2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN)

Accident Detection Using Deep Learning

2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN) | 978-1-7281-8337-4/20/$31.00 ©2020 IEEE | DOI: 10.1109/ICACCCN51052.2020.9362808

Durgesh Kumar Yadav Renu Ankita Iftisham Anjum

Department of CSE Asst. Professor, CSE Department of CSE Department of CSE
GCET Greater Noida, GCET Greater Noida, GCET Greater Noida, GCET Greater Noida,
UP, India 201306 UP, India 201306 UP, India 201306 UP, India 201306
[email protected] [email protected] [email protected] [email protected]

Abstract— Every year around 1.35 million people are cut approx. 13 people die each hour as a victim to it. Looking to
off due to numerous crashes in case of road traffic accident. As the present scenario about 140,000 people die per year
per the statistics 20 to 50 million people suffer as a result of its making India being the top on the list of losing lives due to
injuries. As a consequence of such traffic accidents people pays road traffic accidents. Such accidents undergo in three
off their lives. These conditions are caused due to the lack of
co-ordination among the organizations involving in it. Also not
phases.
properly practising the rules and ways as it to be followed First phase of an accident is considered with the last few
magnifies the graph upwards. The risk factors include minutes or seconds of the victim’s death by the accident, as
speeding, drink and drive, distraction in driving, bad of 10% of such deaths contributed in it. The Second phase
infrastructure, in-proper vehicles, breaking rules and many holds the time of an hour of the accident that is with the
more. As such a system is needed which is perfectly able to co- death percentage as (75% of all). Third phase of it holds 15%
ordinate between the different actions that is to be taken for as a death rate and the duration is a week or more [2].
the quick response at the accident location. As per the research The objective being is to directly the help in critical hour of
such detection system involves different technologies such as need that makes a difference to their lives. Not only the
Global Positioning System [GPS] & Global System for Mobile
Communication [GSM], applications of mobile phones, etc. All
victim but their families suffer its consequences. In order to
the vehicles are included under these detection systems and make difference in lives we need to take proper steps with
other technologies are also considered for the same. As this quick response to the people. We can also join in with the
paper represents an overview related to the technologies that various organizations to accomplish it. Government also
interconnected with that of road accidents by automated road takes important measures along with several big and small
[traffic] accident detection system. organizations in order to tackle such situations.
Nowadays Road Accidents have become very common and
Keywords—Accident Detection, Long Short-Term Memory, the reason behind it are population growth, increasing
Convolutional Neural Network, RNN.
number of automobiles and modes of transport have also
I. INTRODUCTION increased.
Road accidents being the major factor in case of losing lives Many horrible accidents occur daily and these accidents
these days. With the increase in the rate of manufacturing of causes a terrible loss not only to the lives but materialistic
the automobiles, the rate of deaths due to road accidents also things too. The craziness for the speed is one of the main
increases. With going days people are being more and more reasons of human beings for the accidents. In our country
careless which causes severe consequences, the death rate is hundreds of accidents occurs daily.
also large in the large cities. Today there an availability of Road accidents happen due to following reasons, distraction
various types and modes of transports as adding a plus point of drivers while behind the wheel, taking their focus off
to the road accidents. Roads seems more crowded day by from road, long hours of drive also make drivers tired,
day causing traffic. They damage both the lives and the resulting in preventable errors, combination of bad visibility
modes. More attention, awareness is needed among the to unsafe road design. While the reasons of accidents can be
people. People with no vehicles, walking on streets are also different but the results are often the same heading with a
not safe. As with every passing day its going worst and great loss.
increasing every hour. Around 3000 car accidents occur every day around the
Around to 1.35 million people lose their lives due to road world. Drugs and alcohol are found to be contributing cause
accidents with every passing year, according to the facts of in up to 22% of vehicular accidents on the world’s highway
around 20 to 50 million of people comes under the and byways.
consequences of the road accidents. They go through
various injuries caused by it leading to different major
disabilities. The World Health Organization [1] (WHO) also
stated regarding the financial status of the countries, the
developing countries witness more loss as compared to non-
developing countries and among them the poor and middle-
class people get the worst. The developing countries holds
higher loss-rate than that of the developed countries. As the
developing countries having less art mobiles still losses
more than the developed countries having more automobiles Fig. 1. Crash Report
in number. Talking about the India, as per the reports

232
ISBN: 978-1-7281-8337-4/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on November 02,2022 at 16:58:24 UTC from IEEE Xplore. Restrictions apply.
 2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN)

Risk factors for Road traffic accidents be as human factor, units, and each layer consisting of around 256 hidden LSTM
speed, alcohol and drugs, seat belts, visibility, distributed units, followed by the other dense layers.
driving and exhaustion. Such a system that is able to provide
an immediate medical assistance to the victim at the correct Results: The Accuracy metric measures the performance.
accident location can probably reduce the rate of death to a As running model, minimum of 82% and maximum of
greater extent. Therefore, the idea of an alert or detection 98.76%. And average accuracy goes over 92.38%.
system comes into mind and further into existence too
which, after sensing the probability of accident, will alert
the respective guardians and to medical aids by sending a
SOS emergency notification or text message to nearby
medical facility or say hospital and to the nearest police
station.

