Stalker Retrieval on Surveillance Videos Using Spatio-Temporal Coappearance

摘要

Stalking is a distinctive form of criminal activity that consists of repeated following, watching, tracking and/or harassing by one person against another. Stalking victims are reported many daily happenings, thus stalking has attracted widespread public concern in recent years. It results in that stalking analysis becomes a significant social problem in criminology. Meanwhile, nowadays surveillance camera has become a ubiquitous aspect of the modern urban landscape, widely installed for security purposes. Video data generated by surveillance cameras are recognized as reliable sources for crime prevention and investigation, definitely useful for stalking analysis. However, since stalking behaviors have complicated variants, it is too difficult to directly analyze stalking behaviors in surveillance videos in terms of behavior analysis. In the state of the art, even there is no existing technology that can handle such analysis. Hence, motivated by this and inspired by the distinctive characteristics of stalking behaviors, in this paper, we define a new query type, stalker query, to retrieve potential stalkers. We then propose efficient algorithms to handle the stalker query, and conduct extensive experiments based on real surveillance videos to evaluate the efficiency and scalability of our algorithms. The experimental results show that our approach outperforms the baseline method over two orders of magnitude to return stalker candidates. Besides, we also examine the high accuracy of retrieval by a data set containing simulated real stalkers. This data set will be contributed to academic communities for attracting continuous effort to tackle remaining open problems with high social impact.