A Multi-Modal Approach for Fight Detection in Surveillance Systems: Integrating Audio and Video Analytics

The detection of fights in surveillance systems demands public protection yet the current methods heavily depend on visual information which creates weaknesses during poor lighting conditions and when targets become obstructed or their movements are unclear. The proposed work introduces an innovative audio-video integration process to boost the detection performance. Our solution combines CNN-BiLSTM video-based fight detection with an audio transformer that classifies the sounds of fight activities such as shrieks and verbal assaults and hitting sounds. The combination of late fusion between audio and video capabilities produces a strategy that substantially improves detection reliability. Our system shows superior performance compared to single-mode solutions by reaching 89.2% precision together with 87.8% F1-score. The new research platform adopts AI capabilities in violence detection by establishing better standards that lead to smarter real-time security solutions.