Source-linked AI summary
The Oxford Radar RobotCar Dataset: A Radar Extension to the Oxford RobotCar Dataset
Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, Ingmar Posner
TL;DR
Autonomous-vehicle research needs sensing that remains useful in conditions challenging for vision and LIDAR. This paper releases a large Oxford multimodal dataset centered on FMCW radar, with calibrated data, tools, and optimized radar odometry collected across varied urban drives. The collection comprises 32 traversals over 280 km of urban driving.
Problem
The central gap is limited exploitation of FMCW radar for vehicle autonomy and mobile robotics despite its robustness in rain, snow, dust, fog, and direct sunlight.
Method
The paper constructs and releases a multimodal Oxford driving dataset centered on a Navtech CTS350-X radar, with complementary cameras, LIDARs, GPS/INS, calibrations, and software tools.
Results
32 traversals cover 280 km of urban driving across varied traffic, weather, and lighting conditions.
Takeaways & Limitations
The release provides a shared test bed for extending radar research in mapping, localisation, motion estimation, and scene understanding.
Takeaways & Limitations
The radar configuration has no Doppler information, and future work proposes collecting Doppler data for semantic scene understanding.
Abstract
from arXiv · showhide
In this paper we present The Oxford Radar RobotCar Dataset, a new dataset for researching scene understanding using Millimetre-Wave FMCW scanning radar data. The target application is autonomous vehicles where this modality is robust to environmental conditions such as fog, rain, snow, or lens flare, which typically challenge other sensor modalities such as vision and LIDAR. The data were gathered in January 2019 over thirty-two traversals of a central Oxford route spanning a total of 280km of urban driving. It encompasses a variety of weather, traffic, and lighting conditions. This 4.7TB dataset consists of over 240,000 scans from a Navtech CTS350-X radar and 2.4 million scans from two Velodyne HDL-32E 3D LIDARs; along with six cameras, two 2D LIDARs, and a GPS/INS receiver. In addition we release ground truth optimised radar odometry to provide an additional impetus to research in this domain. The full dataset is available for download at: ori.ox.ac.uk/datasets/radar-robotcar-dataset
I. INTRODUCTION
The paper introduces a large-scale Oxford radar dataset to expand autonomous-driving research beyond vision and LIDAR. It targets robust sensing across challenging conditions and includes complementary data and tools for mapping, localisation, and scene understanding.
- Motivation: Radar offers 360° sensing, long detection ranges, and robustness to rain, snow, dust, fog, and direct sunlight.These properties are presented as valuable for higher-speed driving and feature-poor environments.
- Dataset contribution: The release extends the Oxford RobotCar Dataset with radar alongside the cameras, LIDAR, inertial, and GPS data used in autonomous-driving research.The original dataset covered more than 100 traversals of a 10 km Oxford route over a year.
- Dataset contribution: The dataset provides raw recordings, updated calibrations, a radar ground-truth trajectory, and MATLAB and Python development tools.The tools support access to and manipulation of the newly provided sensor formats.
- Research aim: The release is intended to accelerate research into FMCW radar for vehicle autonomy and mobile robotics.It is positioned as a resource for extending competencies developed with more established sensing modalities.
III. THE RADAR ROBOTCAR PLATFORM
The Radar RobotCar platform equips an autonomous-capable Nissan LEAF with a central millimetre-wave radar and complementary 3D, 2D, camera, and navigation sensors. The configuration emphasizes synchronized, calibrated multimodal observations for urban scene understanding.
- Sensor suite: The platform adds one central Navtech CTS350-X FMCW radar and two Velodyne HDL-32E 3D LIDARs to the Oxford RobotCar sensor suite.The Velodyne pair replaced the earlier 3D LIDAR to improve 3D scene understanding and provide full coverage around the vehicle.
- Sensor specifications: The radar operates at 4 Hz with 400 measurements per rotation, 163 m range, 4.38 cm range resolution, and 1.8° beamwidth.
- Sensor specifications: The two Velodyne HDL-32E sensors provide 360° horizontal and 41.3° vertical fields of view at 20 Hz, with 100 m range and 32 planes.
- Sensor suite: The vehicle also carries stereo and monocular cameras, two SICK 2D LIDARs, and a NovAtel inertial and GPS navigation system.
- Data acquisition: Internal sensor drivers provide synchronization and timestamping across the Navtech radar, Velodyne LIDARs, and other sensors.
