Human Violence Recognition in Video Surveillance in Real-Time

The automatic detection of human violence in video surveillance is an area of great attention due to its application in security, monitoring, and prevention systems. Detecting violence in real time could prevent criminal acts and even save lives. There are many investigations and proposals for the detection of violence in video surveillance; however, most of them focus on effectiveness and not on efficiency. They focus on overcoming the accuracy results of other proposals and not on their applicability in a real scenario and real-time. In this work, we propose an efficient model for recognizing human violence in real-time, based on deep learning, composed of two modules, a spatial attention module (SA) and a temporal attention module (TA). SA extracts spatial features and regions of interest by frame difference of two consecutive frames and morphological dilation. TA extracts temporal features by averaging all three RGB channels in a single channel to have three frames as input to a 2D CNN backbone. The proposal was evaluated in efficiency, accuracy, and real-time. The results showed that our work has the best efficiency compared to other proposals. Accuracy was very close to the result of the best proposal, and latency was very close to real-time. Therefore our model can be applied in real scenarios and in real-time.