© 2023 JETIR April 2023, Volume 10, Issue 4 www.jetir.

org (ISSN-2349-5162)

Accident Detection Using Convolutional Neural

Networks
Chadalavada Lavanya Dr.K.Santhi Sri
MCA Associate Professor
Department of IT&CA Department of IT&CA
VFSTR Deemed to be University, VFSTR Deemed to be University,
Guntur,AP, India Guntur,AP, India

Abstract— to develop a CNN model for accident detection, a changes in speed or direction of vehicles, or the presence of debris
large dataset of accident and non-accident images will be on the road. Another approach could be to use sensors or cameras
required for training and testing the model. The dataset should installed on vehicles or on roads to detect accidents in real-time.
be diverse and cover a range of accident scenarios to ensure that Such systems can automatically detect accidents and alert
the model is robust and can detect different types of accidents. emergency services and nearby drivers to take appropriate action.
Once the CNN model is trained, it can be used to process the live It is important to ensure that such systems are accurate and reliable
video feed from the CCTV camera installed on the highway. to avoid false alarms, which could lead to unnecessary panic and
Each frame of the video can be passed through the CNN model confusion. The system should also be designed to ensure privacy
to classify it as an accident or non-accident frame. If an accident and security of the footage, and comply with relevant regulations
is detected, an alert can be sent to the nearest emergency services and laws. Incorporating a system to detect accidents from video
or to a central control room to initiate the required rescue footage can help save lives and reduce the impact of accidents on
operation. One potential challenge with this approach is the need individuals and society.
for real-time processing of the video feed to ensure that accidents
are detected promptly. This may require high-performance The approach of using deep learning algorithms, specifically
computing hardware and optimized software algorithms to convolutional neural networks (CNNs), to analyze video footage
ensure that the CNN model can process the frames of the video for accident detection is a promising one. CNNs are particularly
in real-time. The proposed system to detect accidents based on the well-suited for image and video processing tasks, and can be
live feed of video from a CCTV camera using a deep learning trained to recognize patterns and features that are indicative of
convolution neural network model is a promising approach. accidents. Setting up this system on highways where timely help
CNNs have indeed been shown to be highly effective for image reaching the accident victims is rare is a sensible strategy, as it can
classification tasks and have been successfully used in many help to reduce the time between the occurrence of an accident and
applications, including object detection and recognition. In the arrival of emergency services. By installing CCTV cameras at
summary, the proposed system using a CNN-based model to intervals of about 500 meters, the system can cover a large area
detect accidents based on live video feed from CCTV cameras has and monitor traffic flow in real-time. The use of Raspberry Pi 3 B+
great potential to reduce the number of accident-related deaths in as a portable and remote computer to be set up on a CCTV camera
India by enabling timely help to reach accident victims. However, is a cost-effective and practical solution. The Pi Camera can be
the development and implementation of such a system will directly attached to the Raspberry Pi to capture video footage,
require significant resources and expertise in computer vision, which can then be analyzed by the CNN-based accident detection
deep learning, and real-time processing. (Abstract) model. It is important to ensure that the system is accurate and
reliable, and can detect accidents within seconds of them
Keywords— Convolutional Neural Network; Accident happening. This requires careful design and training of the CNN
Detection; Deep Learning; Video Classification; Recurrent model, as well as rigorous testing and validation to ensure that it
Neural Network performs well under different conditions and scenarios. Overall,
the proposed system has the potential to significantly improve the
I. INTRODUCTION safety of highways by detecting accidents in real-time and alerting
emergency services to take appropriate action.
Each year, road accidents result in over 1.3 million deaths globally,
with an additional 25 to 65 million people sustaining mild injuries. That sounds like a great application of machine learning and
Shockingly, developing countries bear the brunt of these fatalities, computer vision to help detect accidents. It's impressive that you
with a death rate of 23.5 per 100,000 population - significantly were able to achieve such high accuracy with your Inception v3
higher than the rate of 11.3 per 100,000 in high-income countries. model. Implementing the model on a Raspberry Pi using
This is evident in the World Health Organisation's (WHO) survey, TensorFlow, OpenCV, and Keras is also a great way to make it
which highlights that more than 90% of road traffic related deaths accessible and portable. Running each frame of the video through
occur in developing countries, despite these countries possessing the model and predicting whether it's an accident frame or not is a
only half of the world's vehicles. India, for instance, has recorded good approach. However, it's important to keep in mind that even
an alarming average of 13 deaths every hour due to road accidents, with high accuracy, there is always the possibility of false positives
making it a potential candidate for becoming the top country with or false negatives. It's crucial to evaluate the model's performance
the highest number of road accident deaths. Unfortunately, this on a diverse set of test data to ensure its reliability before
number may be an underestimation, considering that numerous deploying it in real-world scenarios.
cases of accidents are not reported. Overall, a road accident can be
split into three phases, each presenting an opportunity to rescue This project aims to not only detect accidents but also notify the
victims. relevant authorities in real-time, providing them with important
information such as the timestamp, location, and frame at which
The objective of incorporating a system to detect accidents from the accident was detected. The use of a GSM module to send
video footage is a noble one as it can help save lives by alerting notifications is a good way to ensure that the authorities are alerted
authorities and emergency services in a timely manner. There are promptly. The emergency light that lights up when an accident is
various ways in which such a system can be designed and detected can also be helpful in alerting other drivers and
implemented. One approach could be to use computer vision pedestrians in the area to be cautious and avoid the accident site,
algorithms to analyze the video footage and detect the presence of thereby preventing further accidents or injuries. It's important to
an accident. This could involve detecting certain patterns or note that while the accuracy of the system is 95%, there is still a
movements that are indicative of an accident, such as sudden chance of false positives or false negatives. Therefore, it's crucial
JETIR2304585 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org f646
© 2023 JETIR April 2023, Volume 10, Issue 4 www.jetir.org (ISSN-2349-5162)

