Research And Implementation Of Model-Based3D Pose Tracking

In the natural world, almost all objects have three dimensions, thereby how to recover their motion parameters which can be used in motion analysis from a2D planer image quickly and accurately is becoming more and more attractive. Model-based3D pose estimation methods have a low requirement of equipments, and an appropriate model as well as several initial parameters is enough to start3D pose estimation using a monocular camera. Hence, technology of3D tracking with a single camera has a perspective outlook in face recognition, expression recognition, gesture understanding, virtual reality, game and entertainment. This paper endeavors to study correspondence theories and implement relevant algorithms.Firstly, we use stereo-vision algorithm to reconstruct the real3D model for the target tracking object. The whole reconstruction procedure include camera calibration, stereo rectification, image filtering, stereo matching and a forward-backward scanning technology which can be used to detect stereo matching errors. When obtain the parallax image after stereo matching, we use a pre-projection method to get the whole three-dimensional coordinates of target object’s appearance and map them to obtain the final three-dimensional model;Secondly, we study the correspondence algorithms of model/image registration with a monocular camera. We present a RANSAC-based two-step method to estimation pose parameters between two concussive frames in image sequences or real-time video. This method also can be called weighted brightness and depth change constraints function. When we estimate current pose, we should first use a RANSAC-based sampling of2D feature points, then estimate parameters using a BCCE function until we already obtain a comparable accurate pose results, lastly use the weighted brightness and depth change constraints to optimal pose result. Lastly, we automatic update2D feature points on basis of the correspondence confidence of the feature points.Finally, when3D tracker undergoes long-time tracking, the pose estimation may be difficult to recovery due to drift accumulation, object scale change, and face expression change of occlusion. In order to solve these issues, we addressed Cube-based Multi-layer view appearance model and relevant indexing technology. Our novel model successfully solved above problems. We employed Gaussian random variable to measure frame error estimation, and applyed a kalman filtering method to make an overall optimization.