Research On Recognition Of Human Behavior Based On Dynamic Image Sequences

Target behavior analysis based on visual analysis includes target detection, track, descriptionand recognition. The target detection and track are the basis of target analysis. It is on the baselayer of behavior understanding system. The target description and recognition is on the highlayer. The target behavior understanding research has important implications and is used widely.Fives behavior types are researched including Walk, Crouch, Run, Fall and Sit. Target detection,track and behavior understanding are studied in this paper. The results show that the system ofthis paper can recognize these five types. Especially the accuracy rate of abnormal behaviordetection has a good outcome.(1)To avoid the shortcoming of the traditional background subtraction approach and the framedifference, an improved method is put forward combining with the two mentioned. First, the twoforeground images are segmented, and then the two results do the multiplication operation. Thenew algorithm dynamically looks for the threshold, which makes the last foregrounds has anautofit segmentation result and enhances the accuracy of the target detection.(2)On the target track step, this paper takes the DESO and GVF-Snake to track the target. Theinitialized curve of the outline in the track algorithm takes shape from object’s minimumbounding rectangle tracked from the DESO result. It avoids to initialize the curve manually andthe real-time is improved，which realizes the fast track to the target outline.And then the resultfrom GVF-Snake update the DESO. It is a integrated tracking and the robustness and real-timehas a development.(3)During the target description, this paper takes the PCA combining with the shape informationfrom the silhouette outline to describe the target. The description to the shape information mainlytakes the posture ratio and the outline centroid position information. The target motion state isobserved and analyzed. The uniformitarian image data base is created. The PCA is used toextract features of different target behavior to get a further analysis.(4)On the target classification step, the OAA-SVM is set up. The voting mechanism is employed.Different classifiers are put to use according to different characteristic. The classifyresults get a vote. The top vote-getter is the last behavior. This classification system improves theclassification effectiveness and the classification performance. The classification rate isimproved certainly.