Human Action Recognition Based On Conditional Random Field

The human action recognition is an important topic in computer vision. With the development of the modern society and the spread of the internationalization of our life, the traditional video monitoring system, whose aim is only capturing video, will be replaced by the intelligent recognition system to reduce manpower, material resources and time for the captured video analysis by the manual way. The recognition system, for human's movements and gestures analysis, has a wide range of requirements in all the aspects of our life, such as public security, smart home and elder monitoring. Human's movements have the complex nonrigid, and the apperance is diverse with the different people, even the same people. Therefore, the robustness of human action recognition algorithm is very important. Compared with static object images such as faces and cars, human motion sequences have the characteristics of long-time correlation. To describe the motion sequence as the data point of high dimensional space will cause the loss of the temporal information during the movements. Under this context, the human action recognition method based on the graph model, which can build weight relations among human action sequence frames, and capture the relation of the previous and subsequent changes among human action sequence frames, has been paid more and more attention.In this thesis, Conditional Random Field(CRF) and Hidden Conditional Random Field(HCRF) are adopted to build the human action recognition method. To represent human action sequences, two problems, i.e. artificial constraint and excessive freedom, have been handled in the model-building process. The main contributions are summarized as followed.1. The graph model, which has high order neighborhood relationships, was built to make up the non-neighborhood information, which is lost in the chained CRF through analyzing the relations and characteristics among human action sequence frames. The L1-group regular term was introduced to make the graph model sparse, and then the intrinsic structure of the graph model was extracted for the human movement representation.2. For the dispersion properties of human action sequence frames, the Stage-HCRF was proposed based on HCRF. By analyzing the information of the previous layer hidden nodes, Stage-HCRF represents the movement information as the state of the next layer of this model. Moreover, the L1-group regular term was introduced to make the two hidden layers sparse, and the excessive freedom was eliminated when the model was built. With this model, the better intrinsic model structure was found for describing the human action.