Object recognition and localisation from 3D point clouds by maximum likelihood estimation

We present an algorithm based on maximum likelihood analysis for the automated recognition of objects, and estimation of their pose, from 3D point clouds. Surfaces segmented from depth images are used as the features, unlike ‘interest point’ based algorithms which normally discard such data. Compared to the 6D Hough transform it has negligible memory requirements, and is computationally efficient compared to iterative closest point (ICP) algorithms. The same method is applicable to both the initial recognition/pose estimation problem as well as subsequent pose refinement through appropriate choice of the dispersion of the probability density functions. This single unified approach therefore avoids the usual requirement for different algorithms for these two tasks. In addition to the theoretical description, a simple 2 degree of freedom (DOF) example is given, followed by a full 6 DOF analysis of 3D point cloud data from a cluttered scene acquired by a projected fringe-based scanner, which demonstrated an rms alignment error as low as 0:3 mm.