Road Accident Prediction using Deep learning

DOI : 10.17577/ICCIDT2K23-228

Download Full-Text PDF Cite this Publication

Text Only Version

Road Accident Prediction using Deep learning

ISSN: 2278-0181

ICCIDT – 2023 Conference Proceedings

Malavika Prasad(Author)

Dept.of computer science and engineering Mangalam college of engineering, Ettumanoor,India

Nandana K Saji(Author)

Dept.of computer science and engineering Mangalam college of engineering, Ettumanoor,India

Neenu Joseph(Author)

Dept.of computer science and engineering Mangalam college of engineering, Ettumanoor,India

Eldhose K Paul(Author)

Dept.of computer science and engineering Mangalam college of engineering, Ettumanoor,India

Volume 11, Issue 01

Published by,

AbstractA database of the traffic accidents was

data mining R, traffic injury


organized and analyzed, and an intersection accident risk prediction model based on different mechanical learning methods was created to estimate the possible high accident risk locations for traffic management departments to use in planning countermeasures to reduce accident risk. Using Bayes' theorem to identify

techniques BhuvanIeCshCwIaDr T – 20se2v3erCityonference Proceedings

i R

Accident Adnan Bin Uses the alarm The avoidance Faiz, Ahmed pulses and smartphones

  1. system using Imteaj, vibration sensors could IR transmitter Mahfuzulhoq system as the provide false

    Chowdhury first level of data sometimes safety

    environmental variables at intersections that affect

    accident risk levels, this study found that road width, speed limit and roadside markings are the significant risk factors for traffic accidents. Meanwhile, Naïve Bayes, Decision tree C4.5, Bayesian Network, Multi- layer perceptron (MLP), Deep Neural Networks (DNN), Deep Belief Network (DBN) and Convolution Neural Network (CNN) were used to develop an accident risk prediction model. This model can also identify the key factors that affect the occurrence of high-risk intersections, and provide traffic management departments with a better basis for decision-making for intersection improvement.

    Keywords Byes theorem, Deep Neural Network (DNN) and Convolution Neural Network (CNN)

    1. INTRODUCTION (Heading 1)

      A high accident risk prediction model is developed to analyze traffic accident data and identify them priority intersections for improvement. A traffic accident database was organized analyzed. An Intersection Crash Risk Prediction

      Android Dnyanesh

      application for Dalvi, Vinit It is integrated

  2. automatic Agrawal, with Slow in Accident Sagar Bansod, multimodal responding detection Apurv Jadhav, alert

    Prof. Minal dissemination Shahakar


    detection and Md. Syedul

  3. reporting Amin, Jubayer Capture the Limited rate of system using Jalil, location of data transfer GPS,GPRS and M. B. I. Reaz vehicle

    GSM accident


    Real time Hossam M.

    traffic accident Sherif, Long distance It cannot be

  4. detection Hossam M. data collection used for high system using Sherif, Samah and speed wireless sensor A. Senbel transmission


Table 1: Literature survey of accident prediction model


The rapid development and wide application of computer

Model Based on Different Machine Learning methods for technologies, computer network technologies, multimedia and estimating potential high accident risk locations for traffic communication technologies, and the Internet of Things fields [1], has management have been developed department to use in planning driven the recent development of intelligent road traffic management countermeasures to reduce the risk of accidents. Using Bayes systems [2]. Li et al. The Internet of Things allows for the collection of theorem identify environmental variables at intersections that various kinds of information through sensors [3], each of which influence the level of crash risk, this study found that the width represents an independent information source [4] from which data is of the road, the speed limit and the markings along the road are collected at a certain frequency for categorization and analysis.Each significant risk factors for traffic accidents.Meanwhile, Naïve independent information source would sense, measure,capture and Bayes (NBD), Deep Neural Networks (DNN) and Convolutional transmit information anytime and anywhere.The development of Neural Networks (CNN) were used to develop the accident risk advanced chip design and new materials have also increased the utility

prediction model.

and longevity of such sensors [5], while also allowing for anti-

This model can also identify key factors that influence the interference, multi-mode, and self-adapting features [6].