- Platform layout: Figure 2 places the radar at the vehicle centre and defines sensor frames using x-forward, y-right, and z-down axes.The development tools include exact SE(3) extrinsic calibrations, although the illustrated measurements are approximate.
IV. RADAR DATA
The Navtech CTS350-X supplies high-resolution, short-range FMCW radar scans represented as polar power measurements and convertible to Cartesian form. Its configuration is tailored to urban scenes and emphasizes operation in harsh conditions.
- Radar configuration: The radar records 3768 power readings across 400 azimuths at 4 Hz, with 4.38 cm range resolution, 163 m maximum range, and 0.9° azimuth resolution.The sensor configuration excludes Doppler information.
- Data representation: Figure 3 shows raw polar radar power over 0→2π and 0→163 m alongside the corresponding Cartesian scan centered on the vehicle.The supplied SDK performs parsing and polar-to-Cartesian conversion.
- Radar configuration: The dataset favors shorter-range, high-resolution radar because straight-line distances beyond 163 m are rare in urban scenarios.
- Data representation: Each full rotation forms a 2D power matrix whose rows represent azimuths and columns represent range bins.Intensity records the highest power reflection within each range bin.
- Operating conditions: The radar operates from 76 to 77 GHz and is designed for consistent measurements through dust, rain, and snow.Its beam includes a cosec-squared fill-in pattern extending up to 40° below horizontal for detecting objects beneath the main beam.
V. DATA COLLECTION
The dataset comprises 32 manually driven Oxford traversals totaling 280 km in varied traffic, weather, and lighting conditions, with a 4.7 TB download. Route deviations and variable fused INS accuracy motivate use of optimized radar odometry.
- Collection scope: 32 traversals covered 280 km of urban driving in January 2019 under varied traffic, weather, and lighting conditions.The vehicle was driven manually throughout collection, without autonomous capabilities.
- Collection scope: The complete dataset has a total download size of 4.7 TB.Summary statistics are provided for both collected raw data and processed data.
- Collection caveats: Some traversals include infrequent slight route diversions, and two partial traversals do not cover the entire route.GPS/INS data can identify these deviations.
- Collection caveats: The fused INS solution varied significantly during data collection, so optimized radar odometry is suggested as the best available estimate of radar motion.
A. Sensor Calibration
The dataset provides extrinsic calibration for the added radar and Velodyne sensors, with estimates intended as starting points for further cross-modality calibration research.
- A. Sensor Calibration: Extrinsic calibration data are included for the additional Navtech radar and Velodyne sensors.The original sensors’ intrinsics and extrinsics remain unchanged.
- A. Sensor Calibration: Calibration was seeded by manual measurements and refined through pose optimisation using laser–radar co-observations.
- A. Sensor Calibration: The calibration estimates are recommended as initial seeds for further cross-modality calibration research.
B. Data Formats
The dataset is organised by traversal archives and stores radar scans as structured polar PNGs with embedded per-azimuth metadata, including timestamps, sweep counters, and validity flags.
- B. Data Formats: Each traversal archive contains complete sensor data or processed outputs in the directory structure shown for the dataset.Unlike the original Oxford RobotCar Dataset, sensor data are not chunked into smaller files.
- 1) Radar scans:: Radar scans are stored as lossless-compressed PNGs in polar form, with rows for azimuths and columns for range-bin power returns.
- 1) Radar scans:: Each radar file is addressed as <dataset>/radar/<timestamp>.png, with the timestamp denoting capture start time in microseconds.
- 1) Radar scans:: 400 azimuth rows and 3768 range-bin columns represent each radar sweep in the released configuration.
- 1) Radar scans:: Per-azimuth metadata occupy the first 11 PNG columns, including an int64 UNIX timestamp in columns 1–8.
- 1) Radar scans:: The sweep counter is stored as a uint16 in columns 9–10 and converted to an angle in radians using the supplied equation.
- 1) Radar scans:: A uint8 validity flag in column 11 marks infrequently dropped azimuth returns, while adjacent returns are interpolated to provide 400 rows per scan.Users who do not want interpolation can discard rows whose validity flag is zero.
2) 3D Velodyne LIDAR scans:
The two Velodyne sensors are released in raw and binary formats: raw lossless PNGs preserve per-azimuth measurements and metadata, while binary files provide non-motion-compensated point clouds.
- 2) 3D Velodyne LIDAR scans:: Velodyne scans are provided as raw sensor data or binary non-motion-compensated point clouds.
