A Lightweight Network for Violence Detection
Wei Wang, Shuai Dong, Kun Zou, Wensheng Li · 2022
Video violence detection is an application area under the field of action recognition, which refers to the detection of violent behavior in video sequences. Existing methods or deep learning models, while capable of effective detection, are not as efficient when applied to real-time violence detection scenarios. Based on the purpose of reducing model computation and coping with the specificity of violence detection scenarios, a new violence detection network is proposed in this paper. This network uses depthwise separable convolution to reduce the number of model parameters and the number of computations; in addition, the network is based on the idea of mixing 2D and 3D convolution to further reduce the number of parameters. The model proposed in this paper is compared with other mainstream models on several datasets (Movies, Hockey, Crowd, RWF-2000). The experiments show that this model significantly reduces the number of parameters as well as improves the inference speed within an acceptable range of accuracy degradation.