Accident Precaution System For Vehicle In Motion Using Machine Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 04 | Apr 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 787
Accident Precaution System For Vehicle In Motion Using Machine
Learning
1,2,3,4Student at Mahatma Gandhi Mission College of Engineering and Technology, Mumbai, Maharashtra, India.
5Assistant professor at Mahatma Gandhi MissionCollegeofEngineeringandTechnology,Mumbai,Maharashtra,India.
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Nowadays speed and breaching trafficrules
causes many accident. we can save many liveifweprovide
accident information toemergencyserviceandifwereach
in time. heavy road traffic and increasing number of road
accidents are major concern in current scenario rather
than new vehicle have latest technology. Our survey on
this topic is made to construct such a system which is
efficient and reliable to detect danger while our vehicle in
motion. In this paper, we try to overcome the problem by
create system "Accident Precaution System For Vehicle In
Motion Using Machine Learning" using Deeplearningand
machine learning algorithmsuchasConvolutionalNeural
Network (CNN), Artificial Neural Network (ANN), YOLO
(you look only once).Using these algorithm we develop
different model such as Driver Drowsiness, object
detection, pothole detection and traffic sign detection for
decrease the possibility of accident.
Key Words: Deep learning, Machine learning,
Convolutional Neural Network (CNN), Artificial Neural
Network (ANN), YOLO(you look only once),python,
Computer Vision.
1. INTRODUCTION
Automobile has provide a great benefits in our daily life. We
use vehicle to reach destination on time. In 21st century it
hard to imagine life without vehicle. There are various types
of vehicle such as car, bus, truck etc. each used for different
purpose, But every coin had two side that way increasing
number of vehicle on road provide us benefits of fast
transportation and decrease our travel time but it also cause
disaster to us and may kill us through serious accident.
Inappropriate driving and over speeding causes risk to
involve in accident. Many efforts taken by various
organization and government to decreases number of
accident but still so many accident happen daily.Wecansave
many lifeby provideemergencyinformationaboutupcoming
danger while driving. According Data of Ministry of Road
Transport & Highways of India major reason for accident is
Over Speeding, Distractions of Driver, Traffic Light Jumping,
Non-adherence to lane driving and overtaking in a wrong
manner.
As most of accident occurs because traffic rules properly
not follow.so many drivers they actual not seen traffic light
because there vehicle in motion so deep learning algorithm
such as Convolutional Neural Network(CNN) which focus on
particular region can help in fast image processing[1].
Some causes for accident is Driver fall in slip while
driving, it happen mostlyinnightdriving.itmaycauseserious
crash so if we analyze driver face and detect facial feature in
real time it will overcome the danger of an accident[3].
Another obstacles on Road also cause accident such as
vehicle can crash with each other so vehicle can detect using
technology such as R-CNN, CNN, Darknet for detect and
analyze vehicle from surrounding in safe distance [9].
In this paper, weproposed systemsthatpreviouslydetect
Road objects, Traffic sign and Driver Drowsiness using
various Deep learning and machinelearningtechnologywith
tensorflow and image processing. It will help in Autonomous
vehicle creation and make driving safe and secure for Driver
and passenger.
2. RELATED WORK
In Existing system, models are built for accident prevention
such as Drowsiness Detection, road object detection and
traffic sign detection for create pre-alarm system that help
driver in his journey. These models are built separately and
used dataset that is limited. As we built different models we
see different work separately as follow:
Traffic sign detection
In paper[1] they built model on DFG dataset which had
different types of traffic signs, some of which give warning,
mandatoryand prohibitive instructionsonhowandwhereto
drive. To trained and built the model they used region-based
neural network which focus on particular region for detect
traffic sign. In paper to built traffic sign detection we had to
done pre-image processing as mention in paper [2].In this
they used Image enhancement technology for clear shapes
using linear filtering algorithm.
Driver Drowsiness
In previous study [3] To monitor and warn the driver in
real-time, the use of the kernelized correlation filters (KCF)
algorithm is preferred basedon system’s evaluation.Forreal
Saniya Shaikh1, Omkar Prabhu2, Rahul Patil3, Somesh Nikam4, Prof.Sachin Chavan5

