Studies On Posture Estimation And Tracking Of Human Body Based Sparse Point Clouds

With the development of computer vision technology, the target tracking technology forthree-dimensional information, which is based on image sequences, has been widely appliedin many fields. In a complex image background, people can accurately track and identifytargets through their eyes, while the current machine vision technology, which gets interestinginformation from the environment through image sequence and then tracks the target,becomes a hotspot of research.Body posture estimation and tracking method are usually based on the two dimensionalmethod and three dimensional method. This paper is based three-dimensional method,because the method based on the two dimensional needs a great deal of prior knowledge. Thecurrent researches on stereo vision have difficulty in a large amount of calculation andmatching for stereo matching algorithm, and the existence of high complexity tracking andthe lack of three-dimensional information for target tracking. Thus, the estimation andtracking based on sparse point clouds of body posture is proposed to get better results.The method based on sparse point clouds makes stereo matching, which reduces theredundancy calculation of the matching process. The sparse point clouds, which makes use ofbinocular stereo vision, means to extract the sparse feature points, and calculates the threedimensional coordinate information. This paper expounds the stereo matching algorithm,proposes two matching algorithms that one is dependent on limit constraint and feature pointsand the other is based on feature points and region matching. By the experiments oncomparing the two kinds of matching algorithms, the former can match more feature pointsfor a simple target, the latter for complex human target matching effect is better. In addition,the parallax is given based on these methods, and eventually to get three-dimensionalinformation with small errors.The method is based on three-dimensional method to estimate and track body posture,which needs to the human body3D model given firstly and the3D information which belongs to the human body target portion feature points from the2D images to describe the target;then the unscented kalman particle filter algorithm is combined to establish the observationfunction, and the method based on model predictions is used to search out a set of humanbody models which has the best similarity in model configuration space to realize the trackingprocess for the human body skeleton. Finally, the space position of human body center can bequickly located according to the three-dimensional sparse characteristic point clouds for thebody skeleton model. The experiment shows that the unscented kalman filter algorithm hashigher precision than other particle tracking algorithm, and has a good tracking results incomplex background, then has a better posture tracking based on three-dimensional pointclouds in simple background.