Researches On Human Motion Recognition Based On Video

Human motion analysis based on video is currently one of the most active research fields, and human action recognition is a challenging research topic in this field. Human motion recognition has many promising applications such as man-machine interface, intelligent surveillance, athletic training, and content retrieval. This thesis analyses and summarizes some related research work on human motion recognition in computer field and built suitable models to solve problems which is complicated single and continuous human action recognition according to the data with motion capture in video. The main contributions are as follows:Firstly, this thesis proposes a method to describe human action eigen-sequence based on conditional random field in order to solve the problems of the limitation to observation and label-bias exist in traditional probability models. Without modeling the observation, this method can avoid unrealistic independent assumption on the observations given the action class labels, and it has settled the label-bias problem with global normalize in the optimizing process, therefore, this method is more suitable to human action modeling in reality. Besides, conditional random field accommodates overlapping features or long term contextual dependencies in the observation sequence, which make this method has advantage in recognizing complicated human action.Secondly, this thesis proposes a method combined with Conditional Random Field and Condensation (Conditional Density Propagation) to segment and recognize human continuous action, which decompose the continuous action recognition into various single action recognition. Firstly, the Condensation is used to segment human continuous action sequence into single action fragment with the strategy of hypothesis-verification by generating a hypothesis about the border of a single action in continuous action sequence; secondly, Conditional Random Field implements recognition of the hypothetical single action fragment; then, the likelihood of a hypothesis can be evaluated by the probabilistic output of Conditional Random Field. Having removed some transitional frames in the action switching, it reduces the influence on transitional poses and improves the robust and accuracy of recognition.The experiment result shows that the methods of complicated human action recognition and continuous action recognition proposed in the paper are useful and have established a solid foundation for the future research.