TL;DR
This paper presents a highly robust, radar-only odometry pipeline capable of precise ego-motion estimation in challenging environments, from urban streets to off-road boulder fields. By framing data association as a global graph matching problem and using a novel one-parameter keypoint extraction algorithm, the system effectively ignores severe radar noise and false positives. It achieves real-time performance that rivals or exceeds state-of-the-art visual odometry, without requiring motion priors or complex parameter tuning.
What Problem Does This Paper Address?
Autonomous vehicles typically rely on a standard sensor suite: vision, lidar, and GPS. However, these sensors are fragile in adverse conditions—cameras fail in bad lighting, lidar struggles in heavy rain or snow, and GPS drops out in tunnels or urban canyons.
Radar is a highly versatile alternative that easily penetrates weather and operates in all lighting conditions. However, radar odometry (estimating a vehicle's trajectory using only radar) is notoriously difficult. Radar scans suffer from low resolution, heavy speckle noise, clutter, and false positives caused by multipath reflections. Existing radar odometry methods usually require heavy parameter tuning (like CFAR algorithms), rely on restrictive geometric models, need strong motion priors (like Iterative Closest Point methods), or fall back on other sensors when the environment gets too complex. The authors tackle the challenge of building a purely radar-based odometry system that works anywhere with minimal human intervention.
Methods
The authors propose an indirect (feature-based) odometry pipeline divided into two main novel components:
- One-Parameter Keypoint Extraction: Instead of using traditional Constant False Alarm Rate (CFAR) detectors which require multiple tuned parameters, the authors use a gradient-based approach. They apply a Prewitt operator to the radar scan to find peaks (areas of high intensity and low gradient). The algorithm suppresses redundant returns by keeping only one keypoint per continuous bright region. It requires only a single tunable parameter: the maximum number of keypoints ($\ell_{max}$).
- Data Association via Graph Matching: To match keypoints between consecutive scans, the authors abandon iterative local searches (like ICP) in favor of global graph matching.
- Unary Descriptors: They introduce a new rotation-invariant descriptor for each keypoint based on histograms of surrounding angular slices and annuli, normalized using a Fast Fourier Transform (FFT).
- Pairwise Compatibility: The system checks if the geometric distance and relationship between two points in Scan A match the relationship of their proposed pairs in Scan B.
- Optimization: Finding the perfect match is an NP-hard problem. The authors relax the constraints and solve it by finding the principal eigenvector of the compatibility matrix. They introduce a novel termination condition that stops adding matches the moment a new match decreases the overall mutual compatibility of the set.
Finally, standard rigid-body motion techniques are used to calculate the 3-DOF (planar translation and rotation) movement between the matched scans.
Datasets
The system was evaluated on six diverse datasets collected using a millimeter-wave (MMW) FMCW scanning radar, with GPS serving as the ground truth. The datasets span highly varied environments:
- Urban/Structured (UK): Oxford City, Oxford Backstreets, M40 Highway.
- Unstructured/Off-road: Rural Countryside (UK), Heidmörk Forest (Iceland), and Þórsmörk Rocks/boulder field (Iceland).
Key Insights and Ideas
- Simplicity scales: By reducing keypoint extraction to a single parameter and eliminating parameters entirely in the data association phase, the pipeline generalizes beautifully. The exact same setup works just as well on a high-speed UK highway as it does in an unstructured Icelandic forest.
- Global context beats local search: Using graph matching leverages the global geometric structure of the scene. This makes the system incredibly robust to radar's notorious "ghost" objects and speckle noise, because random noise and false positives rarely maintain consistent global geometry between scans.
- Radar can fly solo: The system proves that scanning radar doesn't just have to be a weather-resistant backup sensor; it can provide primary, highly accurate odometry that rivals vision, even without a motion prior.
Key Results
- High Accuracy: Across all datasets, the Radar Odometry (RO) pipeline achieved a median translation error of 5.20 cm and a rotation error of 0.0929 degrees (compared to GPS ground truth).
- Beating Vision: RO slightly outperformed state-of-the-art Visual Odometry (VO), which had median errors of 5.77 cm and 0.1032 degrees.
- Beating ICP: The graph matching approach vastly outperformed standard Iterative Closest Point (ICP) matching across all environments, even when ICP was given a generous motion prior to help it along.
- Real-Time Performance: The system runs efficiently in real-time at ~4 Hz, matching the physical rotation speed of the scanning radar.
Limitations and Future Work
Limitations: Because the radar provides 2D scans, the system only estimates 3-DOF motion (2D planar translation and rotation). In highly unstructured environments (like the Iceland boulder fields), vehicle rolling, pitching, or elevation changes are forcibly projected into this 2D plane. This "squashing" of 3D motion into a 2D estimate causes a slight increase in error in off-road settings compared to flat city streets.
Future Work: The authors suggest three promising directions:
- Streamlining the unary candidate proposal step to further reduce computational complexity.
- Addressing highly symmetrical scenes (which can confuse the graph matcher by offering multiple valid geometric alignments) using tertiary graph matching or by quantifying motion observability.
- Deducing 3D environmental information from the 2D radar scans to account for pitch and roll.
Summary updated