Source-linked AI summary
Masking by Moving: Learning Distraction-Free Radar Odometry from Pose Information
Dan Barnes, Rob Weston, Ingmar Posner
TL;DR
Radar odometry must remain reliable despite sensing artefacts, dynamic objects, and adverse conditions that undermine conventional scan matching. The paper learns masked radar embeddings and uses fully differentiable correlative matching to preserve interpretability and estimate uncertainty. The resulting system reduces odometry errors by over 68% while running an order of magnitude faster than prior state of the art.
Problem
Radar supports navigation in adverse conditions but suffers from ghost objects, phase noise, speckle, saturation, and changing scene content that degrade conventional correlation-based matching.
Method
A neural network learns bounded masks and radar feature embeddings for differentiable correlative scan matching, retaining interpretable correlation-based pose estimation.
Results
Over 68% lower odometry errors and an order-of-magnitude speedup are reported against the previous state of the art.
Takeaways & Limitations
The system combines learned radar representations with interpretable scan matching and provides a principled mechanism for pose-uncertainty estimation.
Takeaways & Limitations
Calibrated uncertainty currently requires tuning a pre-trained model rather than learning uncertainty estimation within the training pipeline.
Abstract
from arXiv · showhide
This paper presents an end-to-end radar odometry system which delivers robust, real-time pose estimates based on a learned embedding space free of sensing artefacts and distractor objects. The system deploys a fully differentiable, correlation-based radar matching approach. This provides the same level of interpretability as established scan-matching methods and allows for a principled derivation of uncertainty estimates. The system is trained in a (self-)supervised way using only previously obtained pose information as a training signal. Using 280km of urban driving data, we demonstrate that our approach outperforms the previous state-of-the-art in radar odometry by reducing errors by up 68% whilst running an order of magnitude faster.
1 Introduction
Radar offers robustness to adverse conditions but is difficult to use because noise artefacts and distractors corrupt scans. This work combines learned masking with differentiable correlative matching to produce interpretable, efficient radar odometry that outperforms prior state of the art.
- 1 Introduction: Radar can operate through fog, rain, snow, lens flare, and direct sunlight, but ghost objects, phase noise, speckle, and saturation complicate navigation.Earlier radar approaches commonly relied on simplifying assumptions, hand-designed heuristics, or features borrowed from other modalities.
- 1 Introduction: The system learns artefact- and distraction-free radar embeddings for efficient correlative matching while retaining interpretability and principled uncertainty estimates.A deep network learns the representation used for pose prediction, while correlation-based matching preserves an interpretable estimation structure.
- 1 Introduction: In challenging urban environments, the network trained on over 186,000 examples from 216 km of driving reduces errors by over 68% and runs an order of magnitude faster.Pose ground truth is gathered in a self-supervised manner, automatically optimizing odometry, loop closure, and location constraints without manual labeling.
- 1 Introduction: Figure 1 applies CNN-generated masks before 2D correlation, FFT-based volume construction, and softargmax pose recovery in a fully differentiable pipeline.The differentiable design enables the filter masks to be learned end to end.
2 Related Work
Prior radar odometry typically separates feature extraction, correspondence estimation, and pose optimization, while newer end-to-end methods improve representation learning but can reduce interpretability. The proposed approach combines learned radar embeddings and masking with correlation-based matching to retain both learning capacity and geometric transparency.
- 2 Related Work: Traditional radar odometry uses separate feature extraction, data association, and loss minimization stages, including ICP-based processing [8, 9].Features may come from vision methods such as SIFT and SURF, CFAR filtering, temporal continuity models, or BASD descriptors.
- 2 Related Work: Fourier correlation methods densely search correspondences and are interpretable and more robust to pose initialization noise than ICP, but assume temporally stationary power returns.Dynamic objects and radar artefacts violate this assumption and can degrade correlation-based estimates.
- 2 Related Work: End-to-end visual odometry learns features and point correspondences jointly, but this entangles feature representation and data association, reducing interpretability; learned lidar embeddings maintain conventional scan-matching interpretability.Vision systems also mask image regions containing non-stationary features or distractor objects that could corrupt odometry estimates.
