Source-linked AI summary
Radar-only ego-motion estimation in difficult settings via graph matching
Sarah H. Cen, Paul Newman
TL;DR
Radar odometry must cope with low-resolution scans, speckle noise, and false detections, while radar’s stable long-range observations make it useful across challenging conditions. This paper proposes a radar-only pipeline with one-parameter keypoint extraction and graph-matching data association. Across diverse settings, the pipeline achieves high accuracy against GPS and performs comparably to, and sometimes better than, visual odometry.
Problem
Radar odometry must remain precise despite low-resolution data, slow updates, speckle noise, and false detections.
Source: §II · p. 1Method
The pipeline combines one-parameter keypoint extraction with data association framed as global graph matching, without a motion prior.
Source: §I · p. 1, C. Data association · p. 3Results
RO differs from GPS by only 5.2 cm and 0.09 deg in translation and rotation, respectively, on average; its overall median difference from GPS is lower than VO’s.
Source: §V · p. 6Takeaways & Limitations
The pipeline achieves high accuracy across diverse settings and performs comparably to, and at times better than, VO.
Source: §V · p. 5, §VI · p. 6
Abstract
from arXiv · showhide
Radar detects stable, long-range objects under variable weather and lighting conditions, making it a reliable and versatile sensor well suited for ego-motion estimation. In this work, we propose a radar-only odometry pipeline that is highly robust to radar artifacts (e.g., speckle noise and false positives) and requires only one input parameter. We demonstrate its ability to adapt across diverse settings, from urban UK to off-road Iceland, achieving a scan matching accuracy of approximately 5.20 cm and 0.0929 deg when using GPS as ground truth (compared to visual odometry's 5.77 cm and 0.1032 deg). We present algorithms for keypoint extraction and data association, framing the latter as a graph matching optimization problem, and provide an in-depth system analysis.
I. INTRODUCTION
Radar’s resilience across weather and lighting, and its ability to detect stable long-range features, make it suitable for odometry and localization. The revised radar-only pipeline targets noisy, artifact-prone scans with new keypoint extraction and data association methods, and is reported to run in real time across UK and Iceland settings.
I. INTRODUCTION
- Radar’s stable, long-range detections and resilience to variable lighting and weather support its use for odometry and localization.Unlike GPS, radar also functions indoors and outdoors.Source: §I · p. 1
- The radar-only pipeline handles significant measurement noise and missing detections without a prior on the temporal or spatial relationship between keypoint sets.Its scan matching is geometrically based, non-iterative, displacement-independent, and approximately optimizes a global objective.Source: §I · p. 1
- The revised system introduces gradient-based keypoint extraction with one input parameter and more efficient, robust data association.It also provides theoretical analysis of system complexity and estimation uncertainty.Source: §I · p. 1
- The pipeline achieves real-time, high-accuracy odometry across UK city streets and off-road Iceland, including challenging terrain.The figure demonstrates scan matching without a motion prior despite high measurement noise and significant appearance changes.Source: §I · p. 1, Figure 1 · p. 1
II. RELATED WORKS
Radar odometry must handle low-resolution scans contaminated by speckle noise and false detections. Existing methods may generalize poorly or depend on tuning, other sensors, restrictive assumptions, or SLAM; this paper presents a radar-only alternative with minimal prior knowledge.
II. RELATED WORKS
- Radar odometry must remain precise despite low-resolution data, slow updates, speckle noise, and false detections from multipath reflections and receiver saturation.Radar’s wide spreadbeam and long range contribute to lower resolution than lidar and susceptibility to clutter.Source: §II · p. 1
- Indirect methods extract salient keypoints before associating them, while direct methods operate on minimally pre-processed sensor outputs and are outside this paper’s scope.Both categories assume most observed objects are static.Source: §II · p. 2
- CFAR detection uses sliding-window thresholding, but CFAR variants generally need at least three tunable parameters and behave inconsistently across datasets.Other approaches use coherent structures, geometric primitives, scan-image features, or models such as NDT.Source: §II · p. 2
- ICP iteratively matches points and relies on a good relative-displacement estimate; other association methods optimize similarity, distance, map quality, scan distortion, or descriptors.These alternatives include NDT, cluster-edge difference, BASD, and SURF.Source: §II · p. 2
- Many methods generalize poorly under radar noise, and adequate performance often requires fine tuning, prior knowledge, restrictive assumptions, outlier detection, other sensors, or SLAM.Sensor dependence can compromise performance when those sensors fail; SLAM adds overhead and model-reliant motion filters.Source: §II · p. 2
- The proposed indirect pipeline uses interpretable, low-redundancy keypoints and data association that needs neither a motion prior nor parameter tuning and tolerates noise and false detections.Its keypoint extraction requires one input parameter and makes no scene-structure or noise assumptions.Source: §II · p. 2
A. FMCW scanning radar
The FMCW scanning radar measures reflected power by range at successive azimuths, then combines a full rotation into a two-dimensional scan. Its scans are information-rich but include noise and false detections.
