A Residual Network for Traffic Accident Images Classification Based on Optimized Deep Learning

Abstract

Reporting Vehicle Accident recognition based on computer vision using deep learning techniques has achieved reasonable results. Two problems regarding these techniques are computational complexity for the generated network, and accuracy of recognition. In this paper, a modified ResNet-based accident image recognition network is proposed. It is used as a feature extractor. Only the most important features will be selected using greedy stepwise. These selected features are used as input data for the KNN accident classifier. Two datasets have been used for this purpose. The results show that the proposed network has outperformed noticeable accuracy of about 96.9% in 29.7167 sec training compared to 93.7 % accuracy and 58.06449 Sec training for ResNet18. 93.9 % accuracy and 150.5573 Sec training for ResNet50 and 95.0% and 270.4034 Sec training for ResNet 101. The accident images are made understandable by adding descriptions using YOLOV2, which can be used for reporting the accident

Keywords- PCA; ResNet deep CNN; Yolov2; KNN; QSVM; CSVM.