A Semantic Segmentation Based Lidar SLAM System Towards Dynamic Environments

摘要

The Simultaneous Localization and Mapping (SLAM) ability is essential for autonomous driving and intelligent mobile robots. A large number of methods have been proposed to solve this problem, and outliers rejection in dynamic environments plays an important role in SLAM system. In this paper, we propose a semantic segmentation based Lidar SLAM system, which introduces semantic segmentation into Lidar SLAM system and improves the accuracy of the SLAM system in dynamic environment. A CNN based deep learning method is adopted for semantic segmentation and understanding of the environment. We use semantic segmentation to get rid of dynamic outliers, and then achieve motion estimation and environment reconstruction. We evaluate our method on the public KITTI datasct, and the results show that our proposed method can efficient reject the dynamic outlier and improve the performance in terms of accuracy.