Robust Object Localization from Segmented Range Images by Matching Intrinsic Line Features

Locating 3-D rigid objects reliably from segmented range images in the presence of occlusion, noise, or discrepancies between image and model is important for many applications. While surface features work perfectly for matching, point and line features are still needed for locating. Fast, non-iterative locating methods usually require accurate point correspondences, and do not tolerate significant occlusion. We present a new method to decide which features are object-intrinsic for a given matching hypothesis. To get precise point correspondences, we apply an extension of Arkin's shape similarity metric to multiple intrinsic 3-D contour sections. The metric is robust with respect to segmentation errors. As features, we use the surface patch boundaries, but the method could be applied to other line features. We demonstrate some experience from a gantry robot test site equipped with time-of-flight laser scanners.