LodeStar: Maritime Radar Descriptor for Semi-Direct Radar Odometry

TL;DR

This paper tackles maritime radar odometry for USVs under challenging radar conditions with limited resolution and sparse landmarks. It introduces LodeStar, a circularly cross-correlated radial descriptor that captures coastal context and enables robust rotation estimation, paired with marine-specific feature extraction (contour-based points and $k$-nearest candidates) and a semi-direct fusion approach to ego-motion via point-normal matching. The method demonstrates significant odometry improvement over state-of-the-art sparse, dense, and hybrid baselines on Pohang Canal and Ulsan datasets, particularly in curved and coastal regions, by reducing rotational drift and improving translation estimates. This maritime radar-oriented framework paves the way for radar-only SLAM by integrating a robust descriptor with feature-driven optimization, offering practical impact for navigation in visually challenging maritime environments where cameras or LiDAR may fail.

Abstract

Maritime radars are prevalently adopted to capture the vessel's omnidirectional data as imagery. Nevertheless, inherent challenges persist with marine radars, including limited frequency, suboptimal resolution, and indeterminate detections. Additionally, the scarcity of discernible landmarks in the vast marine expanses remains a challenge, resulting in consecutive scenes that often lack matching feature points. In this context, we introduce a resilient maritime radar scan representation LodeStar, and an enhanced feature extraction technique tailored for marine radar applications. Moreover, we embark on estimating marine radar odometry utilizing a semi-direct approach. LodeStar-based approach markedly attenuates the errors in odometry estimation, and our assertion is corroborated through meticulous experimental validation.

Figures (9)

• Figure 1: Terrestrial radar-based odometry estimation techniques cannot be directly transposed to marine radar systems. As illustrated in the upper right side, even the most advanced radar odometry methodologies fail to generate accurate trajectories. Nonetheless, our maritime place descriptor LodeStar, incorporated with marine-specific features, effectively captures the rotational dynamics of subsequent frames, leading to enhanced odometry correction.
• Figure 2: The proposed framework for maritime odometry estimation. Exclusively using marine radar imagery, we extract pertinent marine features and the LodeStar descriptor. Subsequently, we construct an initial rotated point cloud and identify correspondences between two consecutive frames via a point normal-based approach. This process derives a rotation-enhanced ego-motion estimation.
• Figure 3: Details of the LodeStar descriptor. From the primary radar imagery, we compute a radial intensity vector for every $\theta$. Each of these vectors is subsequently integrated to constitute a column within the integrated radial intensity matrix. By iterating this process over a period of 2$\pi$, we synthesize the final descriptor.
• Figure 4: (a) In our dataset, certain regions remain un-updated due to the low update rate. An accumulation in these overlapping areas could lead to unexpected convergence behaviors. (b) Given that radar imagery inherently constitutes scan-rotated images, there exists a discrepancy between the commencement and conclusion of the scans, primarily attributed to temporal lags. This discrepancy is especially pronounced during sharp rotational movements. (c) Initially, we delineate the contour, subsequently identifying proximal points to ensure consideration extends to the wide area.
• Figure 5: For the given dataset, a cursory classification of the routes can be made as follows: Pohang-b,c and Ulsan01,02 represent more straightforward routes. In contrast, Pohang-d and Ulsan03 pose greater challenges for odometry estimation.