Object and patch based anomaly detection and localization in crowded scenes

摘要

Detecting and localizing anomalies in crowded scenes is an ongoing challenge for public security. Existing approaches are mainly based on patches and trajectories. However, they fall short in semantic understanding of scenes and tackling the depth-of-field problem, respectively. In this paper, we put forward a novel object and patch based framework for anomaly detection and localization. Specifically, we propose to colorize images for precise object detection in dim scenarios. Categories of the objects are used for appearance anomaly justification. For motion anomaly, we propose a new patch-based algorithm which is robust to the depth-of-field problem, which can also be used to detect location anomalies. Besides, a new object re-targeting method is proposed to find the missing objects in detection. It can also handle drift and occlusions in tracking, which can avoid false alarms. Extensive experiments are conducted on several benchmark datasets for anomaly detection. The results show that the proposed method can achieve comparable accuracy in anomaly detection with state-of-the-arts methods and at the same time, localize anomalies precisely.