Real-Time Violence Detection Using Deep Neural Networks and DTW

With the increased coverage of CCTV cameras across major metropolitan cities, there has been an increased need to continuously monitor these video feeds for them to be effective, this incurs human labor which is prone to boredom and flaws. We propose a system that can augment the capabilities of surveillance systems by automatically tagging sequences of violence in real-time. Instead of relying on a bigger network like CNN-LSTM, our system uses DTW to measure the similarity between two sequences of body joint angles over time and uses this feature as an input to different classifiers, thus reducing the need for a big video dataset to perform action recognition. For person detection on the individual frames we employ YoloV3 and then DeepSort is used to track different individuals over time then use this as an input to OpenPose to obtain the key points, the obtained key points are used to calculate different joint angles. The joint angles can be seen as temporal signals which can be compared with different joint angle signals over time and find the similarity measure and use this as a feature to classifiers like SVM, KNN, and Random forests. The results are at par with the ones that use CNN-based classifiers trained on larger data.