A fast and robust scan matching algorithm based on feature dependent sampling

In this paper, a fast and robust feature based scan matching algorithm for LiDAR systems, especially dedicated to outdoor SLAM applications, is introduced. There are large numbers of variants of scan matching algorithms, and they are mostly based on matching of the whole or randomly sampled points. However, this study shows that the sampling strategy is extremely important for a fast and reliable scan matching. In this study, we give point to four different sampling strategies for scan matching. The first one is the regular sampling which is mostly preferred due to its simplicity, and the second one is the normal random sampling. The third strategy is the grid or voxel based sampling, where the point cloud is divided into fixed cells then the equal numbers of points are sampled from each cells. In this way, it is obtained a more uniform sampling method than the others. The last special form of sampling strategy that we propose is based on the feature extraction. The plane segments are extracted from the input scan, and then the matching is performed based on these plane points. The results show that it is possible to obtain a fast and robust scan matching algorithm by using almost %3 percent of the data in matching. Although the introduced sampling methods are independent of the scan matching algorithm, the Multi-Layered Normal Distribution Transform is used for this purpose. In summary, the proposed method has several advantages over conventional methods; the first advantage is that the method becomes more robust to outliers since the matching is based on planar structures. The second advantage is that the algorithm becomes much faster since the number of points matched is very less with respect to all points. Therefore, the feature extraction can be considered as a special sampling strategy. For performance analysis, the real datasets are used and the proposed sampling methods are compared with each other. The results are very attractive for 3D scan matching in SLAM.