time detection of driver face fatigue MC-KCF algorithm is
helpful for highest accuracy for Drowsiness detection by
tracking eyes fatigue. As mention in previous study in paper
Drowsiness track by face fatigue condition, as modelbuiltfor
detect eye aspect ratio by tracking eyes blinking rate.
Normally eyes blink two to three time a minute. For detect
Drowsiness we must track face for few second then as per
face condition detection result occurs.
In study in paper[3] we mark number of facial key points
as per mark rather than detect upper and lower eyelids we
detect eyebrows and whole eyessocketfortrackeyesfatigue.
According the study of paper[11] Behavioral measures
like yawning, amount of eye closure, eye blinking etc. needto
be consider, but normally activity like yawning is less so we
can say it’s not Drowsy. Normally amount of eyes closure is
also 0.1s to 0.4s so eyes are blink 3 to 4 time a second. But
when driver close to sleep amount of eyes closure is
increases.
Road object detection
As per study that previously done vehicle detection is
challenging for human to calculate, this is where a machine
learning had great benefits to automatically learn and
improve over time tracking, classifying, and counting the
number of vehicles passing by certain area [9].object
detection is beneficiary in prevention of accident therefore
fast detection of object is required for real-time object
detection we used technology Yolo and its different
version(v1,v2,v3,v5).
In the paper [9] we apply Deep learning based Yolo
framework and perform task such as Collection of picture of
interest, labeling and classify image, Train model and run
code, it will draw rectangular shapes on object and give label
to the object. Model specifies by this Provide accuracy 82%.
As per model are built separately and had low detection
Accuracy in some model and some model built had low
accuracy as they used dataset that had limited elements.
Pothole Detection
Major reason foraccident ispotholenotdetectedontime.
As per paper[4] for accurately detect the pothole models are
trainedandtestedwithpreprocesseddataset,includingYOLO
V3, SSD, HOG with SVM and Faster R-CNN. According to
accuracy of Different models Yolo v3 provide highest
accuracy.
In paper[7] study show that pothole detectionsystemcanbe
used for create report on road condition as system
automatically detect potholes so its used for monitoringand
make report whether road condition is good or bad. road
surface monitoring, deep learning, convolutional neural
network, k-nearest neighbor, global positioning system are
technology helpful for create the system. still shape of
pothole are different in different regionandthatcauseshard
to detect the poteholes.so there need to used dataset that
provide accurate different sets of images of potholes.
3. PROPOSED SYSTEM
Aim of this project is previously detect obstacles onroadthat
causes serious accidents and aware driver from incoming
danger on road as early as possible. Road obstacles that
causes most of the accident include Drivers Drowsiness,
Traffic light jumping, Various object on road such as other
vehicle is not detected,potholes.Acausesthatabovemention
is most common reason for road accident to happen. In our
study we found that some model are built previously for
solving this problem, but these model had many limitation
such as Low accuracy, trained on limited data, Detection
speed is slow. Previous system models are built separately
that only address particular problem. For creating perfect
accident precaution system we had to Built various models
that provide fast detection and high accuracy.so we built the
model such as Trafficsigndetection,DriverDrowsiness,Road
object detection and Potholedetectionfordetectallroadsign
and object, Driver Drowsiness come in drivers journey.so as
per our study we had to consider these models:
Traffic sign detection
Nowadays weheard about automated vehicletechnology
in this for avoid accidents we had to consider traffic rule. For
that traffic sign need to detect accurately. As most of traffic
sign are placed in left or right side and that cover small
portion of road its challenging to detect accurately.Detection
and classification is dependents on shape and color of the
traffic sign.so we used dataset that had various Indian traffic
sign including Traffic signal ahead, speed limit, no entry,
crosswalk, speed break, speed limit and many more. Dataset
mostly had image data with their labels.
To achieve high accuracy with real-time monitoring and
detection in this project we used CNN (convolution neural
network) and R-CNNare used. CNN are mostly usefulforreal
time detection as per paper[1].In this project Deep learning
image processing and object detection used CNN with
artificial neural network. convolutional neural network is
very helpful in deal with high definition image data and also
blur image to process image and make
prediction.[2]convolutional neural network is work
differently as they treats data as special. Instead of neurons
being connected to every neuron in previous layer, they are
instead only connected to neuron close to it and all have the
same weight. The simplification in connections means the
network upholds the spatial aspect of dataset.

