Enhanced light detection and ranging simultaneous localization and mapping based on three-dimensional moving object tracking in dynamic scenes

Abstract

Moving objects can cause incorrect feature matches, significantly affecting light detection and ranging (LiDAR) odometry accuracy and mapping quality in simultaneous localization and mapping (SLAM). This paper proposes a learning-based SLAM framework that reduces the impact of moving objects on SLAM by estimating their motion through segmentation and tracking. We design a deep dynamic LiDAR odometry (DyLO) network by introducing a hierarchical attention mechanism and a dynamic mask network in point clouds to improve odometry performance in dynamic scenes. Additionally, we design a dynamic object segmentation and tracking module to extract motion information from dynamic scenes and combine it with DyLO to form a joint factor graph backend optimization for SLAM, thereby developing a complete SLAM system. Our SLAM system is validated on the public dataset with 64-line LiDAR and a self-established campus dataset with 16-line LiDAR. We compare the proposed DyLO with several outstanding iterative closest point (ICP) methods, the deep learning-based method of LiDAR odometry network (LO-Net), and the feature-based odometry approach. Compared with the outstanding LO-Net, the proposed algorithm reduces the translation error by 4.5% and the rotation error by 3.3% on average. Furthermore, our ablation experiments demonstrate that incorporating motion estimation of moving objects and backend optimization can remarkably improve the odometry accuracy.