Intrinsic line features and contour metric for locating 3-D objects in sparse, segmented range images

This paper presents a new, direct method for locating 3-D objects reliably from sparse, noisy data: segmented range images with polygonal patch boundaries. Non-iterative locating methods require accurate point correspondences, and do not tolerate significant occlusion. Therefore, we propose rules to select intrinsic contours relevant to the locating task in a multiple-object configuration. For matching intrinsic boundaries and finding correspondences, we develop a 3-D version of Arkin's contour metric using the signatures of turning and torsion angles, and extend it to contours consisting of multiple parts. The metric is robust to quantization and segmentation errors. We integrate the method into a complete system for 3-D object recognition and report on experience gained from a gantry robot test site equipped with time-of-flight laser scanners.