TIR-VD: bimodal CNN-BiLSTM with attention for robust violence detection in thermal infrared videos
Résumé
Automatic violence detection in videos has become a key challenge in computer vision. The proliferation of surveillance cameras and the rapid growth of video content in the era of big data have increased the demand for systems aimed at enhancing public safety and moderating online content. However, most existing methods rely on RGB images and struggle in low-light conditions. Thermal InfraRed (TIR) imaging, which captures heat signatures regardless of lighting, offers a complementary modality. This is crucial, as much urban violence occurs at night, where RGB-based methods often fail. In this work we introduce TIR-VD, a multimodal approach combining video frames and TIR images to improve detection robustness under poor lighting. In a preprocessing stage, TIR sequences for two public datasets, Hockey Fight and RWF-2000, were generated, using a generative model. Our method uses a bimodal CNN-BiLSTM, enhanced with channel and temporal attention mechanisms, to process each modality separately before applying a weighted fusion strategy for joint prediction. Our experiments demonstrate state-of-the-art performance on both benchmarks, confirming the benefits of thermal data for intelligent surveillance systems and enabling reliable violence detection across lighting conditions.