The Research Of Moving Object Retrieval In Surveillance Video Based On Trajectory

With the development of technology and society, more and more surveillance camerasare installed in public areas, residential, shopping malls, and streets. On one hand,those cameras can monitor many areas in real time. On the other hand, the videoscaptured by those cameras can be used in forensic analysis in criminal cases in thefuture. In the latter scenario, if we simply manually search the videos needed, it willspend huge resources in manpower and time, because of the huge numbers of videos.So the automatically video retrieval becomes an urgent needs. The video contains alot of information, but at the same time the instructed characteristic of video data ledto the difficulty in automated video retrieval. This paper mainly studies the trajectorybased video retrieval, we carried out the relevant research on the video analysis,trajectory data obtained, trajectory preprocessing and representation of the trajectoryand its similarity measurement.The number of the target features obtained frame by frame in the video analysisprocess is huge, and the target’s patterns are usually very limited in the surveillancevideo. So there will exist a great data redundant. To solve this problem, this paperpresents a hierarchical clustering method based on the target’s appearance features toextract the key moving objects. The method utilizes the dominant color histogram andedge direction histogram to compute the similarities between the objects of a target,and it can well reflect the local and global changes of target.Because of the inherent limitations of the current tracking algorithms and theocclusion problems in video, there may exist the situation that more than onetrajectories belong to a same target in the trajectory set obtained by the video analysis.This will influence the quality of the trajectory representation and retrieval. To solvethis problem, this paper propose a bi-temporal continuity of the trajectory fragmentsassociated method based on the key moving objects of trajectory to associate thosetrajectory fragments that belong to the same target. And then to form a long completetrajectory by filling algorithm, in order to improve the reliability of the trajectoryrepresentation and accuracy of the retrieval.Because it cannot effectively measure the similarity of the raw data of thetrajectories, so the main task in trajectory based video retrieval are trajectory representation and its similarity measurement. The trajectory representation is toparameterize every trajectory and uses the parameters to index them. To solve theproblem, this paper uses the Gabor filter to carry out the multi-channel spectralanalysis on each trajectory, and then get the key points of each trajectory. Accordingthe moving direction changes of the target at the key points, the key points will beencoded with a specific character. At the same time, the frequency of each key pointgot through the multi-channel spectrum analysis will be preserved. Then the trajectorycan be represented by a string and the frequency information. So the similaritybetween trajectories can be calculated by computing the minimum edit distance inwhich the frequency information is used as the operation costs between the strings.This method cannot only be used for global trajectory matching, it can also be appliedin the sub-trajectory matching.