Robust and Fast Point Cloud Registration for Robot Localization Based on DBSCAN Clustering and Adaptive Segmentation

This paper proposes a registration approach rooted in point cloud clustering and segmentation, named Clustering and Segmentation Normal Distribution Transform (CSNDT), with the aim of improving the scope and efficiency of point cloud registration. Traditional Normal Distribution Transform (NDT) algorithms face challenges during their initialization phase, leading to the loss of local feature information and erroneous mapping. To address these limitations, this paper proposes a method of adaptive cell partitioning. Firstly, a judgment mechanism is incorporated into the DBSCAN algorithm. This mechanism is based on the standard deviation and correlation coefficient of point cloud clusters. It improves the algorithm's adaptive clustering capabilities. Secondly, the point cloud is partitioned into straight-line point cloud clusters, with each cluster generating adaptive grid cells. These adaptive cells extend the range of point cloud registration. This boosts the algorithm's robustness and provides an initial value for subsequent optimization. Lastly, cell segmentation is performed, where the number of segments is determined by the lengths of the adaptively generated cells, thereby improving registration accuracy. The proposed CSNDT algorithm demonstrates superior robustness, precision, and matching efficiency compared to classical point cloud registration methods such as the Iterative Closest Point (ICP) algorithm and the NDT algorithm.