LAUSR

Abnormal violent behavior by people with mental disorders is common. When patients with mental disorders make some abnormal behaviors in public places, they may cause physical and mental harm to others and themselves. Thus, it is necessary to monitor the behavior of human with mental disorders under surveillance video. However, it is a comprehensive challenge to detect abnormal behavior of human (especially patients with mental disorders) based on abnormal detection and motion recognition technology. To address these issues, in this paper, we propose an end-to-end abnormal detection framework from a new perspective in conjunction with the Graph Convolutional Network (GCN) and a 3D Convolutional Neural Network (CNN). Specifically, we first train a one-class classifier to extract features and forecast abnormal scores in the framework. To improve the performance in abnormal behavior detecting, GCN is used to start modeling toward the similarity between video clips for the modification of noise labels. Then, based on this framework, GCN will recognize the normal behavior clips in the abnormal video and delete them, while the clips identified as abnormal behavior are retained. Finally, we use 3D CNN to extract video features and classify abnormal behaviors. In order to better detect the violent behavior of patients with mental disorders, the paper focuses on the UCF-Crime dataset of violent behavior. By experimenting with the dataset, the classification accuracy reaches 37.9%, which is 9.50% higher than that of the current state-of-the-art approaches. This proves the feasibility of classifying abnormal behaviors with this framework.