Like normal neural network convolution neural network
made up of multiple layer that is convolutional layer, pooling
layer, ReLu layer, fully connected layer:
Fig-3.1: Convolutional neural network
Convolutional Layer: This is first layer of CNN
architecture. thislayer meanly deal with scanning imageand
get maximum input from image pixel.in this mathematical
operation is perform between input image pixel and filter of
size NxN, here n is size of filter. therearevariousfiltersuchas
Edge filter, colour filter, curve filter. when we slide or
convolve filter over image the random valueoffilterjoinwith
image pixel value and give us new set of value which help us
to identify image properties. After convolve it generated
featuremap thatgiveuseinformationaboutcornerandedges
of image. featuremap give asoutput to other layerforextract
more information about image and accurately detect the
image.
Pooling Layer: This layer is reduce the sample size of
feature map. this also make processing faster as it reduce
number of parameter network need to process.it reduce
samples size by decreasing connection between layer and
independly used each feature map. output of this layer is
pooled featuremap. there are two method of doing this fistis
max pooling which take maximum input of particular
convolve featureand averagepoolingwhichsimplycalculate
average of the elements in a predefined sized image section.
pooling layer act as bridge between convolutional layer and
fully connected layer.
ReLu: Relu is stand for rectified Linear Unit. Purpose of
this layer is introduce non linearity in neural network model
during the convolution operation.itsmostpopularactivation
layer using after convolutional operation. Relu is operation
with applied after each convolution operation it convert all
negative value to zero. whichhelp to achieve non-linearity.It
computes the function ƒ(κ)=max (0,κ). activation is simply
threshold at zero when value is negative it convert it into
zero.
Fully Connected Layer: FC layer is nothing but dense
network of neuron and connection between every two
neuron. we use fully connected layer to classify image to
particular categories after we have extracted feature from
image using convolutional layer and max pooling layer.
FC layer are dense network of neurons. this layer is
applied after convolutionalandmaxpoolinglayeritspurpose
is to classify the output. FC layer is representation between
input and output.it is final layer of architecture.
Fig-3.2:Traffic sign detection using CNN
Above fig-3.2 show working of different layer of
convolution neural networkas show in figure high definition
traffic sign speed limit 50 detect clearly and go through
various CNN layers.
Drivers Drowsiness Detection
As face is important aspect of our body it will used for
detection of DrowsinessofDriversBasedoneyesclosingtime
and fatigue detection.as number of vehicle increases per
day.[11]Big vehicles like bus, trucks,transportvehiclewhose
drive at night, so prolonged drive duration and bad working
condition Driver fatigue is major reason for accident. In this
project detection of Drowsiness sign is based on eyes aspect
ratio eyes aspect ratio is amount of duration of eyes blinking
2 to 3 timea second is normal. If eyes aspectratioisincreases
suppose 2 to 3 second eyes are not open then it will detect
the drowsiness.