to continue testing and improving the system to ensure its

reliability. Overall, your project has the potential to make a The existing system has several disadvantages, including:
significant impact in reducing the number of road accidents and
related deaths, especially in areas where emergency response times Manual reporting systems are often slow and may not provide
may be longer. timely assistance to accident victims, leading to increased
fatalities.
II. RELATED WORK Ad-hoc systems for collecting traffic data may be limited in their
It's interesting to hear that you've researched and compared your effectiveness due to challenges with maintaining communication
work with other accident detection techniques. It's true that many and data transmission in different terrains and conditions.
studies in this field focus on enhancing tangible infrastructure, Accident detection systems that rely on expensive sensors can be
such as road design and signage, rather than using intelligent cost-prohibitive and may not be accessible to everyone.
transportation systems. While there are a few existing studies on The dependency on sensors can also limit the effectiveness of the
accident detection using ITS, as you mentioned, they may be system, as the sensors may not be able to accurately detect all
constrained by geographical or demographic aspects, making them accidents or collisions.
less applicable to other regions or populations. It's important to False positives may occur if the system detects a collision or
continue exploring and developing different approaches to accident accident that did not actually happen, potentially causing
detection, including those using machine learning and computer unnecessary panic or emergency response.
vision, as these techniques have the potential to be more flexible
and adaptable to different settings. Overall, by comparing your 3.2 PROPOSED SYSTEM
work to existing techniques, you can gain a better understanding of
the strengths and limitations of different approaches and identify Our proposed model combines CNN and LSTM layers to perform
opportunities for future research and development in this area.: continuous video classification from a camera feed. The CNN
component is based on Inception v3, but we made some
It's interesting to hear about the "Lexus Enform" feature and how it modifications to better suit our training images. We added LSTM
uses impact sensors to detect accidents and notify users via an layers to the existing Convolution Network to incorporate temporal
application. However, as you mentioned, there are some features alongside spatial features. The model consists of
disadvantages to this system, such as the cost of placing sensors in convolution and recurrent parts, with the CNN handling feature
each individual vehicle and the reliance on physical entities like extraction from the images and the LSTM responsible for sequence
smartphones. Using a machine learning and computer vision-based prediction. This architecture is commonly used in tasks such as
system, as you have developed, can potentially overcome some of Activity Recognition, Image Description, and Video Description.
these limitations. By using cameras and analyzing video frames,
you can detect accidents in real-time without the need for physical While the proposed system initially faced challenges in accurately
sensors in each vehicle. Furthermore, your system can predicting the correct class, there are still several advantages to this
automatically notify relevant authorities without the need for system:
manual intervention, making it a more efficient and reliable
approach. Overall, it's important to continue exploring and  The system is capable of detecting vehicle collisions and
developing different approaches to accident detection, taking into accidents, which can be helpful in alerting emergency services
account the strengths and limitations of existing systems, as well as and providing prompt assistance to those involved in the
the potential for new technologies to improve upon them. accident.
 The use of TensorFlow and CNN algorithms in the training
OnStar Corporation, an ancillary company of General Motors, process allows for more accurate predictions over time as the
launched Chevy Star, an accident notification app that provided system continues to learn and improve.
on-field assistance to victims and self-regulated crash response  The system does not rely on costly sensors or manual
options [4]. However, the app was only available through a reporting, making it more cost-effective and accessible for
subscription model, which made it costly. Additionally, reviews wider use.
indicated that the service was of low quality, resulting in an  The system can operate in real-time, allowing for quick
ineffective system. responses and faster aid to those in need.
 The incorporation of noise reduction techniques can help to
SOSmart SpA developed an innovative smartphone app called improve the accuracy of predictions, even in low-quality or
SOSmart [5], which provided prompt and free assistance to unclear images.
accident victims with just a simple click. However, the manual
reporting system had some drawbacks. Ad-hoc systems are
commonly utilized for collecting traffic data, but they are limited IV. PROPOSED METHODOLOGY
by the challenge of maintaining communication and data The proposed system comprises three modules:
transmission in varying terrains and conditions [6][7]. Most
accident detection systems rely on costly sensors either mounted Load and Generate CNN Model: This module involves loading and
on vehicles or embedded in smartphones, rendering them less training a CNN model with a dataset to recognize and identify
effective and more expensive compared to the proposed accident accidents in videos.
detection system. Browse System Video: Users can upload videos to the system for
analysis in this module.
Start Accident Detector: The uploaded video is loaded, played, and
analyzed for accident detection.
The algorithms used in the system are CNN and LSTM. CNN is
III. SYSTEM ANALYSIS
used for image recognition and pixel data processing, making it
3.1 EXISTING SYSTEM suitable for identifying and recognizing objects in images or
videos. LSTM, on the other hand, is used to interpret the features
When comparing our accident detection techniques with others in extracted by the CNN model across time steps, making it useful for
the field, we found that most studies focus on improving physical analyzing video data.
infrastructure rather than Intelligent Transportation Systems (ITS),
which encompass various functions such as detecting traffic In the proposed system, the CNN model is used for feature
congestion, accidents, and events. However, the limited number of extraction, and the LSTM model is used for interpreting these
studies that do exist lack implementation details and are often features across time steps. This approach helps to improve the
specific to particular terrains and demographics. Thus, accuracy of the system in detecting accidents in videos.
geographical and demographic constraints may limit their
effectiveness.

