[Paper Review] MAROAM: Map-based Radar SLAM through Two-step Feature Selection

Dequan Wang, Yifan Duan+4 more|arXiv (Cornell University)|Oct 25, 2022

Robotics and Sensor-Based Localization4 citations

TL;DR

MAROAM is a map-based radar SLAM framework that uses a two-step feature selection strategy—first applying geometry-based filtering to extract surface features, then employing probabilistic filtering to retain only frequently detected features—resulting in a robust, globally consistent map. It achieves state-of-the-art performance, outperforming existing methods by 7.95%, 37.0%, and 8.9% on the Oxford Radar RobotCar, MulRan, and Boreas datasets, respectively, with a 28.6% improvement over scan-to-frames odometry.

ABSTRACT

In this letter, we propose MAROAM, a millimeter wave radar-based SLAM framework, which employs a two-step feature selection process to build the global consistent map. Specifically, we first extract feature points from raw data based on their local geometric properties to filter out those points that violate the principle of millimeter-wave radar imaging. Then, we further employ another round of probabilistic feature selection by examining how often and how recent the feature point has been detected in the proceeding frames. With such a two-step feature selection, we establish a global consistent map for accurate and robust pose estimation as well as other downstream tasks. At last, we perform loop closure and graph optimization in the back-end, further reducing the accumulated drift error. We evaluate the performance of MAROAM on the three datasets: the Oxford Radar RobotCar Dataset, the MulRan Dataset and the Boreas Dataset. We consider a variety of experimental settings with different scenery, weather, and road conditions. The experimental results show that the accuracy of MAROAM is 7.95%, 37.0% and 8.9% higher than the currently best-performing algorithms on these three datasets, respectively. The ablation results also show that our map-based odometry performs 28.6% better than the commonly used scan-to-frames method. Finally, as devoted contributors to the open-source community, we will open source the algorithm after the paper is accepted.

Motivation & Objective

• To address the challenge of low-quality feature points in millimeter-wave radar SLAM, which degrade map consistency and pose estimation accuracy.
• To improve map quality and reduce drift in radar-only SLAM by introducing a two-stage feature selection mechanism.
• To enable reliable scan-to-map matching in ICP-based radar SLAM using only high-frequency, stable features.
• To evaluate the framework across diverse real-world conditions, including varying weather, lighting, and road types.
• To open-source the algorithm post-acceptance as a contribution to the open-source robotics community.

Proposed method

• The system first applies a geometry-based filter using local linearity and local aggregation to extract surface features from raw radar point clouds, discarding points inconsistent with radar imaging principles.
• A point-to-line ICP algorithm uses the selected surface features to estimate relative pose between consecutive scans, improving registration accuracy.
• A probability-based filter dynamically removes features that are infrequently detected across frames, retaining only stable, repeatable features for map building.
• The framework uses Scan Context for loop closure detection, adapted to radar data format, to correct accumulated drift.
• Graph optimization is applied in the back-end to refine the trajectory and map, enhancing global consistency.
• The map is incrementally updated using only the most frequently detected and geometrically consistent features, ensuring long-term robustness.

Experimental results

Research questions

• RQ1 How can feature selection improve map quality and reduce drift in radar-only SLAM systems?
• RQ2 Can a two-step feature selection strategy—combining geometric and probabilistic filtering—enhance the robustness of ICP-based scan-to-map matching in radar SLAM?
• RQ3 How does the proposed map-based odometry compare to traditional scan-to-frames methods in terms of accuracy and drift reduction?
• RQ4 To what extent does the framework maintain performance across diverse environments, including adverse weather and low-texture scenes?
• RQ5 What is the contribution of loop closure and graph optimization to the final trajectory accuracy in radar SLAM?

Key findings

• MAROAM achieves a 7.95% higher accuracy than the best existing method on the Oxford Radar RobotCar Dataset.
• On the MulRan dataset, MAROAM reduces drift error by 37.0% compared to the SOTA algorithm Hong SLAM v2.
• On the Boreas dataset, MAROAM reduces translation error by 8.9% and rotation error by 19.6% compared to HERO.
• The ablation study shows that map-based odometry improves accuracy by 28.6% compared to the scan-to-frames method.
• Removing the probability filter increases odometry error from 1.80% to 3.16% (translation), and disabling the geometry filter raises it to 8.00%, confirming the importance of both filters.
• Loop closure reduces the trajectory error from 1.80% to 1.63% (translation), demonstrating its critical role in drift correction.