Fig-3.3:Driver Drowsiness System
In fig.3 it’s a Driver Drowsiness system in thisDriverstart
driving camera captures its facial expression it will analyze
yawning, fatigue, eyes blinking and based on physical
reaction like postureand head pose it will detectDrowsyand
Buzzer Alarm.
Road Object Detection
while driving vehicle may crash with other vehicle or
other object on road such as cycle person create obstacle for
driver. In this project we create Road object detection model
that detect road object with high accuracy. we used opencv
and Deep neural network to detect object while driver is
driving. Opencv used for analyze all features of computer
vision and draw rectangle shape around object.
Fig-3.4: SSD network architecture: Base network
(VGGNet)
SSD is based on CNN it used to detect objectandVGGNetused
for image classification.
As pothole are major reason for accident so this is our
priority to detect pothole at run time.to achieve this goal we
proposed system thatusedcomputervision.Computervision
with deep learning model help us in fast detection of pothole
while vehicle in motion. major problem forpotholedetection
is pothole not in particular shapeforthatweusedCNN-CUDA
model that help us fast detection and harnessing power of
GPU and increases computing power to perform pothole
detection in real time and draw circle around pothole.it
required so much work ofgraphicalunitNVIDIAcudaToolkit
help us to accelerate GPU-accelerated applications which
provide good GPU utilization.
By combining all models we can create perfect accident
precaution system while vehicle in motion.
Fig-3.5:Architectural diagram for Accident
precaution system while vehicle in motion
As driver start vehicle its start capturing images basedon
pretrained modelsand analyze thefootageandbasedonthat
detect potholes, Traffic sign, Driver drowsiness and various
road objects.
4. METHDOLOGY
This project main concern with built system that detect
object and prevent accident while vehicle in motion. to solve
this problem methodological approach for study are as
following:
Dataset Collection and model building:
To build model for Road object detection The dataset is
collected from Kaggle Repository and was split into training
and testing data after its analysis.to build perfect road object
detection model which help in prevention of accident we
used dataset that include various trafficsignimageswithtwo
hundreds categories and more than seven thousand images,
car which had more than two thousand images, motorcycle
which had more than 3000 images, person's dataset had
Pothole Detection

more than 1000 images. combine dataset is train and test for
model building which had 13,874 images. we build model
using CNN and train the dataset using this architecture. To
make model predictedaccurate result wefirst performsome
operation on our datasettomakedataunderstandablebyour
trained model.
Road object detection model Training and testing:
model training and testing:
To train the model we used model.fit() function it work
well after successfully trained and test model. using 35batch
size we get 94% accuracy on trained model and get stability
using 15 epoch as per each epoch it get more stable.
epochs=15
M1 = model.fit(X_train1, y_train1, batch_size=32,
epochs=15, validation_data=(X_test2, y_test2))
we trained model on 13874 sample and validate on 3500
sample after testing the model we get following accuracy
chart-1:
Accuracy show how accuracy of training data and testing
data increases per epochs.
Loss graph show error will decreases as per epochs.
initially at training data error is high it goes down same in
testing data error decreases almost goes to zero.
Chart-1:Accuracy and loss chart of trained and testing
data
Performance of model:
As we test the model on real time video that detect
various road objectsuchasperson,motorcycle,carandtraffic
sign. After testing the model following is result of model
performance:
Table-1: Model performance table
OBJECT BEFORE
DETECTION
AFTER
DETECTION
ACCURACY_S
CORE
Person 0 6 0.90
Motorcycle 0 4 0.80
car 0 3 0.75
Traffic sign 0 1 0.83
Accuracy score in table 1 is increases as object is near to
camera and getting decreses after object is far away from
camera.
CNN is deep learning architecturewhich help us fastroad
object detection and recognition. CNN model get input from
camera that compare input data with various images in
dataset. It scan input and inputgoesthroughvariouslayersof
CNN and generate clear image of road object by applying
filter such as color filter, edge filter and curve filter. As most
traffic sign are
place on left or right side of road so input image is blur
but using CNN clear image of is provide to train model and it
predicted sign accurately.
To build this we used python and various libraries in
python such as opencv, keras, matplotlib, numpy which help
us to detection model with high detection accuracy. After
training and testing Accuracy of model is 94%.
Main work of Driver Drowsiness detection system model
is detect fatigue, Drowsinessof driver accuratelyandquickly
and alert driver who sit on driver sit. Model play important
role in implemented the system. For develop model for
Driver Drowsiness detection following are model and sub
model need to build:
i] Frame Acquisition: In thisprocessimageofdriverface
is capture via webcam camera it used forcapturinglivevideo
of driver eyes in all visible condition.
ii] Facial landmark detection: Faciallandmarkdetector
mainly used for face condition detection we use dlib face
detector and dlib python library for draw box around face.
iii] Eye Localization and tracking: This check eyes
condition whether eye is open or close.