occurrence of high-risk intersections, and provide operations

These developments provide the technological basis for intelligent

management departments with a better basis for decision-making expressway management systems, integrating Internet of Things intersection improvement. Using the same environmental applications due to the introduction of mass information compatibility. characteristics as high risk intersections for model inputs to High-speed wired and wireless networks have been integrated to create estimate the level of risk that may occur in the future, which can three-dimensional connections, ensuring the accuracy of data be used to prevent traffic accidents in the future. In addition, it information, wider transmission bandwidth, higher spectrum utilization, can also be used as a reference for future intersection design and more intelligent access, and more efficient network management [7]. environmental improvements.In practical applications, our The development of these advanced technologies mainly depends on proposed model can be used to predict probability (or "risk") NGN (Next Generation Network)communication network technologies accidents at different intersections by identifying similar and new wireless communication networks (3G, 4G, ZIGBEE) environmental variables, ie it enables authorities to take practical [8].Expressway construction and traffic is rapidly growing around the steps to effectively reduce incidence and severity accidents world, and the demand for social development is growing together with the costs associated with such accidents. In synchronously [9].

addition, research results identify important environmental

factors that influence the occurrence of traffic accidents. To

Improving the efficiency of existing expressway traffic

effectively reduce the risk of accidents, in recent years traffic infrastructure requires the effective collection and analysis of usage accident management agencies in countries around the world not data [10]. As cars and individual drivers are increasingly linked to only have established standards and operating procedures for wireless transmissions,drivers demand increasingly sophisticated traffic road surveys, but also sought to develop accident risk analysis information,allowing them to assess current local traffic and driving and forecasting methods. The it hoped that longitudinal crash conditions, predict future conditions, and identify optimal driving data would be used to identify and classify high-risk ones routes [11]. Expressway traffic management agencies also need to intersections, allowing efficient prioritization of scarce resources effectively monitor highway conditions andcoordinate timely

to minimize frequency and severity of traffic accidents.


    emergency response including police, rescue and repair units [12].




    ume 11, Issue 0





    Published b

    sensor networks can be applied to control subsystems and guid

    ysu, bwswywst.eijmerst.oirng the execution subsystem, and to improve s

    controller function to implement the bus priority function of


    A road accident

    prediction model using


    Viswanath,Pre ethi K,Nandini

    It helps to

    identify key factory of

    Requires large database, more

    The data to drive such coordination is sourced from sensor networks that monitor traffic and environmental conditions throughout the highway network. Such monitoring data can be used to improve and simplify signal control algorithms and traffic efficiency. Wireless

    ance ignal the

    intelligent transportation system [13]. Besides, the position sensor can help achieve functions such as energy-saving and emission reduction.

    1. METHOD

      1. Bayes theorem

        ISSN: 2278-0181

        ICCIDT – 2023 Conference Proceedings

        Three layers make up the neural network layer: an input layer, an output layer, and a hidden layer sandwiched in the middle. The neural network was designed with the intention of simulating how human neurons function. The output of this layer (matrix multiplication) is the linear combination of the inputs from the

        1. Bayes' theorem serves as the foundation for the Naive previous layer(s), which cannot be separated from the linear Bayes (NB) algorithm. Chiang (1995) suggested a whole data connection if the activation function is not utilised. In order for storage and analysis system for road traffic safety, including neural networks to express real complex models, the non-linear Bayes' theorem as the key analytical tool [7]. A known target activation function is employed to raise the non-linear factor of

          variable's prior probability, which is frequently available neurons [8].

          through training samples, is assumed by NB. Furthermore, the Sigmoid, Softmax, tanh, ReLU, and ELU are frequentlyused activation participating attribute values are presumptively independent of functions. An event (an element in a sample space) is mapped by the one another given any target variable or dependent variable. loss function to a real number that indicates the event's opportunity cost Assuming that training materials have a set of attributes X = or economic cost. The reduction of the loss function is the optimization [[X1, X2,…, Xn]], X does not contain the attribute for the target objective. As a result, the loss function determines how well the neural variable, and C is the set of values for the target variable's network model performs and what the optimization's objective is. attributes,[[C1,C2,…,Cm]].

          P(C|X) denotes the likelihood that a given collection of X C. Convolutional Neural Networks

          traits will be present for the target category C. P(C/X)=(P(X/C)*P(C))/P(X) (1)

          Deep learning has recently piqued the curiosity of academics and researchers across all disciplines. As a deep learning technique, the convolutional neural network has grown in popularity across many scientific disciplines. In the domains of computer vision, image recognition, and speech recognition, CNN is a rapid and effective feed forward neural network that

          According to the Naive Bayes theory, if each feature is has shown great results.

          assumed to be independent of the others, then equation (1)


          In recent years, the CNN model was created as a road traffic

          Where P(Xi/C) is the likelihood that feature Xi appears in accident prediction model for accurately predicting highway

          a class Cm,Cm C,

          P(C/X)=ni=1P(Xi/C)P(C)nj=1 P(Xj) (2)

          The prior probability of the class Cm,Cm C across the

          road traffic accidents, hence promoting the efficacy of

          prediction. In this study, the CNN model outperformed the classic back propagation neural network model in terms of accuracy and efficiency, with a prediction accuracy of 78.5%, 7.7%.