Fig. 2. Road Fatalities in India.

The main moto is to incorporate or propose such a system

that is properly able to detect an accident from the available
video footages provided through the cameras. The system
behaves as a tool that works out at the time of need along
with the quick response time to the accident location also
including the authorities for the same. The main focus is to
have a short or quick response time in order to implement
the further things early as possible. For this we are using
advanced Deep Learning Algorithms.
II. RELATED WORK
In this section, we compared various work study upon the
topics related to accident detection techniques. Aiming at
the tangible infrastructure in spite of the Intelligent
Transportation Systems (ITS) that includes the traffic
congestion detection, accident detection ability, to detect the
occurrence of various events and many more. Several methods
have been proposed by different researches and scientists in Fig. 3. Architecture of the System
respect to automatic detection of accident. The techniques
they include are as detection of road accidents and the Limitations: A vanishing gradient problem causing
accident location via smart-phones, VANET (Ad-hoc problems in feature extractions that directly affects the
networks), GPS and GSM technologies, mobile applications accuracy of the overall system.
and also the deep learning models. B. Object Detection and Tracking System
A. CNN with LSTM In [3], the proposed system uses ODTS (Object Detection
In [2], a system introduced to detect accidents in the videos. and Transfer System). It connects with the live CCTV feeds
In this technique, the various deep learning concepts are to the inside tunnel and tracks the live feeds. The sections of
used. Including the Convolutional Neural Network with the live feeds are cropped and then detects each object in the
LSTM units has been used in order to train a model that frame. The detection process and tracking system over the
could detect accidents in videos as required. time. By the algorithm class and B-Box were attained, it
Convolutional Neural Network (CNN): Featured predicts the next position and the grants ID.
with shared-weights architecture and translation Results: The model is trained with the dataset of the event
invariance. images in tunnels with that of Average Precision (AP) of
values 0.8479, 0.7161 and 0.9085 for the targeted objects: 4-
In this architecture, two convolutional layers being used to whealer, Person, and Fire, respectively.
support featuring of extraction and then 2 layers of LSTM

233

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on November 02,2022 at 16:58:24 UTC from IEEE Xplore. Restrictions apply.
 2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN)

The library has over 2500 structured algorithms, which

includes a comprehensive set of computer and state-of-
the-art computer and machine learning algorithms.
These algorithms are used to find and recognize faces, to
view objects, to segment human actions in videos, to
track camera movements, to track objects, to extract 3D
object models, to generate 3D point clouds from stereo
cameras, and put images together to produce high
resolution picture of the whole event, find similar
pictures in the photo database, get the red eyes out of the
photos taken with flash, follow the eye movement,
recognize the beauty of the place and establish a tag to
attach it to unpleasant truth, etc. OpenCV has over 47
thousand public users and an estimated download
volume of over 18 million. The library is extensively
Fig. 4. Interfaced Controller with modules
used by research groups, companies, and government
Limitations: Improper visibility leads to low object agencies [6].
detection. Reliability also needs improvement. Keras
C. GSM and GPS model Keras is easy to use and free open source neural network
In [4], a module is proposed that detects, and identifies library which is written in Python that runs on
accidents on the occurrence through means of vibrations and TensorFlow. It is designed to be fast, modular and easy
sends a notification to the given number. The hardware’s to use. It is developed by Google engineer Francois
included at the time of an accident is a vibration sensor that Cholet.
is triggered due to vibration, the GPS module is then fetched The Carus API means to build models and define layers,
and the location of the incident is send as a message to the or set multiple input and output models. At this level,
provided mobile number with the help of GSM module that Keras compiles the model with the Optimizer and Loss
is inbuilt into the particular system. The GPS and GSM function, the training process with the Fit function.
module shows all the sensors that are joint to a central Carus does not handle low-level APIs such as creating
Arduino board, that is in control of the overall system. computational graphs, tensors, or other variables
because it is run by the "backend" engine [5,8].
TensorFlow
Tensor Flow is an end-to-end open source platform for
machine learning. It is a flexible ecosystem of tools,
libraries and community resources which is
comprehensive, that helps research scholars to reach the
edge in Machine Learning, and developers can easily
build and implement machine learning driven
applications.
Tensor Flow offers many levels of abstraction from
which you can choose the right one you needs. Build
and train models using high-level KRA APIs that make
Fig. 5. Block Diagram of module
it easy to get started with Tensor Flow and Machine
Learning.
Results: The location area is detected and informed.
Limitations: The hardware aren’t always much reliable B. Dataset
specially the sensors. The GSM Module might take time to The temporal locations of accidents were manually
send a message. annotated. We have a total of 1730 accident videos and
each video is of 4 seconds, which is further divided into
III. IMPLEMENTATION 100 frames of each. The dataset contains 620 videos
positive containing the exact time of accident is in the
A. Software Used last 15 frames, and 1130 negative videos containing no
OpenCV accidents with high video quality (720x1280 resolution).
The dataset was split into training and testing dataset
OpenCV (Open Source Computer Vision Library) is a that is 75% and 25% respectively: 1284 training videos
huge open source library which is used for image and 466 testing videos. The dataset is collected from six
processing, machine learning and computer vision. major cities in Taiwan. The dataset includes: 42.6%
OpenCV was built to provide an infrastructure for motorbike to car collision, 19.7% car to car collision,
computer vision applications and to accelerate the use of 15% motorbike to motorbike collision, and remaining
machine vision in commercial products. OpenCV makes 22.7% are of other types.
it easy to use for businesses and modify the code.

