Anomaly detection in crowd scene using historical information
Shu Wang, Zhenjiang Miao · 2010
Intelligent surveillance system is becoming more and more essential and important in public place because of more concern with people's safety. In this paper, we present a new approach to detect abnormal behavior automatically in public place. Instead of tracking every individual, we extract motion pattern to represent activity based on optical flow of some pixels, and motion pattern is described by a new descriptor we called histogram of motion vector. 3D grid structure is introduced to model the temporal-spatial relationship between motion patterns. We classify motion pattern into normal or abnormal group not only according to the deviation between motion pattern and trained model, but also according to its historical information. To demonstrate effectiveness of our approach, we present the results on public database.