Research Of Pose Tracking Based On The3D Model

The3D pose tracking as a branch of visual problem, has been a more and more important research direction in computer vision and human computer interaction field recent years, its main purpose includes getting3D pose parameters of the target and realizing the target motion analysis and understanding through inputing image sequences containing the target. The pose tracking method based on the3D model has low requirements on the equipment, usually only needs to construct a reasonable3D model and provides the initial depth of the3D model, therefore it has wide prospect in the field of intelligent monitoring, video conferencing, face recognition, virtual reality and other practical application. This paper studies and implements the pose tracking method based on the3D model, mainly including the true3D model construction, posture estimation and tracking, the cumulative error elimination r.The construction of the real3D target model method adopts the binocular stereo vision technology, and the paper do a research on camera calibration, stereo image correction, image preprocessing, stereo matching method of the stereo vision system. We use Zhang Zhengyou binocular camera calibration method, select a high matching precision adaptive weighting correlation window region matching algorithm for disparity computation, and propose the bidirectional consistency detection method to eliminate false matching and obtain more accurate depth information, and get the real3D mode through the reprojection.The research on the pose estimation and tracking algorithms is based on monocular image sequence, this paper presents a two-step iterative method based on the iteration variable weight gray depth constraint random consistency for inter pose parameters estimation and tracking. This part of research contents includes optical flow method based on gray level constraint equations, depth constraint equations and iterative weighted least squares linear problems. For head pose tracking, the initialization phase using BCD method to determine the initial frame, using the improved KLT algorithm for feature point detection and tracking, and initial the current frame pose values using random sampling method.In cumulative error elimination phase, this paper uses multiple frame registration method to resolve, puts forward a spatial hierarchical multi-scale view appearance model, and uses hierarchical indexing model to improve frame registration efficiency. This registration method is effective in reducing the error accumulation of continuous pose tracking process, and solves the problem of inter frame motion amplitude caused by the excessive registration failure.Finally, the paper uses the pose tracking method based on the3D model under the monocular image sequence to realize head pose tracking experiments, the results show that this method is reliable and accurate, and the future works are discussed.