Multi-view Based Clustering of 3D LiDAR Point Clouds for Intelligent Vehicles

Abstract

Abstract3D point clustering is important for the LiDAR perception system involved applications in tracking, 3D detection, etc. With the development of high-resolution LiDAR, each LiDAR frame perceives richer detail information of the surrounding environment but highly enlarges the point data volume, which brings a challenge for clustering algorithms to precisely segment the point cloud while running with a real-time processing speed. To meet this challenge, we innovate a multi-view (bird’s eye view and front view) based clustering method, named MVC. The method contains two stages. In the first stage, we propose a density image based algorithm, PG-DBSCAN, to segment the point cloud in bird’s eye view (BEV), which derives the preliminary division with fairly low computation resources. Then in the second stage, a front view (FV) clustering process is integrated to refine the under-segmented clusters. Our method takes both the speed and precision advantages of BEV and FV clustering, and this coarse-to-fine architecture reasonably allocates the computation resources and shows a real-time outstanding clustering performance. We evaluate the MVC algorithm both on the publicly available dataset with 64-line LiDAR and our own dataset with 128-line LiDAR. Compared with other clustering methods, MVC is able to derive more accurate clustering results. Specifically, toward the 128-line LiDAR with large data volume, our method shows an outperforming running speed, which perfectly fits on the LiDAR perception tasks.KeywordsPoint Cloud SegmentationHigh Resolution LiDARPG-DBSCAN