UD-LIOM

A Universal Tightly-Coupled Direct LiDAR-Inertial Odometry and Mapping Framework


Baosheng Zhang    Lin Zhang*    Shengjie Zhao    Yicong Zhou   
*corresponding author
Tongji University
Paper Code Data

UD-LIOM is a universal LiDAR SLAM technology designed to be compatible with both mechanical and various solid-state LiDARs. It addresses the long-standing challenge of algorithm incompatibility caused by differences in LiDAR scanning patterns. By supporting diverse LiDAR types, UD-LIOM significantly expands the applicability of LiDAR SLAM, particularly for emerging solid-state LiDARs. Its key features include:

  • A novel, tightly-coupled direct LiDAR-inertial odometry and mapping framework for LiDARs with different scanning patterns.
  • An innovative point cloud classification and registration method, designed to work well with diverse scanning patterns, especially those of solid-state LiDARs with repetitive scanning.
  • A new multi-constraint optimization model that combines LiDAR observation residual errors with IMU pre-integration.

The Performance of Different LiDARs on Challenging Environments

We have evaluated UD-LIOM across multiple LiDARs and diverse challenging environments. The results demonstrate its exceptional adaptability, with robust compatibility for LiDARs employing varying scanning patterns and consistent high performance in complex scenarios.

UD-LIOM Performance with RSM1 LiDAR

The figures below demonstrate the mapping results of UD-LIOM using the RSM1 LiDAR.

pipeline
pipeline
pipeline
pipeline

UD-LIOM Performance with Livox LiDAR

The figures below demonstrate the mapping results of UD-LIOM using the Livox LiDAR.

pipeline
pipeline

UD-LIOM Performance in Challenging Unstructured Environments

The figures below demonstrate the mapping results of UD-LIOM in challenging unstructured environments.

pipeline
pipeline
pipeline
pipeline
pipeline
pipeline

Video Demonstration

The following videos showcases the results of the localization and mapping achieved by UD-LIOM using the RSM1 LiDAR (Play at 10x speed).



The following videos showcases the results of the localization and mapping achieved by UD-LIOM in challenging unstructured environments (Play at 10x speed).


Abstract

This paper puts forward UD-LIOM, a brand-new framework specifically crafted to achieve universal compatibility with both mechanical and various solid-state LiDARs. In the front-end, a point cloud classification method based on range smoothness is presented. This method is used to divide the original point clouds into edge points and planar points. Thanks to this approach, UD-LIOM is able to adapt to different types of LiDARs. Moreover, it can establish more feature correspondences between LiDAR scans and submaps during the optimization process. As a result, the proposed framework can efficiently address the performance decline or failure issues that are frequently encountered in existing methods, especially when dealing with solid-state LiDARs that have a limited FoV. Furthermore, in the back-end, we innovatively combine IMU pre-integration with LiDAR residual errors within a multi-constraint optimization model. By leveraging the IMU's advantage in estimating high-precision local transformations in the back-end, the ability of UD-LIOM to handle extreme motion scenarios is effectively enhanced. To verify the effectiveness of UD-LIOM, we carry out both quantitative and qualitative evaluations. The results show that UD-LIOM outperforms in structured and unstructured environments, is compatible with a wide variety of solid-state LiDARs, and exhibits excellent performance.

Experimental Datasets

We conduct experiments on several publicly accessible datasets, namely KITTI, BotanicGarden, M2DGR, and NCLT. These datasets encompass a wide range of scenarios and incorporate multisensor data. To further assess the robustness and practical applicability of UD-LIOM, we also tested it on self-recorded datasets using a Velodyne VLP16 (mechanical LiDAR) and a RSM1 (solid-state LiDAR), alongside a 100 Hz IMU.

pipeline

Comparative Analysis of Universality

UD-LIOM demonstrates compatibility not only with mechanical LiDAR but also with the widely adopted Livox AVIA LiDAR and the RSM1 solid-state LiDAR.

pipeline

Localization Result Evaluation

We compare UD-LIOM with existing LiDAR SLAM methods, including A-LOAM, LeGO-LOAM, and LIO-mapping, on the KITTI dataset. Additionally, to evaluate the effectiveness of the proposed point cloud classification scheme, we also conduct ablation experiments focusing on edge points and planar points.

pipeline

We conduct comparative analyses using LINS, LIO-SAM, LiLi-OM, and FAST-LIO2 as benchmark methods on the M2DGR dataset. Additionally, we also conduct ablation experiments focusing on IMU factor in the back-end.

pipeline

We also conduct comparative experiments with Loam-Livox, LiLi-OM, FAST-LIO2, and UD-LIOM on the BotanicGarden dataset, which records unstructured scenes within botanical gardens, containing point clouds from Livox AVIA solid-state LiDAR.

pipeline

Running Time Analysis

We analyze the running time (ms) and results of each module in the framework with different LiDARs.

pipeline

Framework

In UD-LIOM, raw point clouds and IMU data first undergo preprocessing, including IMU pre-integration and point cloud distortion correction (aided by IMU results), after temporal synchronization between LiDAR and IMU. Then, the system classifies undistorted point clouds into edge or planar points by range smoothness and matches them with relevant submaps. Matching residual errors combined with IMU-derived local pose transformations form a multi-constraint optimization model, which serves as the objective function for local pose estimation and sliding window optimization, outputting the final system pose and global map. Finally, submaps are updated using localization and mapping results for future matching.

pipeline