FEATURE-BASED REGISTRATION OF RANGE IMAGES IN DOMESTIC ENVIRONMENTS

Abstract

Registration is an important step to combine pictures that are taken either from different perspectives (multi-viewpoint), at different points in time (multi-temporal) or even from diverse sensors (multi-modal). The result is a single image that contains the combined information of all original images. We use 2 1 2 -D range data taken by a laser range scanner. During recording occlusions can leave some of the objects scanned incompletely. Thus it is the task of the registration to produce images which give the most complete view of the scene possible. An algorithm often used for the registration of range images is the ICP algorithm (Rusinkiewicz and Levoy, 2001). This point-based algorithm only uses local information, and it has the advantage that it is universally applicable. The drawbacks are that its very time consuming, it may converge in a local minimum, and the result cannot be evaluated absolutely. Therefore, for our domestic scenario we present an approach that first detects certain global features of the room and then uses their semantic and spatial information to register the range images. Thus this approach has the advantage that its tendency to get stuck in a local minimum is reduced and the overall performance is increased.