Semantic Lidar Odometry and Mapping for Mobile Robots Using RangeNet++

Abstract

The point cloud registration methods in the existing Simultaneous Localization and Mapping (SLAM) systems are all based on static assumptions, and dynamic points will affect the registration accuracy. Therefore, it is difficult for SLAM systems to operate in a highly dynamic environment. If only the geometric information of the point cloud data is used, further processing is required to avoid the influence of dynamic objects on the point cloud registration algorithm. To this end, we utilize the RangeNet++ method to identify the semantic information of 3D laser point clouds. By combining with semantic information, outliers caused by highly dynamic objects can be directly filtered out, and the objective function of point cloud registration in laser odometry is optimized. We can build high-quality semantic maps that are semantically and geometrically consistent. The experimental results prove that more accurate lidar odometry results can be obtained. The high-quality full-static semantic maps will be established to ensure the completeness of the traversable area of the mobile robot in subsequent navigation planning.