Enhanced VGG19 Model for Accident Detection and Classification from Video

Abstract

Over the last years, the number of cars used in road traffic growth at a staggering rate. This situation had resulted in a significant increase of accidents and several traffic problems resulting huge losses. One of the most important road safety technologies is to automatically recognize dangerous situations and quickly share this information with nearby vehicles. In this work, we first, analyze various researches in the detection and classification of traffic anomalies and then propose to explore the potential of VGG19, which is a transfer-learning model to classify anomalies (accidents). In addition, we have compared the proposed algorithm to the other methods used. Our experience shows that our enhanced VGG19 model gives the best performance with 96% accuracy, and 0.99 AUC compared to the Convolutional Neural Network (CNN), which is the most widely used deep learning technique for image (accident image) classification, and the VGG19 models proposed over the last researches.