Three-Dimensional Real-Time Object Perception Based on a 16-Beam LiDAR for an Autonomous Driving Car

摘要

Object perception is essential for autonomous driving applications in urban environment. A 64-beam LiDAR is a widely-used solution in this field, but its high price has prevented it from broader applications of autonomous driving technology. An alternative solution is to adopt a 16-beam LiDAR or multiple 16-beam LiDARs. However, 16-beam LiDAR obtains relative sparse data that makes object perception more challenging. In this paper, a new perception method is proposed to tackle problems caused by sparse data obtained from a 16- beam LiDAR. First, a segmentation method is proposed based on 2D grid image where a free space constraint is employed to reduce unreasonable image dilation and some segments are merged based on prior knowledge. Then, selective features of bounding box are employed in association process for a more accurate result given the sparse data. The proposed method is evaluated on an autonomous driving car in real urban scenarios. The results show that segmentation error can be as low as 7.7% with the free space constraint and prior knowledge, and absolute tracking error and the overall classification accuracy are 0.44 m/s and 93.33 % respectively.