Abnormal Behavior Detection in Crowd Scene Using YOLO and Conv-AE
Yajing Li, Zhongjian Dai
Abstract
This paper proposes a weighted convolutional autoencoder (Conv-AE) and a novel regularity score based on the results of You Only Look Once (YOLO) network to detect abnormal behavior in crowd scenarios. The weighted Conv-AE extracts spatial features of video frames. In the training process, a weighted loss function is proposed based on the YOLO detection results, which emphasizes the foreground part, and thus overcomes the impact of complex background. In addition, a novel regularity score is put forward in the anomaly detection process. The regularity score takes into account the three factors of reconstruction errors obtained from weighted Conv-AE, speed information and category of objects detected by YOLO. Three scores respectively based on these factors are integrated to obtain anomaly detection results. The experimental results on UCSD ped1 and ped2 dataset verify that the proposed method achieves better performance than the most of semi-supervised methods.