Local feature extraction and matching on range images: 2.5D SIFT

This paper presents an algorithm that extracts robust feature descriptors from 2.5D range images, in order to provide accurate point-based correspondences between compared range surfaces. The algorithm is inspired by the two-dimensional (2D) Scale Invariant Feature Transform (SIFT) in which descriptors comprising the local distribution function of the image gradient orientations, are extracted at each sampling keypoint location over a local measurement aperture. We adapt this concept into the 2.5D domain by concatenating the histogram of the range surface topology types, derived using the bounded [−1, 1] shape index, and the histogram of the range gradient orientations to form a feature descriptor. These histograms are sampled within a measurement window centred over each mathematically derived keypoint location. Furthermore, the local slant and tilt at each keypoint location are estimated by extracting range surface normals, allowing the three-dimensional (3D) pose of each keypoint to be recovered and used to adapt the descriptor sampling window to provide a more reliable match under out-of-plane viewpoint rotation.