Research on Anomaly Detection with Hierarchical Aggregation of Domain Context Integration
Haowen Jiang, Wei Wei, D.C. Yuan, Yuan Tang, Xin Qi · 2024
Abnormal behavior detection is a focal point in the fields of computer vision and video analysis. Faced with issues of insufficient detection accuracy and sample imbalance in existing models, this study proposes a detection strategy that integrates domain context with hierarchical aggregation. This strategy achieves precise identification of abnormal behavior by predicting future frames and computing the similarity difference with actual frames. The research first introduces a domain context hierarchical aggregation feature extraction module to optimize the integration and extraction of features. Secondly, a lightweight channel attention mechanism is employed to enhance inter-channel information interaction, improve target information capture, and enrich background details. Finally, a new loss function is designed to strengthen temporal and spatial constraints, ensuring consistency between appearance and motion information.The improved method demonstrates outstanding performance on the UCSD Ped1, Ped2, CUHK Avenue, and ShanghaiTech datasets, particularly achieving a 1% and 10.06% accuracy improvement on the CUHK Avenue and UCSD Ped1 datasets, respectively.