[Paper Review] Lidar-level localization with radar? The CFEAR approach to accurate, fast and robust large-scale radar odometry in diverse environments

Daniel Adolfsson, Martin Magnusson+3 more|arXiv (Cornell University)|Nov 4, 2022

Robotics and Sensor-Based Localization4 citations

TL;DR

This paper presents CFEAR, a learning-free, real-time radar odometry method that achieves lidar-level localization accuracy using spinning radar by combining motion-compensated, one-to-many scan registration with robust loss functions and oriented surface point estimation. It reduces drift by 38% compared to prior radar odometry, achieving 1.09% translation error at 5 Hz on the Oxford Radar RobotCar benchmark—surpassing state-of-the-art radar SLAM and approaching lidar SLAM performance without loop closure.

ABSTRACT

This paper presents an accurate, highly efficient, and learning-free method for large-scale odometry estimation using spinning radar, empirically found to generalize well across very diverse environments -- outdoors, from urban to woodland, and indoors in warehouses and mines - without changing parameters. Our method integrates motion compensation within a sweep with one-to-many scan registration that minimizes distances between nearby oriented surface points and mitigates outliers with a robust loss function. Extending our previous approach CFEAR, we present an in-depth investigation on a wider range of data sets, quantifying the importance of filtering, resolution, registration cost and loss functions, keyframe history, and motion compensation. We present a new solving strategy and configuration that overcomes previous issues with sparsity and bias, and improves our state-of-the-art by 38%, thus, surprisingly, outperforming radar SLAM and approaching lidar SLAM. The most accurate configuration achieves 1.09% error at 5Hz on the Oxford benchmark, and the fastest achieves 1.79% error at 160Hz.

Motivation & Objective

• To develop a highly accurate, fast, and robust radar odometry system that generalizes across diverse real-world environments without parameter tuning.
• To overcome challenges in radar odometry such as sparsity, noise, and bias by integrating motion compensation and multi-sweep scan registration.
• To achieve performance comparable to lidar SLAM using only radar, without learning or loop closure.
• To systematically evaluate and quantify the impact of key components like filtering, resolution, loss functions, and keyframe history on odometry accuracy.

Proposed method

• The method uses a conservative filtering strategy to extract oriented surface points from raw radar scans, improving feature quality and reducing noise.
• It performs one-to-many scan registration by aligning the latest scan to multiple prior keyframes simultaneously, minimizing point-to-line distances.
• A robust loss function is applied to downweight outliers and mitigate the impact of speckle noise and ghost objects.
• Motion compensation is integrated within each sweep to correct for platform motion during data acquisition, improving alignment accuracy.
• The system employs intensity-weighted surface point estimation and residual weighting to enhance robustness to low-quality features.
• A new solving strategy accelerates computation, enabling real-time performance at up to 160 Hz on a standard CPU.

Experimental results

Research questions

• RQ1 How does the integration of motion compensation within each radar sweep affect odometry accuracy and robustness in dynamic environments?
• RQ2 What is the impact of keyframe history and multi-scan registration on reducing drift and improving pose estimation in sparse radar data?
• RQ3 How do filtering, resolution, loss functions, and residual weighting collectively influence the performance of radar odometry?
• RQ4 Can a learning-free radar odometry pipeline generalize across diverse environments—outdoor urban, woodland, and indoor warehouses and mines—without parameter changes?
• RQ5 To what extent can radar odometry approach lidar SLAM performance in terms of localization accuracy without loop closure or refinement?

Key findings

• The most accurate CFEAR configuration achieves 1.09% translation error at 5 Hz on the Oxford Radar RobotCar dataset, representing a 38% improvement over the previous state-of-the-art in radar odometry.
• The fastest configuration runs at 160 Hz with 1.79% translation error, demonstrating real-time feasibility on standard hardware.
• The method outperforms the current state-of-the-art radar SLAM method by 40.4% in drift reduction on the Oxford dataset and by 26.7% on the MulRan dataset.
• The system generalizes robustly across diverse environments—including urban, woodland, warehouses, and underground mines—without parameter tuning.
• One-to-many scan registration with robust loss functions and motion compensation significantly reduces bias and noise, especially in sparse or noisy conditions.
• The ablation study confirms that intensity-weighted surface points, residual weighting, and motion compensation are critical for minimizing drift and improving accuracy.