Human Activity Recognition Research Based On Hierarchical Model

Vision-based human action recognition, an application of high-level computer vision, is an important research area of human motion analysis. And it has been widely used in intelligent surveillance, home security systems, intelligent robots, athlete support training system and so on. To understand human action and activity by computers will become the trends of high-level computer vision. However, human activity is very complex and lack of a clear classification model. Moreover, the same human motion may have different meanings in different contexts. Till now, there is no universal model to describe and recognize human activity and most of the researchers use their specific models for the particular environment. Recently, hierarchical model has become a popular human behavior representation, only the levels of the model and their definitions are different in various environment or tasks.In this paper we first discuss the hierarchical models used in recent years, then proposed a hierarchical model for human action recognition which based on practical research purpose and the characteristic of human reasoning and recognition. In this model, human action is divided into three levels: human posture recognition, simple action recognition and activity analysis for video surveillance systems, which are from bottom to up.An algorithm based on the Fourier descriptors and codebook for human posture recognition is proposed in this paper. The human posture model is establishing by using human silhouette parameters. To classify the types of human posture, a hierarchical recognition method and a nearest neighbor matching classifier are adopted. Experiments show that this algorithm is efficient, less complex and needs less storage. The method is also robust. Even for some interferential images, the algorithm also could achieve fine results.In the human action recognition, a key-frame algorithm is proposed for the key posture extraction in the video sequences. The key posture is then matching with codebook to recognition human action, and a window filter function is used to smooth the result. Experiments show that this algorithm could efficiently extract the key-frame and recognize the human action from video sequences.Finally, the human activity analysis technology is applied to an outdoor video surveillance system. In this system, we use the trajectory of feature points and user-defined rules to recognition several simple abnormal behavior. The action analysis module and its results are also proposed.