The range and fields of view for each sensor are shown below, and the dynamic objects detected by the 4D Radar can be observed.
The Seonsor Configuration
4D FMCW Radar (Continental ARS548)
Spinning Radar (Navtech Ras6)
Solid state FMCW LiDAR (Aeva Aeries II)
Stereo Camera (FLIR Blackfly S BFS-U3-16S2C-CS)
SPAN-CPT7 with VEXXIS GNSS-501 Dual Antenna
Xsens MTi-300 IMU
The sensor coordinate information
The orientation (axis) and distance between the sensors have been ascertained using a CAD model. This information is subsequently utilized as the initial estimation in the extrinsic calibration process.
The Continental radar provides both raw radar point cloud data and filtered object point cloud data via ars548 RDI radar driver.
We provide Navtech Radar data not only in Polar image format but also in Cartesian image format.
The Aeva LiDAR operates with relative velocity settings synchronized to the 4D radar sensor.
All sensor data are processed on an industrial PC, the NUVO-9006LP-NX, equipped with an Intel Core i9 processor, 2 TB SSD, and 64 GB DDR5 RAM.
Data Format
Stereo Camera data
8-bit Bayer pattern RGB image (PNG format)
Spinning Radar data
Polar image and Cartesian image (PNG format)
4D Radar data and 4D Radar Detected Object data
Binary file (structures are shown on the right)
FMCW LiDAR data
Binary file (structures are shown on the right)
Individual Sensor Groundtruth
TXT format with [time, x, y, z, qx, qy, qz, qw]
Calibration parameter
TXT format with rotation matrix and translation vector
Inertial data
CSV format for raw measurement from IMU, INS
Stereo Camera Data
The stereo camera data was acquired at a frequency of 15Hz and saved as a png file. Stereo right and stereo left images are stored in separate folders, each containing 1440x1080, 8-bit Bayer pattern RGB images. The filename of each png file corresponds to the acquisition timestamp, and these timestamps are also recorded in stereo_stamp.csv.
Radar Data
The spinning radar data was acquired at a frequency of 4Hz and saved as a Polar image png file. Subsequently, we converted all the saved polar images into Cartesian images, and both formats are provided in separate folders. The filename of each png file corresponds to the acquisition timestamp, and these timestamps are also recorded in datastamp.csv. The file player releases the topics grounded on this timestamp file.
Radar Data
The 4D FMCW radar data was acquired at a frequency of 20Hz and saved as a bin file. As a distinguishing feature of our dataset, we also provide filtered object point cloud data obtained through the ars548 RDI radar driver. The filename of each binary file corresponds to the acquisition timestamp, and these timestamps are also archived in datastamp.csv. The file player releases the topics grounded on this timestamp file.
LiDAR Data
The FMCW LiDAR was acquired at a frequency of 10Hz and saved as a bin file in the format depicted in the above image. The filename of each binary file corresponds to the acquisition timestamp, and these timestamps are also archived in datastamp.csv. The file player releases the topics grounded on this timestamp file.
IMU Data
The data structure for the IMU is organized in the following order: [timestamp, qx, qy, qz, qw, eul x, eul y, eul z, gyr x, gyr y, gyr z, acc x, acc y, acc z, mag x, mag y, mag z]. This information is stored in the 'sensor_data/xsens_imu.csv' file. Additionally, it is broadcasted as a ROS topic under the name '/imu/data_raw', which publishes data in the format [timestamp, quaternion, angular_velocity, linear_acceleration] at a frequency of 100Hz.
INS Data
The INS data is stored in the 'sensor_data/inspva.csv' file, organized in the following sequence: [timestamp, latitude, longitude, height, north velocity, east velocity, up velocity, roll, pitch, azimuth, status]. This data serves as the foundation for creating the ground truth and was acquired at a frequency of 50Hz.
You can find detailed explanations for each Inertial Solution Status stored in 'sensor_data/inspva.csv' at the following link: https://docs.novatel.com/OEM7/Content/SPAN_Logs/INSATT.htm#InertialSolutionStatus
Individual Sensor Ground Truth in the HeRCULES Dataset
A notable contribution of the HeRCULES dataset is its provision of individual sensor ground truth. Given that every sensor has distinct acquisition timestamp and installation points, the scanning positions differ due to both spatial and temporal variations.
Since place recognition evaluations are predominantly decided by range, the precise positioning of these data points holds paramount importance.
Recognizing this, the HeRCULES Dataset offers individual sensor groundtruth. These ground truths are formulated based on the extrinsic calibration of each sensor and the B-Spline Interpolation, which is tailored to each sensor's unique acquisition timestamp.
Illustrative Example: Ground Truth for River Island
Through the illustration, we can observe the sensor-specific ground truth points for two different scenarios at the same location: one moving quickly in a straight line, and another turning slowly to the right.
Blurring Face & LP
Using EgoBlur: A FasterRCNN-based detector for faces and vehicle license plates.
https://github.com/facebookresearch/EgoBlur/tree/main
If there are any issues with the images, please provide information such as the "sequence" & "image name" and send it to the email below.