Accident Emergency Alert System using Deep

Learning
Nidhi Bhatia Yash Dixit Dr. K.M Balamurugan
School of Computer Science and Engi- School of Computer Science and Engi- School of Computer Science and Engi-
neering neering
2023 Third International Conference on Artificial Intelligence and Smart Energy (ICAIS) | 978-1-6654-6216-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAIS56108.2023.10073901

neering
Galgotias University Galgotias University Galgotias University
Greater Noida, Uttar Pradesh Greater Noida, Uttar Pradesh Greater Noida, Uttar Pradesh
[email protected] [email protected] k.baalamurugan@galgotiasuniversi-
ty.edu.in

Abstract—An accident is an unexpected and unwanted event. time by nearly 60 seconds which results in a six percent
One of the primary causes of car accidents is speed. Many lives increase in the number of lives saved.
could be saved if rescue services received accident information
and responds to it quickly. Human lives became simpler with the
Reducing the time gap between the occurrence of an
emergence of technology and infrastructure. Given the
concerning growth in the number of accidents occurring in
accident and when first responders, such as medical staff,
India, this method ensures that authorities are notified prior to summons themselves and take the required actions can
an accident or at the moment of the event. The advancement of reduce the severity of an accident by significant rates . To
technology has also resulted in an increase in traffic dangers and detect the existence of an accident, the Accident Emergency
road accidents which result in huge loss of life and property due Alert System employs object tracking in a camera [8].
to inadequate emergency services. The failure to receive timely
care is the leading cause of mortality in traffic accidents account-
ing for half of all fatalities. Since, after an accident every second Over-speeding In India
counts and immediate action is required to avoid the death of a
person, intelligent transportation systems have lately evolved as a
powerful means of improving the interpretation of transportation
systems and promoting travel safety. One of the most effective
technologies is accident detection systems so that the rate of
death in road accidents can be decreased by treating the victims
as soon as possible. This article is based on computer vision
which is an emergency alert system built on python which uses
object tracking techniques to detect the occurrence of an
accident. An efficient automatic accident detection system
featuring automated reporting of the accident site to emergency
responders is critical for saving valuable human life. This system
intends to identify accidents beforehand and communicate the
information to emergency responders in order to give rapid aid to
the wounded individual. The research work purpose is to
determine the severity of an accident and to warn the rescue crew
in time.
Keywords—accident, alert system, emergency, accident alert Figure 1 Over speeding in India : [1]
based on computer vision.

The likelihood of an accident is calculated based on

vehicle’s speed [5]-[6] and trajectory abnormalities
I. INTRODUCTION before or after colliding with other vehicles. On common
road-traffic CCTV [8] surveillance video, the suggested
The fast growth of economic construction continues to framework
increase people's living standards. Road traffic accidents
occur often, resulting in significant losses of life and proper-
ty for the country and its citizens. In the national interest,
traffic has become a major event. Poor emergency response
is a key contributor to our country's high number of road
fatalities and death rate.

Several technological and sociological advancements have

contributed to the reduction of traffic fatalities over the last
decade; for example,1% increase in seatbelt usage will
approximately save nearly 136 lives [10]. Furthermore, in
every moment a wounded victim of accident is unable to
receive emergency medical treatment in time makes a
significant difference in their survival rate i.e., according to
research, cutting accident reaction
Figure 2 Reasons for death in the year 2015-2018 [13]

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.
provides a dependent strategy for achieving a high rate of III. M ETHODOLOGY
Detection and low rate of False Alarm. Using the suggested
dataset, this framework was tested under a variety of A. PROBLEM FORMULATION
situations including broad daylight, low visibility weather,
rainfall, and snow [9]. This architecture has been proven to This system employs Computer Vision, Artificial
Intelligence, Machine Learning and Deep Learning to
be effective, paving the way for the creation of real-time
predict and detect high-severity traffic accidents [2] on a
general-purpose vehicle accident detection and emergency
range of high- ways and in a variety of traffic
alert systems [10]. circumstances.
Vehicle accident detection has been a popular field for Despite the fact that, we now have a solid emergency
implementing the technology related to computer vision to service system in place to respond to road accidents, we are
solve the difficult issue of delivering timely health services still unable to provide manual notifications if the incidents
without the requirement for a manual human operation to occur after midnight. To identify road accidents
monitor an accident and respond to it timely in recent years. autonomously in a split second, this challenge may be
As a result, this study presents a practical solution to the tackled utilizing Machine Learning and Object tracking
systems. Detecting accidents that would otherwise go
above-mentioned problem by proposing a technique to unreported would result in safer roadways and a more
identify accidents happening on the spot, which is critical efficient civil defense, Accident and Emergency
for paramedics and traffic agencies to quickly send a departments, free of human error.
response to the situation. This research presents a method
that to detect a significant number of well-known road-side The proposed system is intended to be as simple and
items utilizing the supervised deep learning architecture cost-effective to implement as feasible, especially because it
and well-developed training sets. will simply operate on top of current, widespread road
CCTV equipment. As a result, the expenses of this solution
are kept low because it does not need any major paradigm
adjustments before it can be implemented. A rapid
notification system that can automatically detect probable
II. LITERATURE REVIEW accidents and notify relevant Accident and Emergency
For the past two decades, researchers in the fields of deep departments through a mobile or online alert is one of the
learning, computer vision, image processing and machine primary aspects of this solution. We think that in even the
learning have been researching the topic to detect accidents most severe and critical situations, this instantaneous
[12]. As a result, a variety of solutions have been presented feedback mechanism can mean the difference between life
and created to address this issue. and death.

