Correspondence rejection by trilateration for 3D point cloud registration

Recent years have shown increases in virtual 3D perception and applications, many of these applications require 3D model reconstruction from high quality LIDAR scans. High quality 3D models may be acquired from a collection of overlapping LIDAR scans which need to be registered or aligned to a common coordinate system. This paper investigates the use of a novel implementation of trilateration for correspondence rejection in highly accurate 3D point cloud registration. It is shown that from a synthesized correspondence set of size 100 containing 85% outliers, all or most of the remaining 15% inliers can be retrieved. The trilateration problem is solved for all 4-combinations of correspondence elements from which the true correspondence subsets are easily identifiable. It is also shown that this method's performance may be greatly affected by noisy distance measurements, however the method works well for distance measurements typically acquired by LIDAR systems. Lastly, unnecessarily large sizes of correspondence sets can quickly make the method computationally expensive if all combination subsets require to be evaluated.