Point cloud clustering algorithm for autonomous vehicle based on 2.5D Gaussian Map

Multi-beam Lidars have many good features, such as: high accuracy, wide range, strong anti-interference capacity, and the ability of obtain multiple information (depth, reflectivity) and so on, which excellent performance gain greatly attention by the unmanned system industry, and it is very applicable to the environment perception of autonomous cars. But the vast amount of data and the sparsity make its point cloud very difficult to process. In this paper, a 2.5D map based on Gaussian distribution is purposefully proposed, and a composite distance is used to cluster the point cloud. Experiments show that this clustering method can effectively solve the problem that the same obstacle cannot gather together caused by shadow or the sparsity of point cloud. Compared to previous methods, it achieves excellent performance with high accuracy and stable robustness and meet the requirements of real-time system with less than 100ms average running time.