iv] Measuring EAR: measuring eye aspect ratio(EAR) is
most important task in drowsiness detection. when eyes are
open then EAR is constant and when eyes blink then then
EAR immediately reduce to zero.
v] Monitoring of EAR for blinks detection: using
continuousvideocapturewemonitoringEARforDrowsiness,
fatigue detection by continuously detect using eyes using
webcam camera.
vi] Estimation of fatigue periods between blinking: It
is calculate eyes blinking time.
vii] Audio Visual warning on fatigue detection: When
model detect fatigue it sound buzzer to alert driver.
Drowsiness model training and testing:
We used SVM to test the data in SVM algorithm we create
model based on Eye Aspect Ratio. EAR calculate blinking of
eyes. result of drivers fatigue is calculated by SVM classifier.
Following are test case of Drowsiness model training.
Table-2: Testing of Drowsiness
Test cases EyesDetected Eyes closure result
Case 1 No No No result
Case 2 Yes No No result
Case 3 Yes Yes Voice alert
After training the model on SVM classifier we test it using
CNN model for fast detection. Performance of model after
testing show in following table:
Table-3: Performace of Model
I/P Drowsiness Detection Accuracy
Sample 1 Not Detected 0
Sample 2 Detected 94 %
Sample 3 Detected 96 %
As per each sample its improve its accuracy.to calculatealert
following formula is used :
DrowsinessDetectionAccuracy=totalno.oftimesalertcames
as eyes close/(no. of alert not come as eyes is close+ no. of
alert is cames as eyes is close).
So, Detection accuracy is 96%.
As we used python for trained this model we used
computer vision concept for draw circle around eyes and
continuously capturing driver face. another python library
such as pygame used for playing buzzer sound and dlib used
for detect the face landmark.
To make perfect accident precaution system we build
pothole detection model. In this model we detect potholes of
any size in real timescenario.to accurately detect potholewe
used dataset that had various images of pothole of many
size.to build model we perform following steps:
i] Data collection: collect the data from various source
such as kaggle and real pothole image from road.
ii] Clean the dataset: In this we remove some image and
only used the data that help in accurate detection.
iii] Build model: To model building we used YOLO
v4(You only look once) architecture that help us in fast
detection of pothole while vehicle in motion. we used YOLO
v4 which is CNN based real time object detection. we used
YOLOv4-tiny weight with fixed resolution image. Later we
modify YOLOv4-tiny weight for train model on multi-
resolution. carry out final training process.
iv]Test model: test model on real time input and also
with video it gives accuracy 94%.
As we develop pothole detection for real time input its
required significant amount of processing power of GPU. for
that we used CUDA-dnn (Compute Unified Device
Architecture-deep neural network) which is created by
NVIDIA for provide better GPU utilization and perform
Graphical operation faster. We used python programming
with various python library suchascv2forimageprocessing,
time library for time in code for object, os library used for
access functionality of operating system and it's provide
interaction between user and os.
Result of performance of model after training and
testing:
Table-4: pothole model performance
Potholes Initially After detection Accuracy
crack 0 2 98 %
Low severity 0 3 95 %
High severity 0 4 99 %
As per above Table-4 therearevariousshapeofpotholesuch
as crack, low severity, high severity.it will detect pothole on
real time scenario with 14.77 fps.it will detect all pothole in
its range in real time.it happen because of cnn and yolo v4
working to train and test the data . As testing model had
accuracy 97%.

As per our study we build various model using machine
learning and deep learning algorithm for create perfect
accident Precaution system. we used SVM for Drowsiness
detection for compare eyes aspect ratio of previous and
current face condition that check blinking of eyes this model
gives accuracy is 96%.Model of road object detection had
accuracy 94%.Pothile using yolo v4 had Accuracy 97%.
Table-5: Model performance of all model
Model Algorithm Accuracy
Drowsiness SVM 96%
Road object detection CNN 94%
Pothole detection Yolo 97%
As per Table-5 our system had accuracy 95%.
So it will prevent 95% of Accident.
5. RESULT
We implemented various model using various python
libraries: Numpy,Cv2,Matplotlib,dlib,pandas,tensorflow
pygame.
Following are drowsiness model:
Fig-5.1:Before Detected
After capturing Drowsiness sign it detectDrowsinessand
Draw square box around image with “You are drowsy”
message.
Fig-5.2:After Detected
Road object Detection Model Detected various objectlike
car, cycle, Traffic sign and other object:
Fig-5.3:Before Road object Detection
Above figure is show image before road object detection.
Now after detection various road object such as car,
person, traffic sign, motorcycle will detected by accuracy
score show in Table-1.
Fig-5.4: After road object detection
After detection of object it will show overall accuracy
94%.