- 2) 3D Velodyne LIDAR scans:: Raw scans are lossless PNGs whose columns represent sensor readings at each azimuth, stored under sensor-specific timestamped paths.
- 2) 3D Velodyne LIDAR scans:: Raw PNG rows 1–32 store per-laser intensities as uint8 values.
- 2) 3D Velodyne LIDAR scans:: Raw PNG rows 33–96 store per-laser ranges as uint16 values, with conversion to metres defined by the dataset format.
- 2) 3D Velodyne LIDAR scans:: Raw PNG rows 97–98 store the sweep counter as uint16 values used to derive azimuth angle.
- 2) 3D Velodyne LIDAR scans:: Rows 99–106 contain approximate int64 UNIX timestamps obtained by linearly interpolating packet timestamps across azimuth readings.The original received timestamps can be recovered by taking every twelfth timestamp.
- 2) 3D Velodyne LIDAR scans:: Binary scans contain single-precision (x, y, z, I) × N values for 3D coordinates and measured intensity relative to the sensor.
3) Ground Truth Radar Odometry:
The release includes temporally aligned ground-truth SE(2) radar odometry, with trajectories optimised from visual odometry, visual loop closures, and GPS/INS constraints.
- 3) Ground Truth Radar Odometry:: The radar odometry file records SE(2) relative poses between source and destination radar frames using UNIX timestamps and a six-vector Euler parameterisation.The representation retains zero z, α, and β components for compatibility with other pose sources.
- 3) Ground Truth Radar Odometry:: Optimised radar odometry is plotted for all 32 traversals, with each trajectory offset for visualisation.
- 3) Ground Truth Radar Odometry:: The trajectories were generated by optimising robust visual odometry, visual loop closures, and GPS/INS constraints.
VI. GROUND TRUTH RADAR ODOMETRY
The dataset provides temporally aligned ground truth SE(2) radar odometry to support radar motion estimation, mapping, and localisation. These poses are produced through large-scale optimisation using visual odometry, loop closures, and GPS/INS constraints.
- Ground truth SE(2) radar odometry is temporally aligned to radar data for motion estimation, map building, and localisation.
- The pose estimates are generated by large-scale optimisation with Ceres Solver, incorporating visual odometry, visual loop closures, and GPS/INS constraints.
- All 32 dataset traversals are included, with VO, GPS/INS, and loop closures jointly optimised in the radar frame for approximately accurate global SE(2) poses.
- The SDK provides MATLAB and Python tools for accessing and manipulating the released sensor data formats.
A. Radar Loading and Conversion to Cartesian
The development tools load radar and Velodyne scans and convert their native representations into Cartesian forms. They support both interactive playback and direct access to decoded sensor quantities.
- Radar Loading: Radar loading functions return timestamps, azimuth angles, power returns, and range resolution from raw scans.For this release, radar resolution is fixed at 4.38 cm.
- Radar Conversion: Radar polar-to-Cartesian functions convert decoded azimuth, power, and range-resolution data using a chosen Cartesian resolution and image size.
- Radar Playback: Radar playback scripts animate scans while performing polar-to-Cartesian conversion.
- Velodyne Loading: Raw Velodyne loading functions return ranges, intensities, azimuth angles, and approximate timestamps from timestamped scans.
- Velodyne Conversion: Raw Velodyne conversion functions produce Cartesian pointclouds with per-point intensity values from decoded ranges, intensities, and azimuth angles.
- Velodyne Binary Data: Binary Velodyne loading functions directly return Cartesian pointclouds with per-point intensity values.
- Velodyne Playback: Velodyne playback scripts animate available scans from a dataset directory.
VIII. SUMMARY AND FUTURE WORK
The paper presents a large-scale dataset for millimetre-wave FMCW radar research in vehicle autonomy and mobile robotics. Future work targets broader collection conditions, improved calibration, and radar semantic scene understanding.
- The Oxford Radar RobotCar Dataset focuses on exploiting millimetre-wave FMCW scanning radar for large-scale and long-term vehicle autonomy and mobile robotics.
- The authors anticipate that releasing the dataset will foster discussion and enable research areas not previously possible for this relatively little-studied modality.
- Future Work: Future data collection will target new and challenging conditions, while extrinsic calibration parameters may be fine-tuned more precisely.
- Future Work: Future work will investigate semantic scene understanding in radar, potentially using additionally collected Doppler data.