          By converting the gradient of the accident data into a grey

          board is P(cm). For a given set of features, the classifier's image that represents the weight of the traffic accident's

          output is the group with the highest probability. The

          characteristics, a deep learning strategy with a CNN model was

          proposed for predicting the severity oftraffic accidents. The grey denominator can be regarded as a constant because it is image was then fed into a CNN model that predicted severity. independent of C and the value of the features Xi is The Leeds City Council examined the performance of this provided. The probability of each class Cm,Cm C is suggested CNN model using data on traffic accidents from 2009

          computed in equation (2) to yield the maximum class,which to 2016 and found that it performed better than the K-nearest

          is argmax c =P(C = c) ni=1 P(X = Xi | C = c) (3)

          neighbour technique, logistic regression, gradient boosting

          where argmax c is used to represent the function that neural network, and support vector machines. provides the largest class.

      2. Deep Neural Networks


      The degree of information clutter reduction (benefit degree) can be determined by dividing the "expected information

      A deep learning framework called a deep neural network entropy before being partitioned by the target variable" by the (DNN) can be thought of as a neural network with numerous "expected information entropy before being partitioned by an hidden layers (Neural Networks). In a neural network, attribute," and choosing the node attribute that can artificial neurons are used to create a mathematical model produce the greatest benefit. s1,s2,…sm: A finite that resembles a biological neural network. Neurons are set of samples Category "C:" (c1, c2,…,cm)the quantity of typically arranged in layers, and connections are only made samples falling into a particular category Thenumber of samples between neurons in adjacent levels. The first layer receives that fall under a specific attribute value's (av) range (Ak)

      the input low-order feature vector, which is then

      transformed into a high-order feature vector by advancing Sij : The number of samples for a particular category under a the neurons over time. The number of categories is the same particular attribute value (av) of a particular attribute (Ak) (ci) as the number of neurons in the output layer.In order to

      represent the likelihood that the input vector falls into the Pi :The percentage of the sample (si/S) that falls within a specific appropriate category, the output vector is a probability category

      vector. The predicted calculation of one neuron and its

      output description are presented in Eq. (8), where aij is the Pij: The percentage of samples that match a particular attribute jth neuron in the ith layer and Wi is the weight of the value (av) of a particular attribute (Ak) for a particular category.

      Vnoeluurmone's11s,yInssaupese0,1 which connects the jth neuron Piunbtlhiseheidthby,

      layer with the kth neuron in the layer below (i.e. layer i-1) Equation (4) calculates the expected information entropyprior to


      partitioning by the target variable I(s1,s2,…,sm), which

      represents the post-segmentation degree of entropy ofthe target variable of the training set (Target variable,


      attribute values "a1, a2,…, am"

      ISSN: 2278-0181

      pendent variale). Ak: An attribute that contains the




      Input Data

      ICCIDT – 2023 Conference Proceedings

      I(S1, S2,…, Sm) = m pi (pii)=1


      The sample ratio of the training sample divided by the attribute Ak is calculated by equation (5). As an illustration, the attribute "gender" has the following two attribute values: "7male, 3female," and the sample ratio is {7/10, 3/10}.

      E(Ak)=v(S1j+…+Smj)/S*l(S1,S2,…,Sm) (5)

      For a specific attribute value (av) (female) for attribute (Ak), equation (8) generates the information entropy, which is I (s1,s2,…,sm). Prior to being divided by an attribute variable, we multiply the respective sample ratios to obtain the anticipated information entropy.

      I(S1j,S2j,…, Smj = mi=1 pij (pij) (6)

      Username Password

      Fig 2: User Data flow diagram

      1. Level 0-DFD



        Fig 3: User-login Data flow diagram

        Final result

        Gain (Ak) for a specific attribute node is obtained bydeducting E (Ak) from I (s1,s2,…,sm).

        Gain (Ak) = I (S1, S2,…, Sm) E (A) (7)

        Preprocessing Dataset

        Splitting data into training and testing


      Log out

    4. RESULT

      Predicting the likelihood of accidents at particular junctions is the goal of accident risk analysis. The danger level of each intersection is determined based on the numberof accidents and fatalities in the past. In order to estimate thelevel of accident risk at crossings when accidents have not yet happened, a risk prediction model for intersections is built by identifying the important environmental elements that influence the occurrence of accidents at crossroads.

