Abstract: The traditional ICP algorithm has long matching time and is affected by initial values in LiDAR target point cloud matching, which leads to low positioning accuracy and poor robustness when applied to unmanned vehicle SLAM technology. Proposes an NDT-ICP algorithm that combines the KD-tree algorithm. Firstly, voxel grid filtering is used to preprocess the point cloud data obtained from LiDAR, and the method of plane fitting parameters is used to remove point cloud of ground. Secondly, use NDT algorithm for point cloud coarse matching to shorten the distance between the target point cloud and the point cloud to be matched. Finally, the KD-tree proximity search method is used to improve the speed of corresponding point search, and the precise matching of the ICP algorithm is completed by optimizing the convergence threshold. Through experiments, it has been shown that the improved algorithm proposed in this article has significantly improved speed and accuracy in point cloud matching compared to NDT and ICP algorithms, and has better accuracy and robustness in map construction.

Key words: unmanned vehicle, point cloud registration, ICP algorithm, NDT algorithm, laser SLAM