TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.

org

AI BASED ACCIDENT DETECTION AND

ALERT SYSTEM
Dr. LATHA MAHESWARI T, M.E.,Ph.D.,
SUPERVISOR, ASSOCIATE PROFESSOR
(SKCET)

SELVANAYAKI G, SUJI PRIYA R, VINOTHINI M

(CSE & SKCET)

providing medical assistance in real time, saving

ABSTRACT many lives. In most of the city, traditional traffic
Road accidents are a common cause of tragic control systems already have IP cameras and
deaths, and frequently the victim dies as a result sensors installed to monitor and control traffic.
of failing to report them to the appropriate Traffic tickets can be generated automatically by
authorities. The accident was not reported, so these systems. In this paper, propose a more
emergency care was not provided, resulting in advanced traffic monitoring system that uses live
death. A smart city with a traffic monitoring and camera fees to identify and detect moving objects
reporting system based on AI can assist in like cars, bikes, and other vehicles.

1. INTRODUCTION any information that has been superimposed on it;

real-time detection and recognition of current
1.1 VEHICLE DETECTION
automobiles can effectively prevent fatal traffic
The individual drivers of overweight vehicles are accidents like rear-end collisions.
alerted by the vehicle detection system as they
1.2 DEEP LEARNING
approach overhead structures like bridges, tunnels,
and other structures. Speed, vehicle counts, vehicle Deep learning is a subset of machine learning and
classification, stopped vehicle detection, wrong- artificial intelligence (AI) that mimics human
way vehicle detection, and accident detection are learning. Statistics and predictive modeling are two
just a few of the many types of detection that these branches of data science that include deep learning.
kinds of systems can perform, as are more common It is very helpful for data scientists who are
traffic data like occupancy, gap between vehicles, responsible for collecting, analyzing, and
queue detection, and so on. Any subsystems that Interpreting lot of data; Deep learning Algorithms
make up the system are: traffic flow control, gate are stacked in a hierarchy of increasing complexity
control at toll stations, video surveillance of traffic, and abstraction, in contrast to traditional machine
remote facility monitoring, and a central control learning algorithms that are linear. The process that
room. One of the most important research topics is a child goes through when learning to identify a
vehicle detection and recognition, which is also dog is very similar to that of deep learning
essential to the safe operation of vehicles. The computer programs. In order to generate a
process of rectifying a radio wave and recovering statistical model as an output, each algorithm in the
TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 924
 TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.org

hierarchy applies a nonlinear transformation to its LITERATURE REVIEW

input. The process continues until the output is
2.1 MEASURING THE OBJECTNESS
accurate enough to be considered acceptable. The
OF IMAGE WINDOWS
name "deep" came from the number of processing
layers that must pass through data. Bogdan Alexe and others In this paper, present a
generic objectless measure that quantifies the
1.3 CONVOLUTIONAL NEURAL
likelihood of any class of objects being present in
NETWORK
an image window. It is explicitly trained to
Convolutional neural networks are a specialized distinguish between amorphous background
type of artificial neural networks that use a elements like grass and a road and objects with
mathematical operation called convolution in place clearly defined boundaries in space like cows and
of general matrix multiplication in at least one of telephones. In a Bayesian framework, the measure
their layers. They are specifically designed to combines a number of image cues measuring
process pixel data and are used in image characteristics of objects, such as how different
recognition and processing. In deep learning, a they appear from their surroundings and whether
Convolutional neural network (CNN, or Convnet) their boundaries are closed. An innovative cue to
is a class of artificial neural network (ANN), most measure the closed boundary characteristic is one
commonly applied to analyze visual imagery. of these. In this project to demonstrate that the
CNNs are also known as Shift Invariant or Space combined objectless measure outperforms any
Invariant Artificial Neural Networks (SIANN), other cue in experiments on the challenging
based on the shared-weight architecture of the PASCAL VOC 07 dataset and that this new cue
convolution kernels or filters that slide along input outperforms a cuttingedge saliency measure. A
features and provide translation-equivariant HOG detector, interest point operators, and three
responses known as feature maps. recent attempts at automatic object segmentation
Counterintuitively, most Convolutional neural are also compared. In conclusion, to describe about
networks are not invariant to translation, due to the the present two examples of objectless in use. In
down sampling operation they apply to the input. the first, we present a method for using a small
They have applications in image and video number of windows as location priors for
recognition, recommender systems, image contemporary class-specific object detectors by
classification, image segmentation, medical image sampling them based on their objectless
analysis, natural language processing, brain– probability. This significantly reduces the number
computer interfaces, and financial time series. of windows evaluated by the costly class-specific
model, as demonstrated experimentally. Which use
objectless as a complementary score in addition to
the class-specific model in the second application,
which results in fewer false positives. Objectless
has been shown to be a useful mechanism for
focusing attention in a number of other image

TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 925

 TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.org

window-based applications, such as unsupervised 2.3 SELECTIVE OBJECT

pixel wise segmentation, weakly supervised RECOGNITION THROUGH
learning of object categories, and video object SEGMENTATION

tracking. Koen, E. A. van de Sande, and others the current

2.2 TECHNIQUES FOR VEHICLE state of the art for object recognition is based on

TRACKING AND DETECTION: A SIMPLE exhaustive search, as this paper proposes. A

REVIEW selective search strategy, on the other hand, is

required in order to make use of more expensive
Raad Ahmed Hadi and company This paper
features and classifiers and thus advance beyond
proposes that applications of vehicle detection and
the current state of the art. As a result, to reconsider
tracking play an important role in both civilian and
segmentation in order to adapt it to a selective
military applications, such as control,
search: Because an object whose location is never
management, and urban traffic planning for
generated cannot be recognized and appearance
highway traffic surveillance. The road-based
and immediate nearby context are most effective
vehicle detection process can be used for a variety
for object recognition, To propose that many
of purposes, including vehicle tracking, vehicle
approximate locations be generated rather than few
counts, the average speed of each vehicle, traffic
and precise object delineations. With just 1,536
analysis, and vehicle categorization. In this review,
locations per image, our class-independent method
to provide a brief overview of the image processing
covers 96.7 percent of all objects in the Pascal
and analysis tools utilized in the development of
VOC 2007 test set. On the Pascal VOC 2010
the aforementioned traffic surveillance system-
detection challenge, our selective search enables
related applications. To be more precise, and in
the use of the more expensive bagof-words
contrast to other reviews, are divided the
method, which we use to significantly outperform
processing methods into three groups to provide a
the standard by up to 8.5% for 8 out of 20 classes.
better explanation of the traffic systems. Traffic
surveillance is one of the most important uses for 2.4 CPMC: CONSTRAINED PARAMETRIC

video-based supervision systems. As a result, for a MIN-CUTS FOR

AUTOMATIC OBJECT
considerable amount of time, research has been SEGMENTATION
conducted in the Vision-Based Intelligent
Joao Carreira and others, has proposed in this paper
Transportation System (ITS), transportation
that, by utilizing mid-level selection cues and
planning, and traffic engineering applications in
bottom-up computational processes, we can
order to extract useful and precise traffic
generate and rank plausible hypotheses for the
information for traffic image analysis and traffic
spatial extent of objects in images. By solving a
flow control. This information includes vehicle
series of constrained parametric min-cut problems
count, vehicle trajectory, vehicle tracking, vehicle
(CPMC) on a regular image grid, the object
flow, vehicle classification, traffic density, vehicle
hypotheses, which are represented as figure-
velocity, traffic lane changes, license plate
ground segmentations, are automatically extracted
recognition, and other similar data.
TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 926
 TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.org

without prior knowledge of the properties of makes good use of multiscale data. Finally, To
individual object classes. Then diversify the propose a grouping strategy that efficiently
estimated overlap score using maximum marginal explores the combinatorial space of our multiscale
relevance measures to learn to rank the regions to combine them into highly accurate
corresponding segments by training a continuous object candidates. To demonstrate that MCG
model to predict how likely they are to exhibit real produces cutting-edge contours, hierarchical
world regularities (expressed as putative overlap regions, and object candidates through extensive
with ground truth). In the VOC 2009 and 2010 experiments on the
datasets, to demonstrate that this algorithm BSDS500 and PASCAL 2012 segmentation
significantly outperforms the current state of the art datasets.
for low-level segmentation. A segmentation-based
2.6 HARDWARE DESCRIPTION
visual object category recognition pipeline can
successfully employ the algorithm, as we A GSM Module is principally a GSM Modem( like

demonstrate in our companion papers. In the SIM 900) connected to a PCB with different types

VOC2009 and VOC2010 image segmentation and of affair taken from the board – say-so TTL Affair(

labeling challenges, this architecture came in first for Arduino, 8051 and other microcontrollers) and

place. For highlevel vision tasks like object RS232 Affair to affiliate directly with a PC(

recognition, it is important to be able to reliably particular computer). The board will also have legs

identify the spatial extent of objects in images. A or vittles to attach mic and speaker, to take out 5V

region that completely surrounds an object or other values of power and ground connections.

provides a characteristic spatial scale for feature These type of vittles vary with different modules.

extraction, protects the object from background Lots of kinds of GSM modem and GSM

signals that could be confusing, and makes it Modules are available in the request to choose

possible for information to spread from individual from. For our design of connecting a GSM modem

parts of the object to the whole. For instance, if a or module to Arduino and hence shoot and admit

human body is completely covered, it is possible to SMS using Arduino its always good to choose an

spread the person's identity from the face, which is Arduino compatible GSM Module – that's a GSM

easier to identify, to the rest of the body. module with TTL Affair vittles.

