Efficient abnormal behavior detection with adaptive weight distribution

In recent years, there has been notable progress of recognizing and detecting human abnormal behavior in the field of computer vision. However, it remains an exceptionally difficult task due to serious background interference in surveillance videos, multi-scale variations, inter-target occlusions, and constraints on real-time performance. To address these problems, this paper proposes a new framework called efficient abnormal behavior detection (EABD) that simultaneously integrates spatio-temporal feature modeling and long-term dependency modeling. Then we introduce a new block for adaptive weight distribution to avoid noise interference. Meanwhile, we incorporate the detection head with attention and the SIoU loss function to improve network performance for detecting targets at various scales. Finally, the SoftNMS strategy is employed to enhance the prediction effect for overlapping objects of specific video frames in the inference stage. We conduct extensive experiments on four benchmark datasets, i.e., UCSD Ped1, UCSD Ped2, ShanghaiTech, and CUHK Avenue datasets. Our proposed EABD achieves AUC of 97.8%, 98.7%, 86.7%, and 95.4%, respectively. The experimental results show superiority over other related methods and demonstrate the effectiveness of our proposed method. Additionally, it achieves a maximum inference speed of 71.9 fps.