- 2 Related Work: The proposed system learns an artefact- and distraction-free radar embedding while retaining interpretable masks and correlation cost volumes.It combines deep representation learning with correlation-based matching rather than replacing geometric matching with direct pose regression.
3 Deep Correlative Scan Matching with Learnt Feature Embeddings
The system combines learned radar masking with differentiable correlative scan matching to suppress artefacts and distractors while preserving efficient, interpretable pose estimation. It also derives pose uncertainty from the correlation distribution, supporting calibrated confidence estimates for ambiguous scene geometries.
- 3 Deep Correlative Scan Matching with Learnt Feature Embeddings: Radar’s nonstationary power field limits conventional matching because dynamic objects, noise artefacts, and changing occlusion disrupt correspondence.The method addresses this mismatch between correlation assumptions and real radar observations through learned feature embeddings.
- 3 Deep Correlative Scan Matching with Learnt Feature Embeddings: 68% error reduction and order-of-magnitude faster operation result from pose-supervised masking combined with FFT-based correlative matching.The learned representation attenuates moving vehicles and sensor noise while preserving scan-consistent structures such as walls and buildings.
- 3.1 Correlative Scan Matching with Learnt Feature Embeddings: The pose estimator searches a discretized SE(2) grid, scores warped scan pairs by cross-correlation, and applies soft-argmax for sub-grid pose accuracy.Bi-linear interpolation and 2D FFT correlation make the large-area search differentiable and computationally efficient.
- 3.1 Correlative Scan Matching with Learnt Feature Embeddings: The network masks each scan with values in [0, 1], producing a learned representation S = M ⊙ Z that is optimized directly for correlative pose matching.Differentiable pose prediction enables minimizing mean squared error over pose-labelled training pairs and updating the masking network by stochastic gradient descent.
- 3.2 Pose Uncertainty Estimation: The method quantifies uncertainty by interpreting temperature-scaled softmax weights over candidate poses as probabilities and using them to estimate pose means and covariance.This provides a principled mechanism for identifying uncertainty caused by pathological solutions from scene topology.
- 3.2 Pose Uncertainty Estimation: Higher softmax temperature concentrates probability on the highest-correlation pose with lower covariance, whereas lower temperature spreads weight across candidates and yields higher covariance.The training procedure uses a dataset, search region, temperature, learning rate, and initial network parameters while iterating until convergence.
4 Experimental Setup
The experiments evaluate radar odometry on the Oxford Radar RobotCar Dataset using learned masking, pose-supervised training, established baselines, and calibrated uncertainty assessment.
- 4.1 Dataset: The Oxford Radar RobotCar Dataset provides Navtech CTS350-X radar data and ground-truth poses, with traversals split 80% for training and 20% for evaluation.Spatial cross-validation uses splits from different real-world regions.
- 4.2 Network Architecture and Training: The system uses a U-Net-style masking network trained with Adam until validation loss is minimized, with β selected by grid search and set to 1.Experiments cover Cartesian and polar representations, single and dual configurations, and multiple spatial resolutions.
- 4.3 Evaluation Metrics and Baselines: Evaluation compares the method with the state-of-the-art radar odometry baseline, raw-scan matching, pose-supervised alternatives, static-mask supervision, visual odometry, and UnDeepVO [18].The primary radar baseline extracts point features, performs global shape-similarity scan matching, and refines the result by mean squared error.
- 4.3 Evaluation Metrics and Baselines: Performance is measured using KITTI-style residual translational and angular errors over offsets up to 800m, reported as interquartile ranges rather than standard deviations.Timing averages 1000 predictions on a specified CPU and GPU platform.
- 4.4 Uncertainty Evaluation: Uncertainty calibration uses the Mahalanobis error, which should follow a chi-squared distribution with three degrees of freedom when pose predictions are Gaussian.Mean values far below or above three indicate conservative or overconfident uncertainty estimates, respectively; β is tuned to improve calibration.
5 Results
The system combines learned radar feature embeddings with differentiable correlative scan matching to improve odometry accuracy, speed, interpretability, and uncertainty estimation. Across odometry and uncertainty experiments, it outperforms the benchmark and produces calibrated uncertainty that identifies difficult cases.
