An Improved SlowFast Model for Violence Detection

As social safety concerns become increasingly prominent, the importance of violence detection has become increasingly significant. Accurately identifying violent behavior helps facilitate timely interventions, reducing harm and losses, thereby enhancing public safety. To minimize the interference from irrelevant actions (e.g., those from spectators and uninvolved individuals) during violence detection, this study proposes an improved SlowFast method that enhances the robustness of the model by modifying the network structure of the Slow Pathway and lateral connections. In the Slow Pathway, the original residual block is replaced with an SE-Res residual block, which strengthens the weighting of channel importance, making the model more stable in complex scenarios. The lateral connections are improved by integrating a 3D-Convolutional Block Attention Module (3D-CBAM), enhancing feature fusion and increasing the model’s sensitivity to key features across different temporal scales. Furthermore, to address the class imbalance issue within the dataset, the Focal Loss function has been modified, effectively improving the performance of each class. The improved model demonstrates excellent performance in violence detection tasks, achieving an accuracy of $\mathbf{9 6. 6 7 \%}$, which is a $\mathbf{1. 1 2 \%}$ improvement over the classical model.