Adaptive LiDAR-Radar Fusion for Outdoor Odometry Across Dense Smoke Conditions

TL;DR

The paper presents an adaptive LiDAR-radar fusion framework to achieve robust odometry in dense smoke by detecting LiDAR degeneracy and selectively leveraging radar data. A three-stage pipeline preprocesses Radar point clouds, detects LiDAR-degenerated regions, and removes dynamic LiDAR points using radar-derived cues, enabling reliable LIO backends in adverse conditions. Experimental validation on the NTU4DRadLM dataset shows the approach outperforms LiDAR-only methods and radar-only baselines in challenging smoke scenarios, while effectively identifying degraded regions and removing dynamic content. The work highlights the practical value of cross-modality fusion for resilient perception in outdoor environments where LiDAR performance deteriorates, and outlines future work on cross-modality registration and broader testing.

Abstract

Robust odometry estimation in perceptually degraded environments represents a key challenge in the field of robotics. In this paper, we propose a LiDAR-radar fusion method for robust odometry for adverse environment with LiDAR degeneracy. By comparing the LiDAR point cloud with the radar static point cloud obtained through preprocessing module, it is possible to identify instances of LiDAR degeneracy to overcome perceptual limits. We demonstrate the effectiveness of our method in challenging conditions such as dense smoke, showcasing its ability to reliably estimate odometry and identify/remove dynamic points prone to LiDAR degeneracy.

Figures (5)

• Figure 1: Example of LiDAR and radar point clouds in smoke region. Occluded area is indicated by red boxes. As depicted in the figure, LiDAR point cloud exhibits occluded points due to dense smoke, whereas radar demonstrates robust perception.
• Figure 2: System diagram of proposed method. If the algorithm determines the LiDAR Degenerated Area, the radar static point cloud is used; otherwise, the LiDAR static point cloud is used as the input for the LIO method.
• Figure 3: Example of radar dynamic point cloud $\mathcal{P}^R_{d}$ (red) and LiDAR point cloud $\mathcal{P}^L$ (blue). This scene depicts a scenario where a person is in motion.
• Figure 4: Trajectory result of sequence (a) $\texttt{cp}$, (b) $\texttt{garden}$ and (c) $\texttt{smoke}$. $\texttt{LiDAR}$, $\texttt{Ours}$, $\texttt{Radar}$ and $\texttt{GT}$ correspond to red, blue, green, and black, respectively. LiDAR scans were deleted within the Orange box area to simulate the smoke sequence effect.
• Figure 5: Demonstration of the effect of $\textit{Removing}$$\textit{Moving}$$\textit{Points}$$\textit{in LiDAR Point Cloud}$. In (a) left point cloud shows static LiDAR point cloud (gray) and dynamic LiDAR point cloud (blue) together. In (b), we use FAST-LIO2 to obtain map. $\mathcal{P}^L$ is used for input to make left map and $\mathcal{P}^L_{s}$ is used for input to make right map. After removing LiDAR point cloud, the trace (red box) is removed and map quality get better.