Robust Loosely Coupled SLAM System with LiDAR and IMU for Multi-Curve Dynamic Scenarios

With the widespread development of ground mobile robot applications, high-precision localization and mapping capabilities have become crucial. In this paper, a loosely coupled localization and mapping system, PED-LOAM, based on Lidar and IMU is proposed. This system utilizes the high-precision pose output from Lidar odometry and combines IMU data to achieve accurate localization and mapping. Unlike most solutions based on the assumption of static objects to reduce the impact of dynamic environments on localization accuracy, this paper first employs the advanced ground segmentation method Patchwork++ for preprocessing. It then uses Iterated Error State Kalman Filter (IESKF) to fuse point cloud matching odometry data based on distance and normal vector constraints with IMU data. In the backend fine matching process, a method for dynamic point cloud removal is proposed. By removing sparsely dense point clouds and utilizing an iterative point cloud co-visibility approach, dynamic objects are eliminated for iterative matching, thereby outputting high-precision localization information. Testing on the KITTI and M2DGR datasets shows that, for the KITTI dataset, the average percentage improvement in ATE Root Mean Square Error over LEGO-LOAM is 42.48% and over the advanced algorithm LVI-SAM is 33.04%. In the M2DGR dataset testing, the average percentage improvement in ATE Root Mean Square Error over LEGO-LOAM is 73.97%, and over EKF-LOAM in terms of RPE is 28.34%. These results demonstrate that the localization capability of the PED-LOAM system is more robust and accurate.

I. Introduction

With the rapid advancement of 6G communication networks [1], the application of ground mobile robots in emergency rescue, industrial automation, and smart living has been signif-icantly accelerated. Central to these robots' autonomous tasks are localization and mapping, which demand high accuracy and low drift. To meet these requirements, technologies such as LEGO-LOAM [2], FAST-LIO2 [3], and VINS-FUSION [4] leverage Lidar and vision-based SLAM solutions. Despite the high accuracy of visual SLAM, it is susceptible to environ-mental factors like low light, rapid illumination changes, and fast movements, which can degrade performance. Additionally, the high computational demand of visual SLAM can hinder real-time operation in complex environments.