Movenet on frame difference: a motion-aware pose-based preprocessing approach for real-world violence detection

Abstract

This study examines whether preprocessing can improve automated violence detection in surveillance videos affected by background clutter, lighting variations, occlusion, and irrelevant visual information. The proposed MoveNet on Frame Difference method first uses frame differencing to highlight motion and remove static background content, followed by MoveNet multipose estimation to extract human skeletal information. Tested on the RWF-2000 dataset using the same EfficientNet-B0 and LSTM architecture, the proposed method achieved the best performance, with 81.50% accuracy, 88.50% recall, 0.8271 F1-score, and 0.8661 AUC. The findings show that applying frame differencing before pose estimation provides a more effective representation for violence detection than the reverse processing order.