TEAF-Net: Temporal Entropy-Guided Adaptive Fusion for Automated Violence Detection in Surveillance Videos

With the fast growth of video surveillance in public places, manual analysis of video footage to detect abnormal behaviours has become infeasible. Automating the process seems like the obvious solution, yet it remains incredibly challenging. For instance, the way actual violence appears on camera depends on the view angle, crowd density, and the way the events unfold over time. Due to this, models that perform perfectly on curated datasets often fail in real-world scenarios. To address this problem, we proposed TEAF-Net (Temporal Entropy-Guided Adaptive Fusion), a violence detection architecture that combines multiple pretrained CNN backbones with an adaptive fusion mechanism regularized by bounded entropy. Our designed method learns input and time-dependent fusion weights while preventing degenerate behaviors such as weight collapse or near-uniform weighting, thereby promoting complementary specialization across temporal contexts. Furthermore, we use a bidirectional LSTM (BiLSTM) to track the temporal progression of the event, with a temporal attention module to settle on the most informative video segments before making a final decision. TEAF-Net was evaluated on three public benchmark datasets, namely Hockey Fights, RWF-2000, and RLVS, achieving accuracies of 99.50%, 98.92%, and 98.08%, respectively, with consistently high precision, recall, F1-score, and AUC values. Ablation studies confirm that performance gains arise from the combined effect of multi-backbone diversity, hierarchical adaptive fusion, and bounded-entropy regularization. Although the use of four parallel backbones increases computational cost (30.4 GFLOPs per clip), the proposed framework provides a robust and effective solution for automated violence detection in real-world surveillance scenarios.