An Efficient Background Filtering Method for Roadside LiDARs

In the context of connected and automated transportation systems (CATSs), light detection and ranging (LiDAR) is increasingly being deployed on the roadside to detect detailed motions of road users (i.e., vehicles and pedestrians) for real-time applications. To provide real-time detection, it is essential to conduct background filtering of the LiDAR point cloud to eliminate LiDAR points irrelevant to the traffic objects. Background filtering can significantly reduce the computational load of implementing a traffic-detection algorithm. However, existing methods are not sufficiently fast and accurate for real-time applications. This study proposes an efficient method of background filtering method for roadside LiDAR data by solving the problems of existing methods in the background filtering procedure. In the proposed method, the octree is used to aggregate LiDAR frames, which can dramatically reduce storage space compared to simply superposing frames in existing studies. Second, by integrating ray casting and occupancy ratio (RCOR), the background can be extracted according to the spatial relations and statistical probabilities of objects. In the final stage, a sparse voxel octree (SVO) is applied to represent the background, and a GPU-based parallel filtering algorithm can expedite background filtering significantly. We conducted a field experiment to collect LiDAR data using various LiDARs installed at the roadside of a freeway segment in Chengdu, China. The results demonstrate that the proposed method performs best in terms of accuracy and computation speed in a comparison experiment. Its performance can remain robust with various types of LiDARs under various traffic conditions.

I. Introduction

Light detection and ranging (LiDAR) is an optical sensor that measures the distance by scanning an object with lasers. Raw LiDAR data are a set of points on objects’ surfaces, commonly called a 3-D point cloud. This allows us to obtain various information about the object, including position, size, and orientation, with high accuracy. Hence, LiDARs are increasingly being deployed on the roadside to monitor traffic in detail to facilitate the development of connected and automated transportation systems (CATSs). Specifically, they can continuously and timely collect detailed vehicle trajectories, forming a vital foundation for real-time traffic control in CATS, such as vehicle collision warnings, lane-dependent variable speed limits, and trajectory control of connected and automated vehicles (CAVs). However, 3-D point clouds contain a large number of background points beyond road users (i.e., vehicles and pedestrians), such as the ground, buildings, and trees [1], [2]. Object detection and tracking computational load can be hefty if background points are processed repeatedly in each frame [3]. Therefore, background filtering is essential for real-time trajectory collection by roadside LiDARs [1], [4].