Drone Video Anomaly Detection by Future Segmentation Prediction and Spatio- Temporal Relational Modeling

In traffic surveillance, accurate video anomaly detection is vital for public safety, yet environmental changes, occlusions, and visual obstructions pose significant challenges. In this research, we introduce DAD-FSM, an innovative drone-based video anomaly detection system that leverages a spatio-temporal relational cross-transformer to enhance the encoding of visual and temporal features for future segmentation. Additionally, we propose the motion-aware frame prediction loss function (MAFL) to improve the model’s representation and the background and foreground separation of moving objects. Our method achieves state-of-the-art (SOTA) AUC scores of 68.13% on the UIT-ADrone dataset and 73.5% mAUC on the Drone-Anomaly dataset, surpassing previous methods by 2.68% and 5.71% respectively. The approach is further validated on the CUHK Avenue dataset, underscoring its global applicability and effectiveness in diverse traffic scenarios. These results demonstrate the potential of our model for broad use in traffic surveillance applications.