3D modeling is a hot topic in both computer vision and computer graphics communities thanks largely to its high academic value and wide applicability. The thesis is focused on image registration and integration, a key part of the structured-light based 3D modeling system in our group. The main works are as follows: 1. A simple approach is proposed for the reconstruction of triangular mesh from the dataset acquired from our structured-light system. The approach is shown rather practical and effective. 2. A new image-based method is proposed for 3D registration. It is mainly characterized by using both intensity similarity and point-to-point distances as the matching features, by which the accurate starting position, usually a requirement in other imaged-based methods, is relaxed. Its main steps are: firstly, on the first projected image, distinctive corner pixels, which act as salient features for subsequent image matching, are detected via the minimal eigenvalue of the auto-correlation matrix; then, on the second projected image, the correspondent pixels are found by cross-correlation; a verification scheme is used to discard the potential mismatches by thresholding the correlation coefficients and point-to-point distances. The verification process is divided into a coarse to fine stages, and in each one of the two stages, the two corresponding thresholds for both correlation coefficients and point-to-point distances are determined differently. Experimental results show the superiority of our proposed verification scheme to the previous one using only correlation coefficients in terms of registration accuracy and robustness. 3. We have implemented the Consensus Surface algorithm, a robust volume based method for 3D integration in the literature. It has been integrated into our structured-light based 3D modeling system.
修改评论