2.5 COMBINATORIAL GROUPING AT 3. EXISTING SYSTEM

MULTIPLE SCALES In the existing system to use algorithm RCNN and

Pablo Arbeláez and others In this paper, we present YOLO. The traditional traffic monitoring system is

Multiscale Combinatorial Grouping (MCG), a only intended to monitor or control traffic;

unified strategy for bottom-up hierarchical image however, it does not offer a solution for reducing

segmentation and object candidate generation for the rate of fatal accidents resulting from a lack of

recognition. First, to create a quick normalized cuts immediate medical assistance. Take for instance a

algorithm for this purpose. After that, it suggest a situation in which an accident took place but no one

hierarchical segmented with high performance that was present to report it; the victim is in critical

TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 927

 TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.org

condition, so every second counts, and any delay

Image with
could result in disability or death. Additionally, accident and
non accident Ye
Y
there are numerous sensor-based systems on the s
es
Alert
market that require vehicle owners to install those messa
ge
sensors. These systems operate on the basis of any Accident
image
damage detected by the installed sensors; a system dataset

that will alert nearby medical assistance or an Trai

ning
I
emergency contact number will be triggered by s
Road CNN a
surveill
these signals from the sensors. But what if an Algo
rith
c
c
accident occurred with a vehicle without such a i

sensor-based system? It’s need a cutting-edge

N
surveillance system based on artificial intelligence o
that can not only detect an accident but also
4.1 USER LOGIN MODULE
instantly notify nearby hospitals, ambulances, or
traffic police. Module for user login using this module, a user can
log in by entering the correct user name and
4. PROPOSED SYSTEM
password.
In India, accidents have been a major cause of
4.2 LOAD AND GENERATE CNN MODEL
death. More than 80% of deaths caused by
MODULE
accidents are not caused by the accident itself, but
rather by the lack of prompt assistance for the The CNN model module can be used to load the
victims. Accident victims may be unattended for a training data set and generate the CNN model.
considerable amount of time on highways with
4.3 START ACCIDENT DETECTION
extremely light and rapid traffic. A system that
MODULE
would use the live video feed from a highway-
mounted CCTV camera to identify an accident is Start the accident detection module by entering

the goal. The process involves running each video video The user can start the accident detection

frame through a deep learning convolution neural module by entering video.

network model that has been trained to classify 4.4 LOSS AND ACCURACY GRAPH
video frames as accidental or not. Classifying MODULE
images using Convolutional neural networks has
Loss and accuracy graph module this module's
proven to be fast and accurate. When compared to
overall loss and accuracy in finding accidents
other image classifying algorithms, CNN-based
image classifiers require less preprocessing and 4.5 EXIT MODULE
have an accuracy of over 95% for relatively smaller
Exit module the proposed system's home page can
datasets.
be used by the user to exit.

TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 928

 TIJER || ISSN 2349-9249 || © May 2023 Volume 10, Issue 5 || www.tijer.org

5. EXPERIMENTAL RESULTS 6. CONCLUSION

The use of Deep Learning's visual and temporal Today, road traffic accidents account for a
features in the construction of road accident significant portion of daily fatalities. The driver's
analysis and automatic detection models is crucial error and the emergency services' tardiness in
for the timely implementation of road accident responding are the primary causes. A reliable road
detection systems. The proposed model is applied accident detection and information transfer system
to each frame of the video that is being captured by is required to save injured individuals. The
the Pi-camera when the system begins to run, and proposed approach is based on video analytics
if an error is detected, the GSM module is methods. In particular, architectures of deep
immediately used to send a message. Additionally, learning neural networks that have been trained to
it transmits the frame at which it detected an recognize signs of a car accident are used.
accident along with the accident percentage. approach is based on video analytics methods. In

Fig 1. TRAINING AND VALIDATION ACCURACY particular, architectures of deep learning neural
networks that have been trained to recognize signs
of a car accident are used.

7. REFERENCE

[1] "Measuring the objectness of image

windows," by B. Alexe, T. Deselaers, and V.
Ferrari, TPAMI, 2012.
Fig 2. TRAINING AND VALIDATION LOSS
[2] J. R. R. Uijlings, K. E. A. van de Sande, T.
Gevers, and A. W. M. Smeulders, "Selective
Search for Object Recognition," International
Journal of Computer Vision,
Cilt.

[3] Guzel, MS, "Versatile Vehicle Tracking

and Counting Application," KaraElmas Science
Fig.3 IR PROXIMITY SENSOR CONNECTED WITH and Eng Journal, 7(2), 622- 626, 2017.
ARDUINO BOARD
104, s. 154–171,2013.

[4] "CPMC:" by J. Carreira and C.

Sminchisescu, Automatic object
segmentation using constrained parametric min-
cuts,” 2012, IEEE Transactions on Pattern
Analysis and Machine Intelligence, Cilt.34, pp.
1312–1328.
[5] "Multiscale combinatorial grouping" by P.
Arbelaez, J. Pont-Tuset, J. Barron, F.
Marques, and J. Malik, CVPR, 2014.

TIJER2305225 TIJER - INTERNATIONAL RESEARCH JOURNAL www.tijer.org 929