LiDAR-SLAM Using Semantic Information How to Deal with Dynamic Objects?

Abstract

Localization of mobile robots is needed in all automation tasks, from indoor to outdoor and from automotive over agriculture to exploration. Most Simultaneous Localization and Mapping (SLAM) algorithms based on LiDAR data use an occupancy grid representation. LiDAR measurements are used to populate the occupancy grid and clear the intermediate cells through ray tracing. Additionally, the spatial information is annotated with intensity or RGB information. In this paper, we propose a two-track SLAM process based on the semantic classification information. On the one hand, LiDAR measurements of dynamic objects are filtered to avoid mislocalization and achieve an accurate global localization. On the other hand, all information in the LiDAR point cloud is used for short term localization and odometry estimation. The proposed approach therefore utilizes all information to the highest extent.