Motion Abnormal Detection Base On Dense Trajectory-aligned

The abundant semantic information contained in trajectories can express the interaction between moving objects in time and space.In the field of intelligent video surveillance,how to effectively propose motion targets and describe the pattern of the motion of the target has been the focus and difficulty of the research.After a summary of previous research work,we found two main research directions.Firstly,we extract the spatial trajectory of the target,and the trajectory is modeled to find typical motion patterns in scene structure.Secondly,based on the grid division of image frames,the appearance and movement feature including the spatial information is extracted from the grid,and the typical mode of motion in the video scenes is found by clustering.We effectively combine the advantages of the above two methods,that is,the powerful tracking ability of the trajectory feature and the powerful description ability of the local spatiotemporal features.Several experiments are carried out on the two sub tasks of motion proposal and anomaly detection respectively,and the main work is summarized as follows:(1)In order to overcome the shortcomings of existing motion proposal algorithm is generally complicated and inefficient,we suggest skipping the video segmentation steps directly,and we propose a method based on dense trajectory-aligned and fusion descriptor to capture the key information of the video target.Firstly,the space trajectory of the target is extracted,and the length of the trajectory is restricted to ensure the accuracy of the trajectory in the crowded scene,the high density of the trajectory is maintained to ensure the coverage of the video foreground region well.Then,the fusion influence descriptor is extracted along the dense trajectory direction,which can better describe the motion pattern of the target in a period of time.In addition,the combined HOG,HOF and MBH fusion descriptors are also more robust to the change of the target,and can eliminate the background drift caused by the camera jitter.(2)In order to overcome the shortcomings of existing anomaly detection algorithms in terms of target tracking and description in crowded scenes,we apply the proposed motion proposed framework to the anomaly detection task and advance the moving target before detecting the anomaly.It not only can greatly improve the time performance of the detection system,but also further reduce the negative impact of the background area on the detection system.In addition,in view of the typical characteristics of abnormal motion,the motion influence descriptor is extracted along the dense trajectory direction,which can better describe the motion pattern of the target in a period of time.Finally,compared with the other methods,most of them are concentrated in one of the global and local abnormal detection,we further proposed an overall framework in which we can accurately detect both global and local abnormal motion.In the experiment of multiple public dataset,we compared the performance of the proposed method with that of other state-of-the-art methods and showed that the proposed method outperforms these competing methods.