Motion Analysis Of Small Groups In Intelligent Video Surveillance

With the rapid population growth, public security consciousness has beengreatly improved. The video analysis on crowd will be received more and moreattention. Currently, researchers in the public surveillance field have found that mostcrowds consist of small groups rather than isolated individuals. As a usual kind ofcrowd activity mode, the research on the motion analysis of small groups is underincreasing attention.Combined with the latest computer vision research theory, we identify thecrowd groups and detect abnormal behavior using motion trajectories in detectedcrowd regions. The highlights of the dissertation include four aspects:(1) The aspect of motion region detection. The algorithm is mainly consisted oftwo parts of detection techniques. One is about the difference between the currentframe and background frame with the judgment formula in HSV color space. Theother is motion edge frame differential method to extract reliable regions with themethod of multi-directional scan edge regions. This algorithm is not only suitable forsparse population, at the same time in complex small group scene, also canaccurately detect the crowd movement area.(2) The trajectories extraction in crowd region. This paper firstly analyzes thequality of general target tracking algorithm. Then it introduces a kind of longdistance motion estimation algorithm, which applied to estimate motion trend andobtain the crowd trajectories. This algorithm realized that crowd groups can betracked in public.(3) In the aspect of identifying groups, trajectories spectral cluster algorithm isproposed to cluster small groups. According to particle motion trajectories and thesocial cognitive rules, similarity matrix will be constructed between particletrajectories. Then we use the similarity matrix to distinguish different moving groupsby means of spectral clustering method in the scene.(4) The detection of abnormal behavior about crowd groups. In this aspect, thispaper puts forward a kind of feature extraction and description scheme, which isapplied particle information about crowd motion to space-time features cubes. And it has been proved it would be useful in global monitoring system. The detectionalgorithm combined with space-time feature cubes and neural network model isproposed to detect abnormal events. The experiment is shown that our method canachieve ideal results in detection abnormal group events, such as fighting, falltrample, panic scattered.