Study On Abnormal Event Detection Based On The Multi-instance Learning

In video surveillance system, people pay a lot of attention to the abnormal events all the time. However, it is not able to meet the requirement for timeliness and accuracy of video surveillance system to manually mark the labels of the activities. As a result, it is more and more important to handle this problem in intelligent video surveillance based on the technology of automatic detection of activities.There are usually two kinds of approaches for anomaly detection according to the different point of view of dealing with the problem.The two kinds of methods are called motion-based and trajectory-based respectively.The motion-based methods can only model the variation of speed and direction of the pixels or blocks, so they limit to scenes that have obvious regular movements and only contain simple motion patterns such as the exit of the subway. Trajectory-based approaches are widely used in the case where executor of activity needs to be clearly determined such as in traffic surveillance. They are also used in many cases where we cannot obtain videos of the high resolution and good quality or even the videos themselves, for example, in the exploration of plants, the received signals from radars may be only the positions of objects.However, in trajectory-based abnormal activity detection, abnormal trajectory sometimes only has abnormality in some local area of the whole trajectory with the rest being normal. However, most of the previous approaches are not able to detect this kind of abnormality efficiently. To address this problem, we propose a novel approach for abnormal event detection based on multi-instance learning. In this approach,firstly, every trajectory is segmented into independent sub-trajectories. Secondly, the segmented sub-trajectories are modeled by sequence learning model. Finally, within the multi-instance learning framework, the whole trajectories are considered as bags, while normal ones are negative bags,abnormal ones are positive bags, and sub-trajectories are instances in the bags. Experimental results show our proposed approach achieves higher accuracy and lower false alarm rate than most previous ones.