This repository contains links and instructions for a radar odometry dataset associated with the Do we need scan-matching in radar odometry? paper submitted to the ICRA 2024 conference.
The file format is ROS1 .bag files with all sensory data as recorded during the Forest and Mine experiments. The bag files are split into 5GB parts to make the download and handling easier.
The files are available under this link: Download from ORU.se
The dataset provides sensor measurements from these sensors:
in these topics:
/emlid_gnss/fix (sensor_msgs/NavSatFix)
/emlid_gnss/nmea_sentence (nmea_msgs/Sentence)
/emlid_gnss/post_rtk_fix (sensor_msgs/NavSatFix)
/emlid_gnss/time_reference (sensor_msgs/TimeReference)
/hugin_raf_1/radar_data (sensor_msgs/PointCloud2)
/imu/data (sensor_msgs/Imu)
/imu/mag (geometry_msgs/Vector3Stamped)
/imu/time_ref (sensor_msgs/TimeReference)
/ouster/points (sensor_msgs/PointCloud2)
/tf (tf2_msgs/TFMessage) The /tf topic contains all static transforms between the sensors, except for the IMU, which is located behind the Hugin sensor close to the _baselink frame origin. From space reasons, the IMU is rotated 90 degrees clockwise around its Z axis, such that the _Ximu axis points to the right, the _Yimu axis points forward and the _Zimu axis points up. The IMU internal filter in the IMU is set to the VRU General profile which handles the magnetometer readings as a source of yaw drift suppression, however the heading is not absolutely referenced by it.
Please note that the Hugin A3-Sample radar used in our dataset is an early demo model not with the same performance as the forthcoming production-ready model. It is set to the Short range profile, which assures the highest range resolution for the price of 50m range. It produces between 5,000-10,000 points that include the power and Doppler velocity values. The radar messages are timestamped in the sensor, however they arrive with approx 0.1s lag.
The lidar scans are recorded directly as PointCloud2 messages in the bag files to make the re-use easier. The lidar is PTP-synchronized with the recording computer.
In the Forest experiment, a pair of the Reach RTK-GNSS receivers was used to obtain a positioning reference. The RTK solution was computed offline and inserted into the bag files as the _/emlid_gnss/post_rtkfix topic. The original singe-fix solution is stored in the _/emlidgnss/fix topic. The clock of the recording computer is not synchronized with the GNSS clock, there may be some offset between those two. The ROS message stamps are based on the time of the NMEA message arrival over the serial link.
The ground-truth localization was generated by running the HDL slam with the lidar data, supported by the radar odometry as prior. In fact, the lidar odometry initiated by the radar odometry was drifting in yaw faster than the radar odometry and comlicated the loop closure process of the HDL slam, thus in the "forward" pass, the HDL slam relied on the radar odometry skipping the lidar scan-matching. The subsequent loop closure identification and graph optimization was performed the classical way with lidar point cloud matching. The HDL slam used the distance-based heuristic for the loop closure identification. The output of the HDL slam is a point cloud map, also provided here. In this map, the platform was localized using the NORLAB icp mapper set to localization only, without adding or removing points from the map. Its output is a 6-DOF trajectory, stored here as a csv file. We provide two variants, a csv with a comma as a separator with these fields:
seconds, nanoseconds, x, y, z, qx, qy, qz, qwand a tum-style csv with spaces as separators and fields:
seconds x y z qx qy qz qwThis tum-style format can be directly loaded with the tools of the EVO library. However, keep in mind that by putting seconds and nanoseconds into one number, a few decimal positions in the nanoseconds get rounded out (it does not fit into the double floating-point representation, it has too many digits). The map and ground-truth localization are provided in the experiment directories.
This dataset is a part of a work in progress on a radar SLAM solution for harsh environments. We welcome suggestions regarding this dataset or its documentation in the Issues section of this repository.