Pothole Detection Model detect potholes present in road
from any shape and draw square box surrounding potholes.
Model had 97% accuracy of detection.
Fig-5.5:Pothole Detection
As per our study we build various model to make driving
secure, such as object detection, pothole detection and
Drowsiness detection as per there performance and
accuracy we can say this system will provide95%ofsecurity
to driver.
6. CONCLUSION
From this project we conclude that “Accident precaution
system for vehicle in motion"on account of Machinelearning
is proposed, which mainly aims at recognizing signs and
object and help prevents accidents resulting in safety for the
driver and pedestrians. By using image pre-processing, sign
and object detection, recognition and classification, this
method can effectively detect and identify signs and objects
and prevent accidents.Itusedlatesttechnologywithmachine
learning for classification and deep learning for object
detection with help of Yolo, Tensorflow, Python, CNN,R-CNN
algorithms.
REFERENCES
[1]ALEKSEJAVRAMOVIC,DAVOR SLUGA,DOMENTABERNIK,
DANIJEL SKOCAJ,VLADAN STOJNIC,"Neural-Network-Based
Traffic Sign Detection and Recognition of Images Using
Region Focusing and Parallelization “ October14,2020IEEE.
[2]Natalia Kryvinskaa,Aneta Poniszewska-Marandac,Michal
Gregusb,"An Approach towards Service System Building for
Road Traﬃc Signs Detection and Recognition " 2018
ScienceDirect.
[3]WANGHUA DENG,RUOXUE WU,"Real-Time Driver-
Drowsiness Detection System Using Facial Features"
AUGUST 2019 IEEE.
[4]Ping Ping,Xiaohui Yang,Zeyu Gao,"A Deep Learning
Approach for Street Pothole Detection" AUGUST 2020IEEE.
[5]Sohel Rana,Md. Rabbi Hasan Faysal,Sajal Chandra
Saha,Abdullah Al Noman, Kawshik Shikder,"Road Accident
Prevention by Detecting Drowsiness & Ensure Safety
Issues"June 2021 IEEE.
[6]Djebbara Yasmina,Rebai Karima,Azouaoui
Ouahiba,"Traffic signs recognition with deep
learning"NOVEMBER 2018 IEEE.
[7]Ganesh Babu R,Chellaswamy C,Surya Bhupal Rao
M,Saravanan M,Kanchana E,Shalini J,"Deep Learning Based
Pothole Detection and Reporting System",Sep 2020 IEEE.
[8] ouzia, Roopalakshmi R, Jayantkumar ARathod,Ashwitha
S Shetty, Supriya k,"Driver Drowsiness Detection System
Based on Visual Features" April 2018 IEEE.
[9]Muhammad Azhad bin Zuraimi,Fadhlan Hafizhelmi
Kamaru Zaman,"Vehicle Detection and TrackingusingYOLO
and DeepSORT" July 2021.
[10] Zhilong He ,Zhiguo Yan,"Traffic Sign Recognition Based
on Convolutional Neural Network Model", June 2021.
[11]Miss. Ankita M. Bhoyar,Prof. S. N.
Sawalkar"Implementation on Visual Analysis of Eye State
Using Image Processing for Driver Fatigue Detection",April
2019.
BIOGRAPHIES
Saniya Shaikh- Final Year student at
M.G.M COLLEGE OF ENGINEERING
Persuing Computer Engineering.
Omkar Prabhu- Final Year student at
Rahul Patil- Final Year student at

Somesh Nikam- Final Year student at
Prof.Sachin Chavan- Assistant
Professor at M.G.M COLLEGE OF
ENGINEERING Computer Engineering
Department.

Accident Precaution System For Vehicle In Motion Using Machine Learning

More Related Content

Similar to Accident Precaution System For Vehicle In Motion Using Machine Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

Accident Precaution System For Vehicle In Motion Using Machine Learning