Automatic Traffic Accident Detection System Using ResNet and SVM

The rate of road accidents has increased to a large extent over the last few years. This has eventually resulted in a huge loss of lives and property. Hence, the need of the hour has aroused to detect such accident spots as quickly as possible so that proper life-saving measures can be taken and the mishap prone areas can be put on alert. In order to address this problem, we propose a Machine Learning and Deep Learning based model on the concepts of Clustering and Classification that can be used to detect accidents from the traffic surveillance cameras. Firstly all the videos are split up into smaller shots according to scene changes. And then key frames are extracted from each shot based on histogram difference of consecutive frames. Then distance between the vehicles is determined to detect the potential accident. The obtained key frames are passed through a ResNet50 architecture for feature extraction. After obtaining the feature vectors of all videos, K-Means clustering has been applied to obtain Bag of Visual Words(BOVW). Finally, these bag of visual words is sent as input to a Support Vector Machine(SVM) classifier that outputs if a video contained an accident or not. The proposed method has an accuracy of 94.14%.