Research On The Recognition And Understanding Of The Interaction Behavior Based On Graph Model

Visual-based human interaction recognition is one of the most important research directions in the field of computer vision and human motion analysis. Compared to the single person action recognition field, the research on interaction recognition remains limited.However, the human interaction recognition is more urgent for the intelligent video surveillance in some special occasions. In the research framework of the human interaction recognition, the interaction method recognized as a general action is usually unable to express the intrinsic properties of the interaction, and the accuracy is not high. The method based on individual segmentation has a high complexity of modeling with poor robustness to body occlusion. Those issues greatly restrict the applications of the human interaction recognition algorithms.In this paper, the issues of the existing human interaction recognition algorithm are as follow,First, as a whole study on the research achievements of human interaction recognition,the existing human interaction recognition algorithms are classified into two categories from the viewpoint of the recognition framework: the interaction method recognized as a general action and the method based on individual segmentation. Through in-depth analysis, the advantages and disadvantages of these two kinds of recognition framework are listed, and the characteristics of the two kinds of local space-time characteristics and global characteristics are determined.Secondly, a large number of tests have been carried out on the local features(Spatio-Temporal Interest Points,STIPs) and the global features(Histogram of Oriented Gradient,HOG) Experiments results show that the extraction of HOG features is simple and HOG features have a good ability to describe the global characteristics of the region of interest with better recognition accuracy.Thirdly, according to the problem of the recognition framework, a novel recognition framework based on the probability fusion of hierarchical structure is proposed. Theframework can effectively describe the inner properties of the interaction, meanwhile avoiding external factor disturbance through combining above two recognition frameworks.The structure of the proposed framework is simple and can meet the real-time requirement.Finally, in order to improve recognition accuracy, a method of interaction recognition based on the hierarchical HMMs probability fusion is presented by combining the HMMs with better spatio-temporal modeling capability and the proposed hierarchical structure. This method makes full use of the time context of interaction and the global motion characteristics,which obtains the better performance of interactive recognition.