Singh et al. proposed a method of collecting Spatio- B. DESIGN AND ARCHITECTURE

Temporal Video Volumes from a CCTV security camera
feed and placing a strong emphasis on denoising auto -
encoders afterwards to produce a peculiarity score while
concurrently detecting tracking the objects, moving objects,
and finally finding the intersection of their ways to process
the chances of an occurring accident.[7] This approach
might be beneficial in determining car accidents at
crossings with a steady flow of traffic and enough lighting.
However, it has significant flaws in processing low-visibility
accidents, significant blockage in automotive accidents, and
large changes in patterns observed on roads . It also
underperforms since it is completely reliant on trajectory
intersections and traffic flow anomalies. It implies that it
will face difficulties in irregular traffic patterns and
trajectories.

In the same manner, Hui et al. proposed a strategy that

employs the Gaussian Mixture Model (GMM) to identify
vehicles, followed by the mean shift algorithm to track
them. Even while this algorithm does a good job of
addressing occlusions during accidents, it has a fundamental
flaw in that it relies on restricted parameters in
circumstances where traffic patterns change often and severe
weather conditions exist.
Figure 3 video detection system
Although these techniques keep a precise track of vehicle
movements, they underperform when it comes to
parametrizing accident detection criteria. Because they need When a video detection system installed in a CCTV camera
a certain form of data and hence cannot be executed in a detects an accident, an emergency alert is issued, and a mes -
general context, they are inefficient in developing criteria sage with the percentage of collision severity is sent to the
for accident detection. Through parameter customization, relevant accident and emergency departments' computers
established systems are improved for a single CCTV through servers. The proposed framework can detect
camera. The suggested framework's unique feature is its accidents with some Detection Rate and False Alarm Rate
ability to operate with any CCTV camera feed. Aside from on accident footages captured in various contextual
that, deep learning techniques to visual surveillance for circumstances.
traffic congestion and road surface monitoring are also
known.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.
 C. TOOLS AND TECHNOLOGIES USED
Software and Libraries used Mask R- CNN (Region-based Convolutional Neural Net-
The following libraries were used in this project to connect works) is the object identification framework which is
distinct modules: employed here. We split and produce pixelwise masks for
each element in the camera footage using Mask R-CNN.
Python 3 The earlier RoI Pooling yielded 10 to 50% more precise
Python is a scripting language that is high-level, interpreted, mask findings [11]. Mask R-CNN, on the other hand,
interactive, and object-oriented. Python is intended to be a outperforms Faster R-CNN [15] by applying a modified RoI
very understandable language. Python, unlike the other Align technique. RoI Align can achieve precision because it
languages, does not employ punctuation and instead relies eliminates the issue of location misalignment that hindered
on English keywords. Python 3.10.2 is a new version of the RoI Pooling, which aimed at fitting the input feature map
language. into the blocks. As a result, Mask R-CNN gives us the
benefit of the Instance Segmentation notion. It also
OpenCV enhances accuracy because it employs the RoI Align
Algorithm.
OpenCV (Open-Source Computer Vision Library) is an open
– source computer vision software library for machine A dictionary including boundary boxes, detection scores, all
learning. By offering a common framework for emerging the class IDs, and the created masks for the given video
technologies and computer applications, OpenCV aims to frame is the result from this step.
enhance the usage of machine perception in the business
sec- tor. This library allows code to be easily updated and D.2. Vehicle Tracking and Extraction of Features
used for commercial purposes.
We screen out all the detected objects right after the object
detection phase and just leaving the correctly vehicles which
TensorFlow are detected according to their class IDs and scores.
TensorFlow is an open-source software library for high- Following the detection of automobiles in a particular
performance numerical computing. The TensorFlow Object frame, the framework's next important task is to keep track
Detection API is developed on top of the TensorFlow open- of each of the observed items in the video's progressive time
source framework, making it simple to build, train, and frames. This is performed by using Centroid Tracking
deploy object detection models on projects. Algorithm [4], a basic yet very efficient object tracking
technique. This approach works by evaluating the Euclidean
distance between captured vehicle centroids over a series
of frames.
D. MODULES OF THE SYSTEM
This framework's centroid tracking technique is a various
The suggested system is made up of three interdependent step method that satisfies the requirements.The steps are as
modules. They are as follows: follows:
(A) Detection of Vehicles • The centroid of the items is obtained by intersecting the
(B) Vehicle Tracking and Extraction of Features
lines travelling across the mid points of the identified
(C) Detection of Accidents & Emergency Alert vehicles' bounding boxes.
D.1. Detection of Vehicles • Calculating the Euclidean distance between the centroids
of newly discovered items and those of previously
In this Module, the framework of the system detects vehicle discovered objects.
in the video:
• Use the shortest Euclidean distance in between latest set
of centroids and the prerecorded centroid to adjust the
coordinates of existing objects.
• Assign a unique ID to new objects in the area ofview
and save their centroid coordinates in a dictionary.
• De-register any element that have not been observed in
the ongoing field of view for a preset number of frames
sequentially.
Although item will move between consecutive frames of the
film, the gap between the center of the same entity across 2
subsequent images will be less than the gap be- tween the
center of any other object, according to the centroid tracking
method. There are a few factors that may be utilized to
forecast future collisions after the centroid has been
assigned to each vehicle separately.
1: The overlap of vehicle bounding boxes

