Adaptive Denoising-Enhanced LiDAR Odometry for Degeneration Resilience in Diverse Terrains

TL;DR

This work addresses LiDAR odometry drift in challenging, high-noise environments by introducing ambient-aware, degenerationsensitive filtering. The method adds ambient skeleton extraction and degeneration detection to a LOAM-based SLAM pipeline, using a fast range-image based clustering and normal-vector features to dynamically adjust point-cloud denoising. Key contributions include a fast adaptive Euclidean clustering on a range image, skeleton-based scene representation, and a degeneration-motivated thresholding strategy that enhances accuracy and robustness on KITTI and real-world data. The results demonstrate improved localization accuracy and mapping stability, with real-time performance suitable for diverse terrain scenarios and potential integration into other SLAM systems.

Abstract

The flexibility of Simultaneous Localization and Mapping (SLAM) algorithms in various environments has consistently been a significant challenge. To address the issue of LiDAR odometry drift in high-noise settings, integrating clustering methods to filter out unstable features has become an effective module of SLAM frameworks. However, reducing the amount of point cloud data can lead to potential loss of information and possible degeneration. As a result, this research proposes a LiDAR odometry that can dynamically assess the point cloud's reliability. The algorithm aims to improve adaptability in diverse settings by selecting important feature points with sensitivity to the level of environmental degeneration. Firstly, a fast adaptive Euclidean clustering algorithm based on range image is proposed, which, combined with depth clustering, extracts the primary structural points of the environment defined as ambient skeleton points. Then, the environmental degeneration level is computed through the dense normal features of the skeleton points, and the point cloud cleaning is dynamically adjusted accordingly. The algorithm is validated on the KITTI benchmark and real environments, demonstrating higher accuracy and robustness in different environments.

Figures (13)

• Figure 1: The importance of the same object varies a lot in different scenes. In the scene (a), there are already sufficient features provided, and the leaves introduce matching noise. In the scene (b), if the features provided by the leaves are not fully utilized, the scene will be too simple and leads to degeneration.
• Figure 2: We present the overall framework of our algorithm, focusing on the main contribution of ambient-aware point cloud filtering, which is integrated into the SLAM framework of LeGO-LOAM. The ambient-aware filtering consists of two key components: ambient skeleton extraction and degeneration scene detection. The former extracts point cloud data that represents the fundamental structure of the scene, while the latter analyzes the degree of scene degeneration and dynamically adjusts the threshold for point cloud segmentation and denoising. Our algorithm provides the dynamically denoised point cloud to the LiDAR odometry module, resulting in more accurate and robust localization performance.
• Figure 3: The process of projecting the collected point cloud into a 2D range image is achieved through the physical characteristics of the 32-line LiDAR sensor.
• Figure 4: Illustrations of the two range image based clustering methods. Depth clustering uses $\beta > \beta_0$ as the segmentation condition, with the search range limited to directly adjacent points, indicated by the orange area. On the other hand, Euclidean clustering uses $d < \gamma d_0$ as the segmentation condition, with the search range confined within the yellow search box.
• Figure 5: The effect of depth clustering on cleaning the point cloud at different threshold values. (a) Result with $\beta_0=10^{\circ}$: It retains the majority of the point cloud data. (b) Result with $\beta_0=30^{\circ}$: It filters out some points relative to $\beta_0=10^{\circ}$, but the overall effect is similar. (c) Result with $\beta_0=60^{\circ}$: It mainly preserves the structural information of the point cloud.