Human motion tracking and recognition from a image sequence is an important research field in computer vision and image processing. This paper mainly study this problem and can be concluded as following: (1)From simple model to complex model, two human body motion problems are studied: a) face pose tracking under affine model; b) structural human model tracking. In a), we carry out human face pose estimation work from image sequence under affine model. Under pyramid structure, we use Levenberg-Marquard method to minimize gray scale correlation energy to solve the model parameters. We introduce robust estimation and parameter prediction and smoothing techniques to increase precision and continuity. Human face tracking experiments and application in video rewrite show this method is effective. In b), we propose a rectangle connection model for human structure. In order to find out the model parameters in the sequence, we use two techniques. The first is prediction the initial model value through α-β-γ filter. The second is to find the best match through genetic algorithms. Experiments on real sequence show that the algorithm is robust in some degree. (2)We study eigen-curve representation of motion trajectory and action recognition based on this technique. Through skin color detection in HSI space, we apply motion prediction to track human face and hands and compute the motion trajectories. Then we study how to match trajectories under deformation in spatial-temporal space and conduct the experiment of Taiji posture recognition. Not need large number of training data is the main advantage of this method. 3)We propose active shape model based conditional density propagation algorithm and the idea of virtual mouth. We first construct primary component analysis of hands contour. Then new hands can be represented approximately in term of PCA bases. This coefficients can be viewed as deformation parameters. The state can be constructed from these parameters together with the rigid motion parameters(position, scale and rotation). Then using conditional density propagation algorithm, we sample, predict and propagate the state in time axis. And to compute the expectation state. Initialization with this state, we use active shape model to refine the location and deformation of the target. Finally, we simulate mouse action by gesture. The mouth action is judged by the sate transition, and the mouth mouse position is computed from the centroid of the contour. The tracking and recognition can be realized in near real time. (4)We propose a integrative method for tracking and recognition based on dynamic motion trajectory matching. This method introduces feedback mechanics between tracking system and recognition system, and can solve the tracking and recognition in constraint system with better performance. We conduct this study in example of digit writing image. First, through training samples, we develop the moti
修改评论