JETIR2304585 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org f647
© 2023 JETIR April 2023, Volume 10, Issue 4 www.jetir.org (ISSN-2349-5162)

In above screen tensorflow model is loaded and now click on

‘Browse System Video’ button to upload video

In above screen selecting and uploading video and then click on

‘Open’ button to load video

Figure1: Architecture diagram

V. MODEL COMPARISON
This project is trained with images where vehicles collided and
accident occurred and in test video if anything such collision
happens between vehicles then application detect as accident.
Training is done with tensorflow and CNN Algorithm.
In above screen video is loaded and now click on ‘Start Accident
To run project double click on run.bat file to get below screen
Detector’ button to play video and detect accident

In above screen click on ‘Load & Generate CNN Model’ button to

trained CNN with dataset and to load CNN model using tensorflow In above screen video start playing and upon accident detection
will get below screen with beep sound

JETIR2304585 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org f648
© 2023 JETIR April 2023, Volume 10, Issue 4 www.jetir.org (ISSN-2349-5162)

detects a collision in a test video, it identifies it as an accident. To

run the project, double-click on the run.bat file to launch the
application. The main screen appears, displaying three buttons for
the different modules of the system: Load & Generate CNN
Model, Browse System Video, and Start Accident Detector.
Clicking on the Load & Generate CNN Model button trains the
CNN model with the dataset and loads it using TensorFlow. The
Browse System Video button allows users to upload a video for
analysis. Once the video is uploaded, click on the Start Accident
In below screen playing another video without message if normal Detector button to play the video and detect any accidents. If the
driving appear system detects an accident, a message will appear on the screen
along with a beep sound. Otherwise, if the video shows normal
driving, no message will be displayed. The system is trained with a
set of images showing accidents, enabling it to recognize accidents
and collisions accurately.

VI. CONCLUSION
The project has been trained using images that depict vehicle
collisions and accidents. During testing, if the application detects
any collision between vehicles in a video, it is identified as an
accident. TensorFlow and CNN (Convolutional Neural Network)
algorithms were used for the training process.

VII. REFERENCES
[1] "Global status report on road safety 2015", World Health
Organization, 2019. [Online]. Available:
http://www.who.int/violence_injury_prevention/road_safety
_status/20 15/en/. [Accessed: 07- Mar- 2019].
[2] Prabakar, S., et al. "An enhanced accident detection and
victim status indicating system: Prototype." India
Conference (INDICON), 2012 Annual IEEE. IEEE, 2012.
[3] "Lexus Enform", Lexus, 2019. [Online]. Available:
https://www.lexus.com/enform. [Accessed: 07- Mar- 2019].
[4] "OnStar Safety and Security Services", Onstar.com,
In above screen upon collision then accident display message will 2019. [Online]. Available:
https://www.onstar.com/us/en/services/safety-security/.
appear with beep sound [Accessed: 07- Mar- 2019].
In below screen application is trained with below images [5] "SOSmart automatic car crash detection and notification
app", SOSmart automatic car crash detection app, 2019.
[Online]. Available: http://www.sosmartapp.com.
[Accessed: 07- Mar- 2019].
[6] C. Kockan, “Communication between vehicles” PhD
thesis, Istanbul Technical University, 2008
[7] Zeng, Yuanyuan, Deshi Li, and Athanasios V.
Vasilakos. "Opportunistic fleets for road event detection in
vehicular sensor networks." Wireless Networks 22.2 (2016):
503-521.
[8] Szegedy, Christian, et al. "Rethinking the inception
architecture for computer vision." Proceedings of the IEEE
conference on computer vision and pattern recognition.
2016.
The project uses TensorFlow and CNN algorithm to detect [9] Szegedy, Christian, et al. "Going deeper with
accidents in videos. It is trained on a dataset of images showing convolutions." Proceedings of the IEEE conference on
vehicles colliding and accidents occurring. When the system
computer vision and pattern recognition. 2015.

JETIR2304585 Journal of Emerging Technologies and Innovative Research (JETIR) www.jetir.org f649