Human Behavior Classification And Anomaly Detection Based On Video Sequences

In recent years, with growing awareness of social security, video surveillance system hasbeen more and more attention. However, the degree of intelligent video surveillance is nothigh, automatically human behavior classification and abnormal behavior detectiontechnology is still constantly being explored. On the basis of existing research results, weanalyze the moving target segmentation, behavioral characterization, feature extraction,categorization of actions, and abnormal behavior detection analysis.In the aspect of moving target detection, summarized the current commonly methods, analgorithm of combination of background subtraction and edge detection is proposed. Whenthe slow movement of human and only certain parts of the body moving occurs, conventionalalgorithm can’t extract the complete body contour. Algorithm proposed in this paper cansolve the problem.Based on gait energy image (GEI), variance of GEI (VGEI) and image segmentationalgorithm are proposed. VGEI gets higher recognition rate than GEI. Image segmentation candeal with the problems of blocked feet or shadows effectively, which has a better robustness.The feature of "distance of body contour to the centerline" is put forward, which obtainshigher recognition rate than “distance of the contour to the center point”.In the aspect of linear manifold algorithms, both2D2MSDPCA and2D2MSD areproposed。They not only overcome the small sample size problem of LDA, but also improvethe capacity of behavior classification and enhance robustness, the highest recognition rate is100%.In the detection of abnormal behavior, considering the different distance of the movingtarget to the camera, different location in the screen, inclined image and other issues,combination of improved Hu moments and improved Hausdorff distance method areproposed for motion analysis. Here, region-based discrete Hu moments are replaced by thecontour-based discrete Hu moments, which meet the three invariance of the moment andmake data of the algorithm simple. Improved Hausdorff distance is that the maximum as thedistance criterion is modified by the average as the distance criterion, and the largest varianceof the data is removed, the method effectively smoothes the noise, the recognition rate issignificantly increased. But Hu moments extract global variable, global noise often can drownsubtle differences between similar images, resulting in recognition errors. Wavelet moment isintroduced into the detection of abnormal behavior. We put forward the concept of waveletcontour moments. Computation of wavelet contour moments is fast; parameters-m, n, qvalues are analyzed in detail, and n parameter computation is improved. Using wavelet’s time-frequency characteristics to detect the abnormal behavior with improved Hausdorffdistance, abnormal behavior recognition rate reaches97.92%.