6DoF : Registration of Range Data [lecture 1]

Abstract

Registration deals with the problem of aligning sensor readings taken at two different poses, i.e., in the case of 3D range data to determine the 6 degree‐of‐freedom (DoF) spatial transformation that properlyco‐aligns corresponding regions of the data. 6‐DoF registration is hence a core element in 3D perception and world‐modeling, for example for the matching of templates from an object database to a scene or for the integration of multiple scans taken at different poses into a global 3D map. The lecture presents well‐known state of the art methods like the Iterative Closest Point (ICP) algorithm and its variants, and visual methods using interest points, e.g., the Harris corner detector, plus descriptors, e.g., the Scale Invariant Feature Transform (SIFT), in combination with the Random Sampling Consensus (RANSAC) algorithm. Furthermore, novel 6‐DoF registration methods are presented that unlike classical methods like ICP do not use local features. Instead they are based on the postulate that it is beneficial to employ large dominant structures in the data for registration. The first method is Spectral Registration with Multilayer Resampling (SRMR). It is based on decoupling 3D rotation from 3D translation by a corresponding resampling process of the spectral magnitude of a 3D Fast Fourier Transform (FFT) calculation on discretized 3D range data. The registration of all 6‐DoF is then subsequently carried out with spectral registrations in 1D, 2D and 3D using Phase Only Matched Filtering (POMF). The second method is Plane‐based Registration, which features an efficient algorithm to simultaneously determine optimal correspondences and rigid transforms between two sets of planes fitted into range data. It is shown for both methods that they work under very partial overlap between the scans, with very noisy sensor data, and even under the presence of large occlusions. It is also shown for the Plane‐based. Registration that it not only works in indoor environments but also in unstructured scenarios including even hilly landscapes