SEGMENT-BASED LIDAR ODOMETRY FOR LESS STRUCTURED OUTDOOR SCENE

Weitong Wu, C. Chen, J. Li, Yangzi Cong, Bisheng Yang2020 · DOI

Abstract

Accurate registration of sparse sequential point cloud data frames acquired by a 3D light detection and ranging (LiDAR) sensor is a prerequisite for the back-end optimization of general LiDAR SLAM algorithms. This process, known as LiDAR odometry, often suffers from high drift in natural environments. This paper proposes a segment-based sampling strategy designed for robust LiDAR odometry in less structured outdoor scenes. By prioritizing stable physical structures over transient vegetation noise, the proposed method achieves significantly lower drift than state-of-the-art methods like LOAM and LeGO-LOAM.

Introduction

LiDAR odometry estimates a robot's movement by comparing sequential laser scans. While existing algorithms perform exceptionally well in structured urban environments characterized by flat walls and sharp corners, they often struggle in "less structured" natural settings. In these scenes, features such as tree leaves and bushes introduce significant noise because they move in the wind or appear inconsistently from different angles. This paper addresses the challenge by being more selective about which points are used for feature matching, focusing on stable clusters rather than raw point density.

Methodology

The authors propose a two-stage pipeline that integrates point cloud segmentation with feature-based odometry to ensure reliable tracking in messy environments.

The workflow of the proposed method, illustrating the transition from raw LiDAR scans to 6DOF pose estimation via segmentation and feature extraction.

1. Point Cloud Segmentation

Instead of extracting features from the raw point cloud, the system first performs a fast range image-based segmentation. It separates the ground from non-ground points and groups non-ground points into clusters. Only clusters with more than 20 points—likely representing solid objects like tree trunks, poles, or facades—are retained. This effectively filters out the "noise" of individual leaves and sparse vegetation.

2. Feature-Based Odometry

From these filtered segments, the system extracts two types of "low-level" features:

  • Edge points: Sharp corners or vertical lines.
  • Planar points: Flat surfaces like the ground or building faces.

These features are aligned to a local map using a sliding window "scan-to-map" strategy. The Levenberg-Marquardt method is used to minimize point-to-line and point-to-plane distances, while the initial pose for each new frame is estimated based on the constant velocity of previous frames.

Experimental Setup

The researchers evaluated the system using a Clearpath Jackal UGV equipped with a Velodyne VLP-16 sensor and an NUC on-board processor.

The experimental platform: a Clearpath Jackal UGV deployed in a typical less structured outdoor environment used for testing.

Testing was conducted in two typical less structured outdoor scenes in Wuhan, China:

  1. Site 1: A 1.5 km loop road within Wuhan University.
  2. Site 2: A greenway at Donghu Lake, characterized by heavy vegetation and minimal man-made structures.

Key Results

The proposed method demonstrated a clear lead in accuracy and stability over LOAM and LeGO-LOAM:

  • Trajectory Accuracy (Site 1): The proposed method achieved a position drift of only 0.5% (7.5 meters total closure error). In comparison, LOAM had a 1.24% drift (18.6m), and LeGO-LOAM had a 1.62% drift (24.3m).
  • Precision (Site 2): Using Plane Fitting RMSE as a proxy for map quality, the proposed method achieved 3.0 cm RMSE, significantly outperforming LOAM (4.1 cm) and LeGO-LOAM (7.9 cm).

A major finding was that standard LOAM suffers from significant Z-axis drift in these scenes because it attempts to match the chaotic geometry of tree canopies. By focusing on stable segments, the proposed method maintains much higher elevation stability.

Conclusion

The "less is more" approach to feature selection proves that in messy environments, the quality of feature points matters more than quantity. By aggressively filtering out small clusters, the algorithm identifies reliable landmarks that reduce cumulative matching errors. Future work will involve integrating IMU data to handle motion distortion and moving toward a full, globally consistent 3D mapping system.

Summary updated

Abstract

Accurate registration of sparse sequential point clouds data frames acquired by a 3D light detection and ranging (LiDAR) sensor like VLP-16 is a prerequisite for the back-end optimization of general LiDAR SLAM algorithms to achieve a globally consistent map. This process is also called LiDAR odometry. Aiming to achieve lower drift and robust LiDAR odometry in less structured outdoor scene using a low-cost wheeled robot-borne laser scanning system, a segment-based sampling strategy for LiDAR odometry is proposed in this paper. Proposed method was tested in two typical less structured outdoor scenes and compared with other two state of the art methods. The results reveal that the proposed method achieves lower drift and significantly outperform the state of the art.
BibTeX
@inproceedings{wu2020segment,
  author = {Weitong Wu and C. Chen and J. Li and Yangzi Cong and Bisheng Yang},
  title = {{SEGMENT-BASED LIDAR ODOMETRY FOR LESS STRUCTURED OUTDOOR SCENE}},
  booktitle = {˜The},
  year = {2020},
  doi = {10.5194/isprs-archives-xliii-b1-2020-535-2020},
}
Figure