A Fast GPU Point-cloud Registration Algorithm

The purpose of point cloud registration is to find a 3D rigid body transformation so that the 3D coordinates of the point cloud at different angles can be correctly matched. With the current advances of high definition (HD) 3D cameras, several applications have emerged utilizing stereoscopic cameras. For real time objects tracking, fast point cloud registration calculations are required. In the current study we have considered two methods: a standard Singular Value Decomposition (SVD) and a truncated SVD (TSVD) point cloud registration algorithm. The registration process can be summarized in the following steps; the centroids of the chosen point datasets are first found, then they both are aligned to the origin, and then the optimal rotation and translation are determined based on the SVD or on the TSVD technique. Our strategy was firstly to identify the major computational bottlenecks of our code, and secondly parallelize them on the GPU accordingly. Performance tests were conducted on three GPU cards in comparison to a serial version of the algorithm executed on a CPU. Performance comparisons are also conducted between the parallel SVD and the parallel TSVD in order to test the computational efficiency of them on GPU cards. The studies indicated that there is no computational benefit from the parallization of the simple SVD on GPU. On the contrary, there is computational advantage of the parallel TSVD, but it varies with the GPU architecture. Speedup factors were recorded for every registration steps for all GPU cards. The step 2 of registration process was the most computational expensive task for the algorithm, and when it was parallelized on K40m card gave a maximum speed up of ~100 for the maximum number of pixels, while for other resolution sizes the performance of K40m decreased dramatically. The GTX1080Ti card achieved the highest speed up of ~150 for block 2 calculations, for 8K resolution. In overall, for the full registration process GTX1080Ti indicated a linear increase of speedup factors versus the number of pixels, fact that renders it is the most suitable GPU card with respect to the other GPU cards used for the specific application.