Abnormal event detection in video surveillance / Lim Mei Kuan

Lim, Mei Kuan (2014) Abnormal event detection in video surveillance / Lim Mei Kuan. PhD thesis, University of Malaya.

Preview

PDF Download (12Mb) | Preview

Abstract

The recent Boston Marathon bombing and the kidnap of a British boy at Lake Titiwangsa, have ignited a pressing interest for automated video content analysis to assist the law enforcement in preventing such events from recurring. Post-mortem investigations surrounding such cases often found that there were missed opportunities for using technology to detect the abnormality of the suspects, which lead to those tragedies. Therefore, this thesis aims to develop computer vision solutions to identify regions or behaviours, which could lead to unfavourable events, as a cue to direct the attention of security personnel for a more effective and proactive video surveillance. The first contribution of this thesis introduces a robust visual tracking algorithm that is able to locate moving objects in surveillance videos. A great challenge in this domain is the capability of dealing with complex scenarios of tracking abrupt motion, such as switching between cameras, which is very common when the number of CCTV to be monitored is enormous. Conventional sampling-based predictors often assume that motion is governed by a Gaussian distribution. This assumption holds true for smooth motion but fails in the case of abrupt motion. Therefore, by considering tracking as an optimisation problem, the proposed SwATrack algorithm searches for the optimal distribution of motion model without making prior assumptions, or prior learning of the motion model. Experimental results have shown that the proposed SwATrack improves the accuracy of tracking abrupt motion, with an average accuracy of 91.39%, while significantly reduces the computational overheads, with an average processing time of 63 milliseconds per frame. Visual tracking of objects at mass gatherings such as rallies can be daunting due to the large variations of crowd. Hence, the second contribution proposes an alternative solution that deals with dense crowd scenes. A new research direction that identifies and v localises interesting regions by exploiting the motion dynamics of crowd is proposed. Here, interesting regions refer to abnormalities, where they exhibit high motion dynamics or irregularities. This assumption alludes to the social behaviours and conventions of humans in crowded scenes. Therefore, the possibility of abnormal events taking place is considered likely, when there is high motion dynamics and irregularities. Experiment results have shown an average accuracy of 78% on the defined dataset. The third contribution aims to provide an integrated solution to detect multiple events in different regions-of-interest of a given scene. This is very critical in the real-world scenarios where multiple events may take place in a scene at the same time. Existing solutions such as CROMATICA and PRISMATICA are commonly limited to detect single events, at a particular time. On the contrary, the proposed solution provides flexibility to deal with different environments, for a broader degree of scene understanding. The key idea is to conceptually decompose information obtained from a given scene into several intermediate degrees of abstractions. These low-level descriptions are then integrated using a basic set of rule-packages, to discriminate the different events. Experimental results on fives scenarios of abnormal events have shown an average accuracy of 83%.

Actions (For repository staff only : Login required)