2: Estimating Routes and Collision Angles

3: Estimating Speed and their variation in acceleration.
In this phase, the overlapping of the two vehicles plays a
Figure 4 Block diagram of the proposed method [3] major part because there is a good likelihood that the boxes
creating the border will overlap before any two cars collide.
Although there may be instances where the bounding boxes

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.
overlap, this does not always result in an accident. As a Then, using the equation below, we standardize the Scaled
result, false alarms might be created when vehicles are Speed (Ss) of the automobile is calculated by taking (H)
parked very near to each other at a traffic signal. Hence, this which is the height of the video frame and the length of the
framework considers other criteria apart from distributing car's bounding rectangle (h) and multiplying them together.
normal work to individual criteria. Now discussing the Scaled Speeds of tracked vehicles are stored in a dictionary
process that would be used in deciding the overlapping of for each frame.
two vehicles:
Ss = ( + 1)*Sg …………..(6)
Consider the bounding rectangles of two independent
transports A and B. Let x and y be the centroid coordinates
of a transports, and let α and β be the width and height of
the transport's bounding reactangle, accordingly. The vehicle's Acceleration for a certain Interval is then cal-
Overlapping condition of A and B is:
(2×|a.x−b.x| < a.α+b.α)∧(2×|a.y+b.y| < a.β+b.β)............... (1)

culated from the change in Scaled Speed from S1 to S2

The requirements determine the closeness of the centers s s
of A and B's bounding boxes which is done for both the
axes, regardless of whether they cross. The intersection
of the boundary boxes is considered only when both the A= …………(7)
horizontal and vertical axes intersect, which serves as a
foundation for subsequent criteria. The next goal of phase
(B) is to detect the vehicle's trajectories, which can be Next section discusses the usage of variations in
accomplished by calculating the difference between the
tracked vehicle's centroids every 5 frames, which is acceleration to identify vehicle collision. Workflow of the
accomplished by collecting the middle of each vehicle in
whole procedure of emergency alert systemis depicted as
each frame until the vehicle's centroid is recorded using
the previously described centroid tracking technique. This follows:
provides us the 2D vector, which indicates the motion's
direction. The magnitude of the vector may be calculated
as follows:

magnitude = (( μ . i )2 + ( μ . j )2) ...............(2)

To normalize this vector, we must first divide its magnitude

by its scalar division. We will retain the original magnitude
of the tracked vehicle in the dictionary of normalized
vectors if it exceeds the set threshold value, else we will
discard it. This is done to limit the number of false alarms
generated.

