Abnormal behaviors recognition in crowd environments based on semi-supervised deep learning and hierarchical approach

This paper tackles the complexity of detecting abnormal behavior by focusing on temporal patterns, such as social force and optical flow, rather than spatial patterns. The CycleGAN system trains on normal behaviors, ensuring that abnormal behaviors are not reproduced during testing. Hierarchical supervised learning and transfer learning techniques are applied to recognize abnormal behaviors. Given the scarcity of abnormal patterns, geometric techniques are employed to augment these patterns, followed by training with temporal models. Experimental results on the UCSD and UMN datasets demonstrate that the proposed method significantly outperforms existing approaches, achieving average AUCs of 99.90% on Ped1, 99.89% on Ped2, and 99.82% on UMN.