Research On Human Motion Tracking And 3D Reconstruction

The detection, tracking and analysis of motion objects are important fields of computer vision. They are widely applied in security monitoring, video compression, video edit, sport training and many other fields. This article conducted a research in depth on two important points of human motion analysis which are motion tracking and 3D reconstruction. After a detailed introduction of existing algorithms, we depicted new algorithms that we proposed on the following aspect:(1) Motion tracking is a key point of human motion analysis. Further processing can not be conducted until accurate trajectories of human joints were extracted. Most of the traditional image-matching or statistical theory based methods have the drawback of poor stability and slow speed. What's more, they are less robust to self-collision phenomenen. Therefore, after an analysis of the human vision system, based on the shape analysis technique, we proposed a new algorithm for human motion tracking. Our algorithm does not need special markers, avoids heavy searching burden and consequently improved the flexibility and speed. The algorithm is very robust because it overcomes the affection of object rotation, translation and illumination variation.(2) After the 2D trajectory of motion object was extracted, we need to reconstruct the 3D configuration of the object. But the 2D to 3D projection is a one-to-many transform because of the information loss of its inverse procedure. If multiple cameras are used, the problem can be solved by adding additional equations except that it needs accurate calibration among these cameras. In this article, based on Biomedical constrains and parametrical human model, we proposed a novel algorithm which is more efficient and need not calibration of camera to reconstruct 3D gestures from a single image. The algorithm is feasible to a vast range of image types and the gestures it generated are always reasonable.