The angle between the tracks is then determined using the

traditional formula for determining the angle between two
vectors moving in the same direction. Throughout this case,
the overlapping vector orientations will be 1 and 2. The
angle of collision between the two trajectories is then
determined using the formula.
( μ1.μ 2) ...................................
θ = cos (3)
| μ1 | | μ 2 |
So, the next stage is to compute the vehicle's speed, which
may be done in sequence of steps. To do so, we must
calculate the difference between video frames, which is
inversely proportional to the Frames per Second.

τ = 1/FPS ................................................ (4)

The distance travelled by the vehicle may then be calculated

using its centroid from the first and last frames, represented
as c1 and c2. Using the centroid difference over a five-
frame interval, we can calculate the Gross Speed (Sg).
c2 − c1
sg = …………..…………..….(5) Figure 5 Proposed Flowchart
τ*Interval

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.
D.3. Detection of Accidents & Emergency Alert frame rate of 30 frames per second (FPS) are put into
consideration. This model's data samples are all CCTV
As the name implies, we will explore the process of recordings taken at traffic crossings across the world.
detecting accidents and generating emergency alerts in this Accidents under intense sunlight, daytime hours, snow, and
module. When the vehicles overlap, or condition C1 is night hours areall included in the dataset.
satisfied, as illustrated in the process diagram above, an
We estimate the collision between two cars and draw a
accident is identified. We shall present three metrics (α, β,
circle in the frame to represent the collision zone of interest.
γ) in order to detect abnormalities caused by accidents.
Existing clip accident detection methods use a tiny number
Here, α —> Acceleration Anomaly, of surveillance cameras in contrast to the dataset in this
β —> Trajectory Anomaly,
γ —> Change of Angle Anomaly.
When two cars collide, we compute the
vehicle's acceleration based on their documented velocity in
the dictionary. We compute the average acceleration of the
automobiles 15 frames before the overlapping condition work.
(C1) and its maximum acceleration 15 frames afterwards
C1. The difference between the highest and average
The suggested framework has a detection rate of 71 percent
accelerations under overlapping conditions is used to and a false alarm rate of 0.53 percent. The suggested
calculate the difference in accelerations of individual approach's effectiveness is due to its analysis of the various
independent vehicles (C1). The Acceleration Anomaly is
conditions that might lead to a collision.
used to identify collisions by comparing this difference to a
series of pre-defined parameters. This statistic records the
significant change in velocity that happens following
intersection, allowing accidents to be identified based on its
fluctuations.
The Trajectory Anomaly is defined as the angle of
intersection of the vehicle trajectories (θ) when the
overlapping condition C1 is fulfilled.
1) If θ ∈ (θL θH) , β is determined by a series of pre-defined
criteria on the value of
2) If not, β is calculated using θ and the distance between
the place of intersection of the trajectories based on a series
of pre-defined constraints.
Then, we provide a new parameter that accounts for
irregularities in a vehicle's orientation during an accident.
This parameter is calculated with the help of measuring the
angle (θ) of a vehicle regarding its own trajectory during a
five-frame interval. Because a vehicle rotates to some
degree about an axis in an accident, the trajectories become
the tangential vector in relation to the axis. We may
estimate this degree of rotation, and hence the amount of Figure 6 Deaths and Injuries in accidents for the year 2014-2018
orientation shift, by measuring the change in angles of the
vehicle's trajectories. We estimate the Change in Angle
Anomaly based on this angle for every one of the
vehicles in concern using a pre-defined set of values.
Finally, we apply a method to combine all the individually
discovered abnormalities in order to determine if an
accident occurred. This function f (α, β,γ) generates a score
between 0 and 1 by calculating the weighting factors of each
of the different limits based on their respective values. A
score greater than 0.5 implies a traffic collision; otherwise,
it is eliminated. The most crucial concept in accident
detection is this.

IV. ANALYSIS OF THE SYSTEM

We completed the analysis, and as previously said, the soft-
ware is developed in Python, with Keras and TensorFlow
libraries being utilized as well. OpenCV was used to process
the footage.
This research is based on YouTube footage of automobile Figure 7 Deaths and Injuries in accidents predicted for future
collisions from various regions. Surveillance films with a India

