Real-Time Point Cloud Clustering Algorithm Based on Roadside LiDAR

Abstract

Light detection and ranging (LiDAR) is a crucial roadside intelligent perception device in cooperative vehicle infrastructure systems, which can generate a large amount of disordered 3-D point cloud data. Point cloud clustering serves as a prerequisite for road target identification, trajectory tracking, and traffic conflict prediction. However, due to limitations in data collection methods and clustering algorithms, a pronounced delay in target clustering exists. In this work, a real-time point cloud clustering algorithm for roadside LiDAR (RTPCC-RL) is proposed, which primarily comprises three aspects: online point cloud capture, background point cloud filtering, and real-time clustering of target point clouds, implemented using a C++ program. Initially, point cloud information can be captured and extracted by decoding user datagram protocol (UDP) packets. Upon receiving a UDP data packet, background filtering is achieved through the point distance difference method, reducing computational complexity and enhancing processing speed. Furthermore, the point cloud information of the background frame is stored in various matrices based on different vertical angles for efficient point location. Real-time clustering of target point clouds is executed based on voxel data features. The spatial region of the lane can be divided into voxel grids, and voxel point cloud coordinates are sequentially organized. The proposed algorithm was compared with other methods and the experimental results with a 32-channel LiDAR demonstrated that the RTPCC-RL effectively clustered road vehicle cloud points online with high precision and recall rates, achieving a processing time of 100 ms. The code can be found on the website https://github.com/480196239xiaoman/LiDAR/blob/main/online_cluster.cpp .