Abstract
Violence detection can improve the ability to deal with emergencies, but there is still no data set specifically for violence detection. In this work, we propose VioData, a datasets specialized for detection in complex surveillance scenarios, and to more accurately assess the efficacy of these datasets, we propose a violence detection model based on target detection and 3D convolution. The model consists of two key modules: spatio-temporal feature extraction module and spatiotemporal feature fusion module. Among them, the spatio-temporal feature extraction module consists of a spatial feature module that extracts key frames using ordinary convolutional networks and a temporal feature extraction module that establishes temporal features using 3D convolution. The spatio-temporal feature fusion module Channel Fusion and Attention Mechanism (CFAM) fuses the temporal and spatial features. The experimental results indicate that the precision of the suggested detection model on UCF101-24, JHMDB behavioral detection datasets, and our proposed violence detection datasets, VioData, is improved compared to other violence detection models, which not only verifies the validity of the datasets, but also provides a baseline for the subsequent research and improvement in this area.