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.
The above figure number 6 shows the graphical Speech, and Signal Processing (ICASSP), vol. 6, M arch
representation of data of accidents from year 2014 to 2018. 1999, pp. 3305–3308.
The red line rep- resents the number of fatalities due to
accidents. The peak reaches up to 3.0 Lakhs fatalities in [5] R. J. Blissett, C. Stennett, and R. M . Day, “Digital cctv pro-
cessing in traffic management,” in Proc. of IEE Colloquium
year 2016. The loss of lives due to accidents has shown a on Electronics in Handling Road Capacity Demand, Nov
sharp inclination fromthe year 2014 to 2018 due to higher 1993, pp. 12/1-12/5.
involvement of peoplein transportation through vehicles.
[6] F. Baselice, G. Ferraioli, G. M atuozzo, V. Pascazio, and G.
With increase in the number of vehicles the chances of roads Schirinzi, “3d automotive imaging radar for transportation
accidents will also increase which might result in increase in systems monitoring,” in Proc. of IEEE Workshop on Envi-
fatalities. ronmental, Energy, and Structural Monitoring Systems, Sep
The figure 7 on the other hand shows the proposed graphical 2014, pp. 1–5.
representation of the data of accidents in the upcoming years [7] https://www.researchgate.net/publication/30480806 5_Evalua-
with the accident alert and emergency system, which clearly tion_of_Roadside_Wrong- Way_Warning_Systems_with_D-
shows the declined rate of fatalities. ifferent_Types_of_Se nsors
A quick response to the accidents and treatment of the [8]. A. Franklin, “The future of cctv in road monitoring,” in Proc.
injured ones in time will avoid the loss of lives up to some of IEE Seminar on CCTV and Road Surveillance, M ay 1999,
extent. With increase in time the annual fatality rate might pp. 10/1–10/4.
even come near to 0.
[9] Z. Hui, X. Yaohua, M . Lu, and F. Jiansheng, “Vision-based
Hence, the suggested framework could help in saving the real-time traffic accident detection,” in Proc. of World Con-
precious lives of the people. gress on Intelligent Control and Automation, June 2014, pp.
1035–1038.
[10] https://www.geographyandyou.com/disaster/disaste r-events/
V. CO NCLUSIO N track-troubles-road-accidents-in-india/
[11] https://www.researchgate.net/publication/32241275 0_Auto-
A basic framework has been established in this study matic_road_accident_detection_techniques_ A_brief_survey
for sensing vehicular crashes and transmitting an
emergency notice to the nearest Accident and [12] W. Hu, X. Xiao, D. Xie, T. Tan, and S. M aybank, “Traffic
Emergency Department. Local characteristics such as accident pre- diction using 3-d model-based vehicle
trajectory collision, velocity estimation, and their tracking,” in IEEE Transactions on Vehicular Technology,
vol. 53, no. 6, pp. 677–694, M ay 2004.
different anomalies are used to build this framework.
The major goal of this project is to reduce the numbers [13] https://timesofindia.indiatimes.com/india/last-years-road-ac-
killed in car accidents caused by excessivespeeding, to cident s -w ere-mos t -fat al-in-5-y ears /art icles how /
enhance public safety, and create a better system for 87574502.cms
regulating traffic on the roadways. The technology is [14] D. Singh and C. K. M ohan, “Deep spatio-temporal representa-
cost-effective, scalable, and rapid, and it can be simply tion for detection of road accidents using stacked autoen-
integrated into existing live surveillance systems. Our coder,” in IEEE Transactions on Intelligent Transportation
system's dependability is enhanced using numerous Systems, vol. 20, no. 3, pp. 879–887, M arch 2019.
factors to assess the risk of an accident. One of the [15] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards
study's shortcomings is its function in high-traffic real-time object detection with region proposal networks,” in
areas and in low-light settings, which causes problems IEEE Transactions on Pattern Analysis and Machine Intelli-
with vehicle identification and tracking inaccuracies, gence, vol. 39, no. 6, pp. 1137– 1149, June 2017.
which will be solved in future studies. Large
impediments in the way of the cameras' range of vision
may also hinder vehicle tracking and, as a result,
collision detection. The suggested framework can
accurately detect accidents with a 71 percent rate of
detection and a 0.53 percent rate of False Alarm. The
experimental findings are encouraging and illustrate
the framework’s effectiveness.

REFERENCES

[1] “Road traffic injuries and deaths global problem,” https://

www.cdc.gov/features/globalroadsafety/index.html.

[2] K. He, G. Gkioxari, P. Dollr, and R. Girshick, “M ask r-cnn,”

in Proc. of IEEE International Conference on Computer
Vision (ICCV), Oct 2017.

[3] “Object detection for dummies part 3: R-cnn family,” https://

lilianweng.github.io/lil-log/assets/images/rcnn-family- sum-
mary.png.

[4] J. C. Nascimento, A. J. Abrantes, and J. S. M arques, “An algo-

rithm for centroid-based tracking of moving objects,” in
Proc. of IEEE International Conference on Acoustics,

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on November 07,2024 at 07:24:10 UTC from IEEE Xplore. Restrictions apply.