Video Violence Detection Method Based on Multi-Feature and Graph Convolutional Network

In the task of detecting violent video content, data scale is typically massive, presenting a challenge to performance. This study introduces a novel approach to address this issue by employing the DEtection Transfomer (DETR) model for processing Regions of Interest (ROI) in video frames. This reduces redundancy and decreases the model's dependence on the scene, thereby enhancing detection performance. Subsequently, we extract RGB, optical flow, and audio features from the video and employ a temporal fusion strategy to integrate these multimodal features sequentially. Finally, we employ a classification prediction method based on Graph Convolutional Network (GCN) models. This comprehensive approach overcomes the data scale challenge, improves performance, and opens new avenues for future research in video content analysis.