Research Of Abnormal Events Detection In Videos

Video surveillance is an important direction of application of the image processing field, but till now the intelligence of video surveillance system aiming at detecting abnormal events is not high enough to satisfy the demand of people in modern society. Our former laboratory research group implemented the crowd gather event detection, the riots panic event detection and human gesture recognition in crowd scenes after their research on intelligent video surveillance technology. This paper mainly discuss the technology of abnormal events detection in intelligent video surveillance, and attempts to achieve more event detection function on the basis of the former group, including objects abandoned, objects removed, people loitering, people falling. The main research work is as follows:First, the detection method of target of interest, including static objects and dynamic humans is studied, and brief analysis of the algorithms for target detection is performed. Considering the detection of static and dynamic objects at the same time is necessary, background subtraction method is used to extract foreground region, and motion features as well as appearance is merged to judge the foreground region corresponds to object or body. On this basis, region growing algorithm is employed to distinguish abandoned object from removed object, and abnormal events relating objects is detected. Secondly, this paper studies the tracking method of moving human body, and a method based on boosting learning is adopted to track the foreground area which previously identified as human target. After the motion trail of human target is acquired, entropy of motion direction change and absolute motion direction of the motion trail are computed to determine whether it is loitering trajectory or not, and thus loitering event is detected. In order to make up for the deficiency of the target detection, human location based on color histogram is introduced, weakening the effects of occluding to a certain extent. On the foundation of target detection, people falling is recognized based on motion feature.Experiments have been carried out for foreground detection, tracking and abnormal events detection in the Xiamen university video datasets based on the above research work, respectively. Results show that under the particular scenario, the method in this paper can identify and differentiate events of objects abandoned, objects removed, people loitering and people falling relatively effectively, enriching the event detection function of the video surveillance system developed by our laboratory research group.