Computer vision techniques for complete three-dimensional model reconstruction

This dissertation addresses the problem of automatic 3D model reconstruction of real objects. It has a number of applications in both computer vision and computer graphics areas such as industrial inspection, reverse engineering, Internet Web content and E-commerce. Current 3D modeling techniques require significant manual intervention and use expensive hardware systems for data collection. In this research, we present a complete and low-cost digital vision system to create photo-realistic 3D models. In our approach, the reconstruction of 3D models involves four major steps: (1) data acquisition, (2) registration, (3) surface integration, and (4) texture mapping. Partial 3D shapes and texture information are acquired from multiple viewpoints using rotational stereo and shape from focus (SFF). Rotational stereo model is first introduced in this work to acquire the depth information. Rotational stereo provides the flexibility of controlling the effective stereo baseline and it can incorporate existing stereo matching techniques. It provides a fast and accurate approach for 3D shape recovery. The resulting range images are registered to a common coordinate frame according to the acquisition viewpoints. The accuracy of the initial registration is improved by finding the rotation and translation matrices iteratively by a least-squares error minimization approach. The registered range data are then integrated into a surface model using three different approaches. A new algorithm named Region-of-Construction is developed for fast surface integration. It directly exploits the structure of the raw range images and determines regions corresponding to non-redundant surfaces which can be stitched along the boundaries to construct the complete 3D surface model. The algorithm is computationally efficient and has low sensitivity to noise and registration error. A final photo-realistic 3D model is obtained by mapping the texture information recovered by SFF onto the complete surface model representing 3D shape. A digital vision system has been built and the algorithms have been implemented on the system. The results of 3D model acquisition for several real objects are presented.