A. FMCW scanning radar
- For each azimuth, scanning radar returns a power-range spectrum, and a full rotation combines these spectra into a two-dimensional scan.The radar sequentially observes narrow angular regions as it rotates.Source: A. FMCW scanning radar · p. 2
- Radar scans are information-rich and data-efficient, but contain noise and false detections and have lower resolution and slower updates than lidar.Source: A. FMCW scanning radar · p. 2
B. Keypoint extraction
The keypoint extractor selects meaningful, nonredundant features from radar scans while requiring only a maximum-keypoint parameter and no prior knowledge of scene structure or noise. Its peak- and region-based procedure is designed to adapt across settings and suppress isolated or distracting detections.
B. Keypoint extraction
- The design seeks interpretable keypoints, one maximum-count parameter, adaptation without noise-characteristic priors, reduced redundancy, and rejection of bright swathes from rolling, pitching, or clutter.Its only input parameter is the upper bound on desired keypoints.Source: B. Keypoint extraction · p. 3
- The extractor processes high-intensity, low-gradient regions, keeps one strongest return per marked range sequence, and removes isolated detections lacking azimuth neighbors.It caps selected regions using ℓmax and uses a Prewitt gradient.Source: B. Keypoint extraction · p. 3
- 250×250 m Iceland forest scans illustrate the progression from scan and intensity gradient to selected regions and a keypoint set capped by ℓmax.The resulting set is intended to capture scene structure while minimizing redundancy.Source: Figure 3 · p. 3
- The authors contrast the method with CFAR, which requires multiple tunable parameters, returns redundant detections, and struggles with bright rolling, pitching, or clutter patches.They report that their new algorithm has fewer parameters than their previous work and operates in unstructured environments such as forests.Source: B. Keypoint extraction · p. 3
C. Data association
Data association is framed as graph matching: pairwise geometric compatibility refines unary keypoint matches into a globally consistent set. A greedy approximation avoids requiring a motion prior or parameter tuning and uses a new incompatibility-based stopping condition.
C. Data association
- The pipeline is designed to handle measurement noise, false detections, and appearance changes while using global scan information to form correspondences.The figure illustrates unary proposals, pairwise consistency, the greedy termination condition, and the final alignment in Warwick, UK.Source: C. Data association · p. 3, Figure 4 · p. 4
- The method uses global geometric constraints to select mutually consistent correspondences without a motion prior or parameter tuning.Unlike ICP, it does not require an approximate scan alignment.Source: C. Data association · p. 3
- A rotation-invariant descriptor proposes unary matches using angular-slice and annulus histograms, with range weighting to correct radar’s range-density bias.Angular-slice counts are Fourier transformed and phase-normalized; annulus counts are normalized.Source: C. Data association · p. 4
- Pairwise compatibility scores compare relationships between proposed match pairs, and maximizing their global score forms an NP-hard graph-matching problem.The method relaxes the binary constraints and uses the principal eigenvector of the compatibility matrix to approximate the constrained solution.Source: C. Data association · p. 4
- A greedy algorithm ranks candidate matches by eigenvector value and stops when the next match is incompatible with those already selected.The revised termination condition does not depend on the magnitude of eigenvector elements.Source: C. Data association · p. 4
A. Algorithmic complexity
The system’s average computational cost is dominated by unary candidate matching in data association. The user-set maximum keypoint count is its only tunable input parameter.
