Spotting the Aggressor: Pose-Based Violence Detection Through Spatial-Temporal Deep Learning Techniques

The automatic detection of violent behaviour in video sequences has emerged as a critical area of research in public safety and surveillance, as the early detection of aggressive actions enables rapid intervention and can significantly mitigate potential harm. This paper proposes a novel methodology for the automatic detection and identification of violent behaviours in video sequences, with particular emphasis on the recognition of the specific individual responsible for such actions. A significant innovation of our approach is the integrated extraction of human pose features using a You Only Look Once-based (YOLO) model that efficiently captures critical key points which serve as essential cues for the detection of violent interactions. The proposed approach integrates human pose estimation techniques used to extract spatial features with temporal analysis models designed to capture the dynamic nature of aggressive behaviour. To assess the effectiveness of the method, two temporal architectures, a Bidirectional Long-Short-Term Memory (BiLSTM) network and a transformer-based model, were evaluated on the AirtLab dataset. Experimental results demonstrate the robustness and reliability of the proposed approach, highlighting high accuracy alongside real-time applicability. Furthermore, by relying on pose-based representations that can be processed in distributed edge-cloud architectures, the methodology offers enhanced privacy preservation compared to raw video processing approaches.