This dissertation presents an automatic registration system for aligning combined rangeintensity scans. The approach is designed to handle several challenges including extensive structural changes, large viewpoint differences, repetitive structure, illumination differences, and flat regions.;First we present a technique for pairwise registration split into three stages: initialization, refinement, and verification. During initialization, intensity keypoints are backprojected into the scans and matched to form candidate transformations, each based on a single match. We explore methods of improving this image-based matching using the range data. For refinement, we extend the Dual-Bootstrap ICP algorithm for alignment of range data and introduce novel geometric constraints formed by backprojected imagebased edgel features. The verification stage determines if a refined transformation is correct. We treat verification as a classification problem based on accuracy, stability, and a novel boundary alignment measure. Experiments with 14 scan pairs show both the overall effectiveness of the pairwise algorithm and the importance of its component techniques.;We also develop a method for directly combining images with range scans for keypoint detection, description, and matching. We extend a 2-D image-based detection and description framework to 3-D using an image backprojected onto a range scan. A key feature of the framework is a physical scale space for detecting keypoints, which eliminates errors in scale during both detection and matching. We develop smoothing, differentiation, and description techniques that are focused on making the keypoint invariant to viewpoint, sampling, and intensity changes. We integrate physical scale keypoints into our pairwise registration algorithm, in turn developing a physical scale keypoint based registration verification measure. We present a new technique for improving physical scale keypoint match ranking by combining an image and a range metric. We demonstrate the power of our algorithm with comparisons to variants of the SIFT algorithm and show that it is able to find and match keypoints in a variety of challenging scan pairs.;Finally, we present an algorithm for the registration of an unordered set of combined range-intensity scans. We use a Bayesian network to combine pairwise verification criteria with a many-scan cycle consistency measure for determining the true overlap in the set range-intensity scans. We finish by refining all pairwise registrations at once to produce a single transformation for each range scan that brings all scans into alignment under a single coordinate system. We demonstrate this algorithm on three different datasets.
展开▼