A. Algorithmic complexity
- The analysis computes intensity gradients, ordering, and the principal eigenvector with the Prewitt operator, merge sort, and power iteration, respectively.The no-prior complexity case is also considered.Source: A. Algorithmic complexity · pp. 4–5
- Overall, system computation is generally dominated by unary candidate matching in Stage 1 of data association, with average complexity Θ(ℓ²_max(A + R)) ∼ Θ(ℓ³_max).To estimate average complexity, the analysis assumes A < ℓ_max < R; for most radars, A < R.Source: A. Algorithmic complexity · pp. 4–5
- The user sets ℓ_max as the system’s only tunable input parameter because the computational bottleneck depends on it.A mild relative-motion prior can reduce unary descriptor comparisons during real-time operation.Source: A. Algorithmic complexity · pp. 4–5
B. Theoretical analysis and performance measures
The analysis examines how incorrect proposed matches affect compatibility scores and the quality of the selected match set. It also defines measures for selected-match consistency and solution robustness to perturbations.
B. Theoretical analysis and performance measures
- G(m, C) overestimates the true global compatibility score unless all proposed matches are correct.Because selection can therefore be overly optimistic, the algorithm seeks to resolve conflicts and avoid adding poor matches.Source: B. Theoretical analysis and… · p. 5
- Under perfect measurements and no perfect symmetry, Proposition 2 identifies the optimal binary match vector as maximizing the relaxed global compatibility objective.The analysis treats incorrect matches as perturbations to the compatibility matrix.Source: B. Theoretical analysis and… · p. 5
- I_M measures mutual compatibility among selected matches, and the greedy algorithm terminates when adding a match would decrease it.The measure uses the angle between the selected-match vector and its compatibility-weighted vector.Source: B. Theoretical analysis and… · p. 5
- The normalized eigengap I_eg measures solution robustness to perturbations; it considers all proposed matches and reflects confidence over the next-highest-scoring solution.I_M is introspective about agreement among selected matches, whereas I_eg considers the full proposal set.Source: B. Theoretical analysis and… · p. 5
V. RESULTS
The new radar odometry pipeline achieves high accuracy across UK and Iceland settings, with average GPS differences of 5.2 cm in translation and 0.09 deg in rotation. It outperforms ICP across settings, generally approaches VO accuracy, and has higher variability than VO.
V. RESULTS
- A very mild motion prior caps possible acceleration at 8 m/s2 to remove impossible matches, but the system rarely needs it: 2 scan matches out of 10,000.Incorrect alignments can arise in highly symmetric scenes with many possible alignment solutions.Source: §V · pp. 5–6
- RO outperforms ICP in every setting and generally achieves accuracy comparable to, though slightly lower than, VO; it greatly outperforms VO on the highway.The lower rotational error ICP occasionally shows is misleading when its translation estimate is highly inaccurate.Source: §V · p. 6
- RO’s 3-DOF estimate can absorb rolling or pitching, and discarding elevation in 2D scans can affect estimates in less structured environments.These factors are consistent with slightly higher RO error in countryside, forest, and boulder-field datasets.Source: §V · p. 6
- RO differs from GPS by only 5.2 cm in translation and 0.09 deg in rotation on average; VO’s overall median GPS difference exceeds RO’s, while RO’s difference standard deviation is larger.Source: §V · p. 6
- The new pipeline improves on prior work in Oxford city and consistently outperforms it, including in unstructured settings the earlier system cannot handle.Source: §V · p. 6
VI. CONCLUSION AND FUTURE WORK
The paper presents radar as a reliable sensor for challenging conditions and proposes keypoint extraction and scan-matching algorithms designed to need little prior knowledge or human intervention. It reports high accuracy across diverse settings and identifies future work on candidate proposal, scene symmetry, and 3D information from 2D scans.
VI. CONCLUSION AND FUTURE WORK
- The paper motivates radar as an information-rich, long-range sensor robust under challenging conditions.Source: §VI · p. 6
- Its keypoint extraction and scan-matching algorithms are designed to require minimal a priori knowledge and human intervention.Source: §VI · p. 6
- The pipeline achieves high accuracy across diverse settings, performing comparably to and sometimes better than VO.Source: §VI · p. 6
- Future work includes streamlining unary candidate proposal, addressing scene symmetry, and deducing 3D information from 2D scans.The paper also discusses theoretical underpinnings, proposes two performance measures, and analyzes system complexity.Source: §VI · p. 6