[Paper Review] Masking by Moving: Learning Distraction-Free Radar Odometry from Pose Information

Dan Barnes, Rob Weston+1 more|arXiv (Cornell University)|Sep 9, 2019

Robotics and Sensor-Based Localization51 citations

TL;DR

The paper presents an end-to-end radar odometry system that uses a learned artefact-masked embedding with differentiable correlation-based matching to achieve robust, fast pose estimates and uncertainty quantification, trained self-supervised from pose data.

ABSTRACT

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.

Motivation & Objective

• Motivate robust ego-motion estimation in challenging weather and clutter where radar is advantageous over vision/LiDAR.
• Develop an end-to-end, differentiable radar odometry pipeline that preserves interpretability via correlation-based matching.
• Learn radar feature embeddings that suppress artefacts and dynamic objects while retaining stationary scene structure.
• Enable principled uncertainty estimation for pose predictions to support robust deployment.

Proposed method

• Use a fully differentiable, correlation-based scan matching over a learned radar embedding S = M ⊙ Z, where M = fα(Z) is a mask produced by a CNN (U-Net).
• Train α by minimizing the mean squared error between predicted pose p̂ and ground-truth pose p on consecutive radar scans (self-supervised via pose information).
• Compute 2D cross-correlation between masked scans using FFT for efficient exhaustive pose search and apply soft-argmax to obtain sub-grid pose estimates.
• Quantify pose uncertainty by interpreting softmax weights over candidate poses as probabilities and deriving a Gaussian-like covariance Σ from the weighted pose distribution.
• Provide uncertainty calibration by tuning a temperature parameter β in the softmax so that the mean Mahalanobis distance aligns with expected chi-squared distribution (calibrated uncertainty).
• Optionally compare Cartesian vs Polar radar inputs and single vs dual masking configurations to study performance versus interpretability.

Experimental results

Research questions

• RQ1 Can a learned masking representation improve the robustness of radar odometry against clutter and dynamic objects while preserving interpretable, correlative scan matching?
• RQ2 Does end-to-end training with pose supervision yield substantial gains over masking methods trained on proxy tasks or raw radar data for odometry accuracy and speed?
• RQ3 Can we obtain calibrated, per-pose uncertainty estimates in a radar odometry system using a differentiable softmax-based weighting over pose candidates?
• RQ4 What are the comparative benefits of Cartesian versus Polar radar representations and single versus dual input masking for pose estimation and computation speed?

Key findings

• The proposed method reduces odometry translation and rotation errors by up to 68% relative to the prior state of the art on urban radar data.
• The system runs an order of magnitude faster than the previous best radar odometry approach, achieving high-frequency operation (over 100 Hz in some configurations).
• Learning a radar feature embedding optimized for pose prediction substantially improves accuracy, outperforming correlative matching on raw scans by a large margin (over 85% relative reduction in error).
• Masking learned from pose supervision yields better performance than masking trained via a proxy task (static occupancy labelling).
• The approach provides principled, calibrated uncertainty estimates for pose predictions through the softmax-weights-based covariance, enabling detection of pathological cases and improved reliability.