A CNN-Based Road Accident Detection and Comparison of Classification Techniques

Abstract

The likelihood of accidents rises due to an increase in the number of vehicles on the road. Traffic collision detection can help to lessen fatalities and catastrophic injuries. This paper considers the different frames in road accident videos, creates a deep learning feature extraction algorithm, and classifies images as accident or non-accident. The last pooling layer output from the Convolutional Neural Network (CNN) is used to extract features that should be used for detecting an accident. Different layers of CNN with filters are used to extract features and also padding is applied to avoid any information loss in the convolution process. The fully connected layer classifies after receiving the extracted features. CNN overcomes the different shortcomings of conventional approaches for accident analysis. Also, different machine learning classification algorithms are used with the feature extraction by CNN to improve the accuracy of the classification of images as accident or non-accident.