A Dynamic Clustering Algorithm for Lidar Obstacle Detection of Autonomous Driving System

Lidar is an important sensor of the autonomous driving system to detect environmental obstacles, but the spatial distribution of its point cloud is non-uniform because of the scanning mechanism. For adaption to this spatial non-uniformity, a dynamic clustering algorithm is proposed based on the spatial distribution analysis of the point cloud along different directions. The proposed algorithm adopts an elliptical function to describe the neighbor, whose semi-minor and semi-major are adjusted dynamically according to the position of the core point. Base on the relationship analysis of different clustering parameters, they are further designed quantitatively by KITTI dataset considering comprehensive clustering performances. To validate the effectiveness of the proposed algorithm, several comparative experiments with different clustering methods and projection planes have been conducted in the campus by an electric sedan equipped with three IBEO LUX 8 lidars. The experimental results show that the proposed elliptical neighbor can deal with the uneven point cloud more effectively, the performances of over-segmentation, under- segmentation and missed detection all are improved and accordingly a higher detection accuracy is achieved.