The Research On Human Motion Capture Algorithm In Video

Human motion capture is the basis of behavior recognition and analysis. It is still an open problem in computer vision and has broad application prospects in the security surveillance, human computer interaction, and computer animation and so on. This paper do research on human motion capture through the video analysis, including moving target detection and tracking. In detection, a background subtraction method based on gaussian mixture model with adaptive number of components and adaptive learning rate of parameters is proposed. In tracking, a tracker which integrates detection results and tracking results using particle filter is proposed.Firstly, the background modeling method based on gaussian mixture model is improved in two aspects. For a fixed number of gaussian components resulting in waste of system resources and storage space, this paper select the appropriate number of components for pixel observations using Dirichlet distribution and likelihood probability. For the invariant learning rate used to update the model parameters in background changing pharase leads to slow convergence, this paper update the parameters based on the statistical characteristics of the pixel’s observed values. After obtaining the target or targets, this paper combined the geometry characteristics of human and human skeleton model to identify human target. For tracking, this paper take the detection results to initialize the tracker and use the weighted sum of posterior probability of detection and the probability of particle propagation to approximate the importance density probability, and then compute the importance weights of particles according to the importance density to increase the accuracy of particle. In the tracking process, if new human is detected, the information can be used to establish a new observation model so that multi-target can be tracked. Experimental results show that the proposed method based on gaussian mixture model can get more accuracy detection results, and the tracker based on particle filter can track multi-target.