| Recognition of human behavior is an important research aspect in artificial intelligence. It can be widely applied in many areas such as smart surveillance, human-computer interface, analysis of movement, virtual reality, content based retrieval, medical diagnosis, video conference, etc, thus creating immense economic value. Especially, with the increasing public security requirement, the smart surveillance system based on computer vision has been broadly used in many areas of national life.Considering the requirement of intelligent surveillance in retailing, this paper has designed and realized two algorithms to recognize human interactive actions in video sequences captured from a single fixed camera. Both methods focus on the paying action of the consumer. One of them is based on image segmentation, and the other is based on motion vectors. Finally, a comparison is made between the two algorithms with the respect to the accuracy and the processing speed.The one based on image segmentation combines motion information between adjacent frames with color distribution of each frame. It makes use of frame difference, edge detection, morphologic processing and projection method to obtain the motion mask. At the same time, it uses color histogram and clustering method to segment frames. Based on the segmentation result, the motion mask is utilized to choose effective color blocks and to get the moment feature for further classification. The one based on motion vectors takes into account the space-time motion information included in consecutive frames. Then using motion vectors, which are computed by block-matching algorithm, it calculates the local histograms of motion and the average velocity corresponding to the each bin of the histogram for each chosen space-time sub-patch. At last, the histograms and velocities are used as the HOM feature to make the classification. Both the algorithms use the support vector machine as the classifier. |
PDF Full Text Request