CAO-RONet: A Robust 4D Radar Odometry with Exploring More Information from Low-Quality Points

TL;DR

This work addresses robust ego-motion estimation with 4D radar under sparsity and noise, proposing CAO-RONet, a radar-odometry network tailored to low-quality points. It combines local completion to densify sparse radar data, context-aware association for resilient multi-scale matching, and a bi-directional clip-window optimizer to enforce temporal continuity using historical priors. On the View-of-Delft dataset, CAO-RONet achieves around a 50% reduction in RMSE compared to prior methods and attains accuracy competitive with LiDAR-based odometry, while running at about 50 FPS. The approach advances all-weather, radar-only localization capabilities and demonstrates robust performance in degraded conditions; code release is planned.

Abstract

Recently, 4D millimetre-wave radar exhibits more stable perception ability than LiDAR and camera under adverse conditions (e.g. rain and fog). However, low-quality radar points hinder its application, especially the odometry task that requires a dense and accurate matching. To fully explore the potential of 4D radar, we introduce a learning-based odometry framework, enabling robust ego-motion estimation from finite and uncertain geometry information. First, for sparse radar points, we propose a local completion to supplement missing structures and provide denser guideline for aligning two frames. Then, a context-aware association with a hierarchical structure flexibly matches points of different scales aided by feature similarity, and improves local matching consistency through correlation balancing. Finally, we present a window-based optimizer that uses historical priors to establish a coupling state estimation and correct errors of inter-frame matching. The superiority of our algorithm is confirmed on View-of-Delft dataset, achieving around a 50% performance improvement over previous approaches and delivering accuracy on par with LiDAR odometry. Our code will be available.

Figures (7)

• Figure 1: Comparison of odometry accuracy. The black points are LiDAR map constructed using ground-truth poses, while red points denote radar map assembled from predicted poses.
• Figure 2: The overview of our proposed CAO-RONet. At first, the two frames of radar features derived from backbone are fed into LCM to densify sparse points. Then, CAM implements feature-assisted registration to associate point pairs in different scales, followed by correlation balancing to suppress outliers. Finally, COM with sequential state modeling applies historical prior from clip window to constraint the current ego-motion prediction and smooth trajectory.
• Figure 3: Average translational and rotational errors on the test sequences of VoD in the length of 20, 40, ..., 160$m$.
• Figure 4: The trajectory visualization of our CAO-RONet with other methods on sequences 00, 03, 04, 07 and 23, respectively.
• Figure 5: The effect of different modules on VoD dataset.