A SlowFast-Based Violence Recognition Method
Xingfan Diao, Yong Dong Xu · 2022
With the massive installation of cameras in cities, the requirement of equipment in an urban safety system has been basically satisfied. As a result, an available intelligent video-based safety system is very important. Video-based violence recognition method, which plays an absolutely important role in urban safety, seems to be a useful component of this system. However, a large part of existing violence recognition methods encounter the problems of low efficiency and inaccuracy owing to enormous challenges in accurately identifying the various violence events. In this paper, we propose a violence recognition algorithm based on the SlowFast model. In order to obtain a good performance, we modify the SlowFast model by both improving the convolutional structure and inserting plug-and-play modules. These two schemes can improve the speed and accuracy of violence recognition. The improved model achieves a high accuracy on several publicly available violence datasets surpassing some previous works and can be applied to violence detection in real-world scenarios, which is beneficial to fight crime and maintain social security.