study of these elements which
could result in deterioration in the driving skills and overall performance of a human driver is out of the
scope of this paper. After describing the history of deep learning and its algorithms, it presents an
in-depth research of latest deep learning-based systems, algorithms, and strategies for the detection of
Distraction, Fatigue/Drowsiness,and Aggressiveness of a human driver, achieves a complete knowledge
of HDIADB detection through offering an in depth comparative analysis of all the latest strategies.
Moreover, it highlights the essential requirements.

Siddharth Tripathi, Uthsav Shetty, et.al.[14] ,In this paper an intelligent device known as CBITS
has been proposed. It is like a Holistic method as it's far properly ready with a network of
sensors that offer real-time emission tiers as properly because it signals. .A Unique Vehicle
Identification Number can be assigned to each automobile via which the motors are recognized
through the tracking authorities. The collective statistics can be shared to the cloud the usage of the Wi-
Fi port of Raspberry Pi. CBITS is a relatively powerful, real-time, mild weight, reliable, low-power
consuming and a value effective device for the automobile-proprietors in addition to the tracking
government. GSM and GPS modules are used for sharing the location to the cloud. Raspberry Pi (B
Plus) is the motherboard used, which has 40 GPIO pins and four USB slots which makes the interfacing
easy. CBITS combines the different processes into one single and holistic device in addition to its
functionalities which show to be highly value-effective and fast responding. CBITS makes use of the
cloud-computing platform known as Thingspeak, beneath the class of Software as a Service (SaaS).
4
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)

III. ANALYSIS TABLE

Table 1 shows the analysis of survey of the existing system by stating the
Title of Paper, Technology Used, Dataset used and Accuracy.

Table 1 analysis survey of the existing system

Sr. Title of Paper Technology Used Datasets Accuracy
No.

1 Road Accident AdaBoost ARI of BUET 80%

Analysis and (Accident Research
Prediction of Institute of
Accident Severity Bangladesh
by Using Machine University of
Learning in Engineering and
Bangladesh[3] Technology
2 Traffic Density HDNN Images of 0.7429(epoch 1)
Investigation & four different 0.7830(epoch 5)
Road Accident traffic
Analysis in India scenarios
using Deep
Learning.[4]
3 Performance Sobel edge detection Images of various MSE:- 197.77
Analysis of Edge accidents PSNR:- 50.368
Detection
Algorithms for
Object Detection
in Accident
Images.[5]
4 An application Faster R-CNN This dataset is Ranging from 0.7
of a deep composed of 70,914 to 0.8
learning video images by
algorithm for dividing 45 videos
automatic into frames.
detection of
unexpected
accidents under
bad CCTV
monitoring
conditions in
tunnels.[6]
5 A Deep R-CNN 3 types of dataset 80%-90%
Learning R-CNN with their frame
Approach for counts
Vehicle
Recognition in
Traffic
Surveillance
System[7]

5
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
6 Accident Masked Rcnn Provided by the 80%-90%
Detection, Government of
Severity

Prediction, India
Identification of
Accident Prone
Areas in India and
Feasibility Study
using Improved
Image
Segmentation,
Machine Learning
and Sensors[8]
7 Vehicle accident CNN Traffic Net Approx 80%
and traffic 90%
classification
usingDeep
Convolutional
Neural Network
[9]
8 Traffic Accident ANN (Hybrid learning GES (General 85%-90%
Analysis Using approaches , support estimation system)
Machine Learning vector machines, automobile
Paradigms[10] decision trees) accident data.
9 Traffic Traditional Bootstrap sample 88% to 89.91%
Accident information and and Out of bag
Detection Using communication (OOB) dataset
Random Forest technologies, Ad hoc
Classifier.[11] Networks and
random forest
Classifier.
10 Deep learning Deep Learning N.A Upto 85%
based Detection of
One Way Traffic
Rule Violation of
Three Wheeler
Vehicles.[12]
11 Detecting Human Deep Learning (CNN, COCO datasets 92%
Driver Inattentive DBN).
and Aggressive
Driving
Behaviour using
Deep
Learning: Recent
Advances,
Requirements and
Open Challenges.
[13]

6
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
12 Cloud Based Holistic method Accelerometer N. A
Intelligent Traffic (Network of sensors). and Force-
System to GPS and GSM resistive sensors
Implement Traffic module. Value > 5.0
Rules Violation
Detection and
Accident Detection
Units. [14]

The various algorithms and features used for accident detection are analyzed in the above table. It
includes the Machine Learning and Deep Learning-based technique used for Real-time object detection.
Even there is a various model which uses the IoT based technique which includes various hardware
components and software connectivity, From the analysis table above and studying it can be seen that
best algorithm object detection are CNN algorithm and RCNN gives better results and accuracy while
performing and hence using an RCNN or CNN based model would provide a better performance as
compared to other algorithms used for the same problem.

