Semantic-Assisted LIDAR Tightly Coupled SLAM for Dynamic Environments

Abstract

The Simultaneous Localization and Mapping (SLAM) environment is evolving from static to dynamic. However, traditional SLAM methods struggle to eliminate the influence of dynamic objects, leading to significant deviations in pose estimation. Addressing these challenges in dynamic environments, this paper introduces a semantic-assisted LIDAR tightly coupled SLAM method. Specifically, to mitigate interference from dynamic objects, a scheme for calculating static semantic probability is proposed. This enables the segmentation of static and dynamic points while eliminating both stationary dynamic objects and moving environmental blocking objects. Additionally, in point cloud feature extraction and matching processes, we incorporate constraint conditions based on semantic information to enhance accuracy and improve pose estimation precision. Furthermore, a semantic similarity constraint is included within the closed-loop factor module to significantly enhance positioning accuracy and facilitate the construction of maps with higher global consistency. Experimental results from KITTI and M2DGR datasets demonstrate that our method exhibits generalization ability towards unknown data while effectively mitigating dynamic interference in real-world environments. Compared with current state-of-the-art methods, our approach achieves notable improvements in both accuracy and robustness.