3D-Lidar based obstacle detection and fast map reconstruction in rough terrain

Abstract

Obstacle detection based on 3D-Lidar is an important sensing means of automation navigation. In rough terrain, due to the unstructured characteristics of the field environment and the bumpy road, it is difficult to accurately detect obstacles by Lidar in a single scan. Meanwhile the reconstruction of obstacles in the blind area of the sensor is also a key problem of autonomous navigation for safety. We propose a 3D-Lidar based obstacle detection and fast map reconstruction method to solve these problems in rough terrain. First, point clouds are projected onto the 2D grid map of overhead view. we use gradient statistics in the height direction to obtain a basic obstacle detection result. Then according to the obstacle detection result in single frame, the proposed method uses Bayesian probability calculation based on the first-order Markov model for perception map reconstruction. The final perception result consists of the current frame perception result and historical results by map reconstruction. And a probability map on the confidence value of obstacle is generated rather than the traditional binary grid map of obstacle and obstacle-free area. This method has high robustness for obstacle detection in rough terrain and can effectively solve the problem of environmental perception in unstructured scenes. In addition, obstacle in the blind area of the sensor is retained by map reconstruction to ensure the safe driving of the UGV. The calculation period of the method is within 100ms and meets the real-time requirement for autonomous navigation of the UGV.