234

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on November 02,2022 at 16:58:24 UTC from IEEE Xplore. Restrictions apply.
 2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN)

TABLE I. DISTRIBUTION OF DATASET is used for encoding of the first NN to reflect either a video
Positive Negative Total contain accidents or not.
Videos Videos Through this strategy, the HRNN joins a time reliance of
Training Dataset each video outlines inside each video to anticipate how
455 829 1284
likely it may be a modern video contains a car mischance.
Testing Dataset
165 301 466 LSTM (Long Short-Term Memory) is used here to control
Total the time dependency of each video.
620 1130 1730 The training dataset is further divided into validation and
train dataset, 40% and 60% respectively of the total training
dataset. The training dataset is used to train the model and
C. Implementation the validation dataset is used to validate the model while
training process.
IV. RESULTS
The Model was saved with an accuracy of 80% validation
accuracy as Model_final.hdf5. Now the checkpoint file is
used to test the trained model. And the test accuracy of the
model on the testing dataset was 92% which is a quite good
result for our model. Validation Accuracy = 80%.
REFERENCES
[1] Renu, Durgesh Kumar Yadav, Iftisham Anjum, Ankita, “Accident
Detection Using Deep Learning: A Brief Survey”, International
Fig. 6. Work Flow of Model Journal of Electronics Communication and Computer Engineering
(IJECCE), vol. 11, no. 3, pp. 16-25, May 2020.
Each of these pictures could be a 2D cluster of pixels [2] Ghosh, S., Sunny, S.J. and Roney, R., 2019, March. Accident
(1280x720) where each pixel has data around red, green, detection using Convolutional Neural Networks. In 2019 International
and blue (RGB) colour levels. To decrease the Conference on Data Science and Communication (Icon DSC) (pp. 1-
6). IEEE.
dimensionality at the person picture level, we change
[3] Lee, K.B. and Shin, H.S., 2019, August. An application of a deep
over the 3-D RGB colour clusters to grayscale. Moreover, to learning algorithm for automatic detection of unexpected accidents
form the computations more efficient on a CPU, we under bad CCTV monitoring conditions in tunnels. In 2019
down test case of each picture by a calculate of 5 - International Conference on deep learning and machine learning in
in impact, averaging each five pixels emerging applications (Deep- ML) (pp. 7-11). IEEE.
to decrease the measure of each picture to a 2-D cluster of [4] T Kalyani, S Monika, B Naresh, Mahendra Vucha., 2019, March.
Accident detection and alert system. In 2019 International Journal of
256x144 (as contradicted to the starting 3-D cluster of Innovative Technology and Exploring Engineering (IJITEEE) ISSN:
1280x720x3). (As sequence of events is depicted in the 2278-3075.
below figure 7). [5] François Chollet (2020), Keras [Online], Available:
https://keras.io/about/
[6] Willow Garage (2020), OpenCV [Online], Available:
https://opencv.org/about/.
[7] Prabakar, S., et al. "An enhanced accident detection and victim status
indicating system: Prototype." India Conference (INDICON), 2012
Annual IEEE. IEEE, 2012.

Fig. 7. Different Frames of Video

Pre-Processing Videos and Images

Working with video has many challenges and one of the
biggest challenges is the amount of data. Number of videos
are not much more but when these videos are broken into
frames these are many images over 25,000 individual
images and each frame is broken up into its individual
frames which is to be analysed separately.

Training and Validation

The algorithm works with two layers of short-term memory
NN. The time-dependent sequence of the images in the
video are analysed by the first neural network and the other

235

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on November 02,2022 at 16:58:24 UTC from IEEE Xplore. Restrictions apply.