      Collecting Dataset

      Training Dataset

      Model training using CNN

      Input Data

      Trained model

      Input to model

      Trained model

      Result prediction


Testing Dataset

This study analyses traffic accident data and identifies priority intersections for improvement using a high accident risk prediction model. There has been a significant increase in pedestrian injuries, as well as fatalities, over the past few years. For accident data for provincial highway intersections, risk grouping in terms of CBI was carried out. Different mechanical learning techniques were then employed to create a prediction model for high-risk intersections. The findings indicate that environmental factors including road width, the posted speed limit, and the existence of roadside markings are important indicators of the likelihood of an accident. It was simpler to pinpoint the environmental characteristics of low- and medium-risk crossings based on the frequency of accidents there. The relative lack of data hurt prediction accuracy for high-risk accident intersections, while decision tree rules and detection models were shown to offer respectable prediction accuracy for clusters of high-low and high-medium risk intersections. Additionally, it was discovered that the DBM model performs best for model training with unbalanceddata, whereas NB performs best for intersection risk prediction. The findings of this study can serve as a guidefor traffic management organisations to reduce the probability of accidents at intersections. This study will aim

Testing Accuracy

Fig1: Architecture of road accident prediction model

Volume 11, Issue 01

Published by,

    1. Level 1-DFD

ISSN: 2278-0181

ICCIDT – 2023 Conference Proceedings

to provide a platform for high-risk accident analysis and prediction based on intersecting environmental elements. The following objectives will be accomplished by data collecting and analysis based on the locations of traffic accidents:

  1. This platform's system can combine and analyse traffic information about the GIS layer and accident data, thus understanding the site of the accident as a whole. Afterward, by examining the impact of environmental elements at the scene of the accident and its causation multiple intersections allow us to create useful enhancement approaches as a guide for upcoming intersection design and improvements to the environment.

  2. Use predictive models to estimate the likely locations of high-risk accidents to allow traffic management authorities to better prevent high-risk road accidents or serious casualties.


[1] Intelligent Traffic Accident Prediction Model for Internet of Vehicles With Deep Learning Approach," by Da-Jie Lin, Mu-Yen Chen, Hsiu-

Sen Chiang, and Pradip Kumar Sharma, Senior Member, IEEE, 2021

[2] A road accident prediction model employing data mining technique, Dhanya Viswanath, Preethi K, Nandini R, and Bhuvaneshwari R, 2021

[3] "Accident Avoidance System Using IR Transmitter," IJRASET, Volume 5 Issue IV, April 2017, by Miss Priyanka M. Sankpal and Prof. P. P.

[4] Android Application for Automatic Accident Detection, 2017; Dnyanesh Dalvi, Prof. Minal Shahakar, Vinit Agrawal, Sagar Bansod,Apurv Jadhav.

[5] "Real time traffic accident detection system employing wireless sensor network," Md. Syedul Amin, Jubayer Jalil, and M. B. I. Reaz (2014).

[6] "Real time traffic accident detection system employing wireless sensor network," Md. Syedul Amin, Jubayer Jalil, and M. B. I. Reaz (2014).

[7] Q. Cai, M. Abdel-Aty, J. Yuan, J. Lee, and Y. Wu, Real- time crash prediction on expressways using deep generative models, Transp. Res. C, Emerg. Technol.,

[8] Data Mining Concepts and Techniques, J. Han and M. Kamber Morgan Kaufmann, San Mateo, California, USA, 2001, pp. 284287

[9] Deep spatiotemporal graph convolutional network for traffic accident prediction, L. Yu, B. Du, X. Hu, L. Sun, L. Han, and W. Lv, Neurocomputing, vol. 423, pp. 135147, Jan. 2021.

[10] L. Zheng and "A unique method for real-time crash prediction,"T. Sayed at intersections with signals," Transp. Res. Emerg, C. Technology, volume 117, 2020 August, Art. No. 102683.

[11] "Analysis of real-time crash risk for expressway ramps using traffic, geometry, trip generation, and socio- demographic variables," Accident Analysis and Prevention, vol. 122, no. 1, January 2019, pp. 378384. L. Wang, M. Abdel-Aty, J. Lee, and Q. Shi.

[12] "Crash prediction based on traffic platoon features utilising floating car trajectory data and the machine learning approach," Accident Analysis and Prevention, vol. 133, Dec. 2019, Art. no. 105320, by J. Wang, T. Luo, and T. Fu.

[13] Key feature selection and risk prediction for lane-changing behaviours based on vehicles' trajectory data: T. Chen, X. Shi, and Y.

D. Wong, Accident Analysis and Prevention, vol. 129, pp. 156169, August 2019.

Volume 11, Issue 01

Published by,