Harnessing U-disparity in Point Clouds for Obstacle Detection

Abstract

Lidar is an indispensable equipment in autonomous vehicle for environment perception, in which obstacle detection plays an important role in collision avoiding and route planning. The main challenge of Lidar based obstacle detection is that processing the disordered and sparse point clouds would be difficult and time-consuming. Hence, this paper presents a novel usage of U-disparity to locate obstacles indiscriminately with point clouds, which makes obstacle detection effective and efficient. The proposed method firstly uses cross-calibration to align point cloud with reference image, so that a depth map is formed. Then, the U-disparity map is introduced to process Lidar based depth map. Due to the particularity of Lidar based U-disparity, we select local peaks in column in U-disparity to identify relevant disparities of obstacles. After applying filtering and clustering steps on these salient peak disparities, the corresponding obstacles can be precisely localized. Quantitative and qualitative experimental results on KITTI object detection benchmark and road detection benchmark reveal that the proposed method achieves very encouraging performances in various environments.