This paper presents an end-to-end 3D registration algorithm for relative navigation between known objects based on a point-to-CAD iterative closest point (ICP) principle. The objective of this method is to take in a measured point cloud extracted from a depth or disparity map – such as the ones obtained from stereo cameras, time-of-flight cameras, LiDARs, or depth from defocus sensors - and calculate the rigid body transformation that best aligns the measured data with a corresponding 3D CAD model. By leveraging the geometric information encoded into stereolithography (STL) files, it is sought to address the computational intractability imposed by the naïve generation of dense target point clouds solely based on the target's known surface. To this end, the proposed approach computes a bijective projection onto the known triangular mesh to obtain a target point cloud with which to use ICP techniques for incremental alignment; the projection step is then carried on recursively until the convergence criteria are met, yielding a relative 6DOF pose between the two objects to be used within the estimation pipeline. Demonstrations of the algorithm are presented using simulated datasets; results include time complexity analyses for real-time operation cases, performance variation assessments with respect to CAD model complexity, and sensitivity analysis for determining the tolerance to distinct noise levels and spurious measurements. The design and implementation of the algorithm makes use of the open-source Point Cloud Library, and access to its source code is included within this work.
Point-to-CAD 3D Registration Algorithm for Relative Navigation Using Depth-Based Maps
2019-03-01
749092 byte
Conference paper
Electronic Resource
English
JTIDS relative navigation and data registration
IEEE | 1992
|Registration method based on ICP algorithm for 3D surgical navigation
British Library Online Contents | 2009
|