Algorithms for capturing human body motion and structure.

摘要

The tasks of a typical computer vision application include: recovering scene structure, detecting/recognizing object/human, recognizing activities, responding to events, etc. These tasks involve critical computer vision problems including establishing correspondence, stereo, structure from stereo, optical flow computation, structure from motion, object/human tracking, event detection/understanding, etc. The research work in this dissertation covers four of these problems.; For solving correspondent problem, we propose a fast hierarchical stereo matching approach using discrete wavelet transform. We show that the compactly supported wavelet basis with n vanishing moments can be regarded as the nth derivative of the signal after some scaled smoothing operation. The area- and feature-based methods are then combined using the multi-resolution framework of wavelet. For computing optical flow, we propose three methods. The first is the extension of the stereo matching algorithm. The second is also based on wavelet multi-scale analysis but under the brightness constraint. A novel coarse-and-find method is designed to alleviate aliasing and error propagation problem. Both of these methods are computationally efficient. The third method achieves greater accuracy. Here, the optical flow problem is formulated as a total least square estimation based on 3D structure tensor analysis. A parametric model and an adaptive neighborhood adjustment are integrated to improve the accuracy and dynamically handle the aperture problem.; For the structure from motion (SfM) problem, we solve the nonlinear SfM system by decomposing it into two linear subsystems, a motion subsystem and a structure subsystem, in a multi-resolution framework. We prove the convergence of our algorithm. The statistical analysis shows that error variances are decreasing along the coarse-to-fine iterations when the optical flow estimation is reasonably good and the structure of the object is relatively flat.; For the tracking problem, the optical flow is implicitly used as a bridge to fill the gap between the brightness constraint and the kinematic chain. The articulated body tracking problem is then formulated as a linear system. A re-initialization scheme is proposed to facilitate system linearization. A new statistical constraint is derived to achieve more accurate tracking.