Abnormal Behavior Detection in Video Surveillance Via Improved Patch Transformation

Abnormal behavior detection is a crucial area of research in video surveillance systems; it demands highly efficient and accurate techniques. Unfortunately, most of the available methods are computationally expensive, which can lead to significant performance degradation. In response to this challenge, we proposed a single-stream network for frame prediction as well as object-centric spatial and temporal transformation techniques to enrich the learning of regular features during the training stage. The proposed model has proven highly effective in empirically assessing three benchmark datasets. The model achieves impressive frame-level AUCs of 97.7%, 89.3%, and 73.1% on the UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets, respectively. These results highlight the exceptional performance of the proposed framework and demonstrate its potential to create a significant impact in the field of abnormality detection in video surveillance.