IV. CONCLUSION
Since there is a major growth in accidents of vehicles and deaths related to it, it is very vital to detect such
incidents. From the literature survey, we came to know various models use various machine learning and deep
learning-based techniques such as AdaBoost, HDNN, CNN, ANN with different features to detect the accident
and locate the incident in the footage. But the above-mentioned papers do not have good accuracy and alerting
systems to the nearby emergency unit. From different studies of paper, we came to know the Faster RCNN gives
better results and accuracy.

Acknowledgements
We would like to express a deep sense of gratitude towards the Computer Engineering Department of
VIVA Institute Of Technology for their constant encouragement and valuable suggestions. The work that
we are able to present is possible because of their timely guidance. We would like to pay gratitude to the
panel of examiners Prof. Sunita Naik, Prof. Dnyaneshwar Bhabad, & Prof. Vinit Raut for their time,
effort they put to evaluate our work and their valuable suggestions time to time. We would like to thank
Project Head of the Computer Engineering Department, Prof. Janhavi Sangoi for her support and
coordination. We would like to thank the Head of the Computer Engineering Department, Prof. Ashwini
Save for her support and coordination.

REFERENCES
[1] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6482791/ ( last accessed on 30-01-2021)

[2] https://www.who.int/violence_injury_prevention/publications/road_traffic/world_report/summary_en_rev.pdf(last
accessed on
30-01-2021)
[3] Md. Farhan Labib, Ahmed Sady Rifat, Md. Mosabbir Hossain, Amit Kumar Das, Faria Nawrine, "Road Accident Analysis and
Prediction of Accident Severity by Using Machine Learning in Bangladesh", IEEE 2019.

[4]Nikhil Sharma, Rajat Rathi, Chinkit Manchanda,"Traffic Density Investigation & Road Accident Analysis in India using Deep Learning"
IEEE 2019

[5]Elijah Blessing Rajsingh, Salaja Silas , P.Joyce Beryl Princess,"Performance Analysis of Edge Detection Algorithms for Object
Detection in Accident Images", IEEE 2019.

[6]Kyu Beom Lee, Hyu Soung Shin,"An application of a deep learning algorithm for automatic detection of unexpected accidents
under bad CCTV monitoring conditions in tunnels" IEEE 2019.

7
www.viva-technology.org/New/IJRI
VIVA-Tech International Journal for Research and Innovation Volume 1, Issue 4 (2021)
ISSN(Online): 2581-7280 Article No. X
PP XX-XX
VIVA Institute of Technology
9th National Conference on Role of Engineers in Nation Building – 2021 (NCRENB-2021)
[7] Murugan. V, Vijaykumar V.R and Nidhila. A,"A Deep Learning RCNN Approach for Vehicle Recognition in Traffic
Surveillance System" IEEE 2019.

[8]Vipul Gaurav, Sanyam Kumar Singh, Avikant Srivastava,"Accident Detection, Severity Prediction, Identification of Accident
Prone Areas in India and Feasibility Study using Improved Image Segmentation, Machine Learning and Sensors", IJERT 2019.

[9] Bulbula Kumeda, Zhang Fengli, Ariyo Oluwasanmi, Forster Owusu, Maregu Assefa, "Temesgen Amenu,Vehicle accident and
traffic classification using Deep Convolutional Neural Network", IEEE 2019.

[10]Miao Chong, Ajith Abraham and Marcin Paprzycki1,"Traffic Accident Analysis Using Machine
Learning Paradigms", IEEE 2017.

[11] Nejdet Dogru, Abdulhamit Subasi,"Traffic Accident Detection Using Random Forest Classifier", IEEE 2018.

Helen Rose Mampilayil, Rahamathullah K,"Deep learning based Detection of One Way Traffic Rule Violation of Three
[12]
Wheeler Vehicles", IEEE 2018.

[13]Monagi H. Alkinani, Wazir Zada Khan, Quratulain Arshad,"Detecting Human Driver Inattentive and Aggressive Driving
Behaviour using Deep Learning: Recent Advances, Requirements and Open Challenges", IEEE 2020.

[14]Siddharth Tripathi, Uthsav Shetty, Asif Hasnain, Rohini Hallikar,"Cloud Based Intelligent Traffic System to Implement Traffic
Rules Violation Detection and Accident Detection Units", IEEE 2019.

8
www.viva-technology.org/New/IJRI