- 5 Results: The embedding masks noise and distractors while preserving temporally consistent features, and uncertainty rises for weakly constrained scenes or motion blur.Pathological solutions from missing x-direction constraints produce high covariance, whereas stationary well-constrained scenes produce low covariance.
- 5.1 Odometry Performance: 68% lower translation and rotation errors with the best model, while running over 4 times faster than the current state of the art.The fastest model exceeds 100Hz while reducing translational error by 28% and rotational error by 20%.
- 5.1 Odometry Performance: Cartesian inputs typically outperform Polar inputs, and dual masking configurations typically outperform single-observation inputs.The experiments cover both input representations and single or dual configurations across multiple resolutions.
- 5.1 Odometry Performance: The learned radar embedding reduces errors by over 85% versus correlative scan matching on raw radar power returns.Pose-optimized masking performs better than direct stationary-scene mask supervision, while directly adapting deep odometry performs significantly worse.
- 5.2 Uncertainty Prediction: Temperature tuning yields calibrated uncertainty, with β = 2.789 producing a mean Mahalanobis distance of d̄2 = 2.99 and most errors inside predicted bounds.Temperatures above and below this value are overly certain and conservative, respectively.
6 Conclusions
The paper concludes that learned radar embeddings combined with correlative scan matching improve odometry accuracy and speed while preserving interpretability and enabling principled uncertainty estimation. Calibrated uncertainty remains dependent on tuning a pre-trained model.
- 6 Conclusions: The learned embedding and correlative scan matching reduce odometry errors by over 68%, run an order of magnitude faster, and remain interpretable.The method also provides a principled mechanism for estimating pose uncertainty.
- 6 Conclusions: Calibrated uncertainty currently requires tuning a pre-trained model rather than learning uncertainty jointly during training.The authors leave incorporation of this tuning process into the training pipeline for future work.
A.1 Masking Network Architecture
The masking network supports Cartesian and Polar radar inputs, with experiments comparing single- and dual-mask configurations. Cost-volume resolution exposes a trade-off between pose accuracy and runtime.
- A.1 Masking Network Architecture: The masking architecture outputs masks constrained to [0, 1] by a final single-channel sigmoid layer and supports Cartesian or Polar radar representations.
- A.1 Masking Network Architecture: The dual configuration concatenates sequential radar observations and produces two masks instead of one.
- A.2 Speed vs Accuracy Trade Off: Reducing grid resolution shortens correlation-based pose prediction time but decreases pose accuracy for the same coverage.
- A.2 Speed vs Accuracy Trade Off: Figure 5 varies translational error, angular error, and runtime with cost-volume resolution in degrees and metres per pixel.
B Data
The paper compares direct pose-supervised masking with masks generated from temporally static occupied cells. It also illustrates automatically generated baseline masks that identify structure useful for odometry.
- B.1 Baseline Masks: Directly training masks for pose estimation improves performance over supervising the proxy task of predicting temporally static occupied cells.
- B.1 Baseline Masks: Baseline masks are automatically generated from raw radar scans to identify structure useful for pose estimation.
B.2 Dataset Splits
The evaluation uses temporally separated training and testing traversals and additionally tests spatial generalisation with cross-validation across traversal-loop splits. The spatially held-out model remains more accurate than the benchmark and substantially faster, despite reduced training-data diversity.
- B.2 Dataset Splits: The dataset includes 25 training and 7 evaluation traversals covering varied traffic and challenging urban conditions, with an additional spatial cross-validation split.
- C.1 Spatial Cross Validation: Spatial cross-validation trains on two of three traversal-loop splits and evaluates on the held-out split, averaging results across splits.
- C.1 Spatial Cross Validation: The best spatial-cross-validation model reduces benchmark translational errors by over 25% and rotational errors by 11% while running over 15x faster.
- C.1 Spatial Cross Validation: Using the medium-resolution model reduces effective training-data diversity by a third and may explain the slight performance reduction relative to the main results.
C.2 Additional Evaluation Examples
Additional qualitative examples show that the learned masks suppress radar noise and distractors while preserving temporally consistent features used for pose prediction.
- The predicted masks filter noise and distractor objects while preserving temporally consistent features such as walls for pose prediction.