Research On Unsupervised Abnormal Events Detection In Video

With the increasingly enrichment of material life, security issue has become a topic concerned by the public. As an important part of the security system, abnormal events detection attracts more and more attentions from the researchers. The abnormal events detection includes supervised approaches and unsupervised approaches. Manual labels of normal events and abnormal events are required for supervised abnormal events detection. But it is difficult to obtain the manual labels. The supervised methods can only detect the events which are same or similar in the training examples. The detection result for other abnormal event is poor. Thus, the supervised abnormal events detection is highly desired. Contrary to supervised abnormal events detection, unsupervised abnormal events detection can detect various abnormal events without manual work. Unfortunately, the existing unsupervised methods are designed for specific scenes, and the detection performance will greatly degrade for another scene. Besides, the computational complexity of unsupervised abnormal events detection is very high, which is difficult to satisfy the real-time requirements for the surveillance video. Without considering some particular video scenes, we extract video features and establish a basic event model to detect and update the abnormal events. At last, an abnormal events detection module is proposed for the video surveillance system.The main contents focus on the following aspects:Firstly, an improved trajectory extraction algorithm is proposed. In order to extract completed objects, we analysis the inherent characteristic of the video surveillance. By combining the current mainstream moving object extraction and tracking methods, a moving object extraction algorithm is proposed based on the location of the block and edge feature matching frame difference. For tracking, we add the edge and direction information to achieve more accurate and faster tracking results.Secondly, by tracking moving object and analyzing the regular pattern of trajectory, we extract the features of trajectory and apply Gaussian mixture model to the detection of abnormal events. The result is in the form of the performance of the possibility, no longer a simple normal events or abnormal events.Thirdly, in order to analyzing the video as a whole, we see the video segments as a unit to describe the degree of the intensity of exercise with exercise intensity and the degree of movement disorder with the entropy. By drawing and modeling the curves of the exercise intensity and directional entropy, we establish the correspondence representation of the probability of events to detect the group abnormal events.Fourthly, the concept of the degree of acquaintance is proposed to describe the relationship of the moving object. The synchronization and the communication are used to measure the degree of acquaintance to predict the abnormal events, making the detection of abnormal events in advance, rather than after the occurrence. It provides a possible research direction to the development of the abnormal events detection.We demonstrate the effectiveness of the proposed methods on the real video. Experimental result shows that the proposed features can describe the event accurately, the events detection model can be adapted to a variety of video scenes and complete the task of detecting abnormal events in real time.