Traffic Accident Detection and Classification in Videos based on Deep Network Features
Zain Ul Arifeen, Jang‐Eui Hong, Bo-Seok Seo, Jae‐Won Suh · 2023
The number of vehicles on the road has increased significantly, which pose a number of challenges to cope with for traffic management. Especially road accidents need instant attention to reduce the loss of people life and their property. In this paper, we propose a traffic accident detection and classification framework, which automatically detects accident in traffic videos using deep networks features and also classify that accident into car-car and car-bike collisions. The proposed framework works in two phases. Accident anomaly detection: We explore three convolution neural networks (CNN’s) named GoogLeNet, AlexNet and VGGNet, where deep features are extracted using these networks and a one class support vector machine (OCSVM) is trained on each network deep features, which are used to detect accident anomalies in a outlier fashion. Accident anomaly classification: where a multi-class SVM model is trained using the features of the detected accident frames and is used to classify accident into car-car and car-bike collisions. The experimental results on UCF-Crime road accident video sequences show that the proposed approach achieves high accuracy on both traffic accident anomaly detection and classification.