Dynamic Vehicle Detection With Sparse Point Clouds Based on PE-CPD

Abstract

Detecting dynamic vehicles is of great significance in the field of autonomous vehicles. In the literature, a few vehicle detection methods are proposed to detect vehicles within 50 m from the Lidar, where the point clouds are relatively dense. It is a great challenge to detect vehicles that are far from the Lidar because of sparse point clouds. Fewer returned point clouds will result in a larger fitting randomness and lower detection rate. To tackle this issue, a dynamic vehicle detection method based on likelihood-field-based model combined with coherent point drift (CPD), which includes the steps of dynamic object detection and dynamic vehicle confirmation, is proposed in this paper. An adaptive threshold based on the distance and grid angular resolution is applied to detect the dynamic objects. The pose estimation based on CPD (PE-CPD) is proposed to estimate the vehicle pose. The scaling series algorithm coupled with a Bayesian filter that is improved by PE-CPD is utilized for updating the vehicle states. Finally, comparative experiments of vehicle detection based on KITTI data sets are conducted. The results show that the proposed method improves the detection rate, especially the detection in the radius of 40–80 m, which is termed as distant area, compared with the method based on pose estimation with modified scaling series.