VIDETEC-2 - A Multi-Modal UWB-Radar-Camera Dataset for Vulnerable-Road-User Sensing at Urban Intersections

Dataset purpose and scope This dataset provides synchronized multi-modal measurement data for vulnerable road user (VRU) sensing at an urban intersection in Garching-Hochbrück near Munich, Germany. It was created to support research on infrastructure-based perception, RF sensing, radar-based detection, multimodal fusion, and UWB-based joint communication and sensing in realistic traffic conditions. Location in Google Maps here. Measurement setup and sensor modalities The dataset combines a dense static mesh of 15 UWB transceivers, two infrastructure-mounted FMCW radars, a network of calibrated HD cameras, and RTK-GNSS reference data from a test VRU. Recordings were collected for approx. one hour on the 8 and 9 October 2025, with the test VRU moving through the intersection as a pedestrian or cyclist while other traffic participants were present in the scene. Date Start Time (local CEST) Test VRU type 08.10.2025 14:57 Pedestrian 08.10.2025 15:23 Cyclist 08.10.2025 15:50 Pedestrian 09.10.2025 14:44 Pedestrian 09.10.2025 15:27 Cyclist Hint: Germany has CEST: Local time = UTC+2 Dataset layout The dataset is organized by sensor subsystem. The data is compressed in either zip or tar.gz files. UWB The recorded UWB measurements including the channel impulse responses (CIR) can be found in uwb.zip. The measurements for each of the 16 nodes can be found inside sub-folders Measurements[NODE_ID] in the CIR-txt folder. Nodes 1 to 15 are static mounted at a height of about 1.5 m. Node 16 is carried by the VRU. The text files with the mesaurements inside the Measurements[NODE_ID] sub-folders are divided in runs following this nameing convention: UWBMeas[MEAS_RUN]_Anchor[NODE_ID]_Mod0_[YY-MM-DD]-[HH-mm-ss].txt Five measurement runs were recorded on the 8th of October 2025 and five on the 9th of October 2025. Along the CIR_txt folder a text file uwb_nodes.txt contains the 2D position of the 15 static UWB nodes deployed on the northern part of the intersection. For each node the position in decimal degrees latitude and longitude and in UTM easting/northing coordinates in meters are given. FMCW RADAR Two stationary radar sensors were mounted on a gantry above the intersection and recorded range–Doppler maps and point detections at 10 Hz during daytime, cloudy/misty conditions. The infrastructure-based automotive radar dataset collected at a road intersection, focusing on vulnerable road users (VRUs) such as pedestrians and cyclists, together with regular vehicle traffic can be founf in Radar_dataset.zip. The compressed file contains: 2 × HDF5 files, one per radar sensor (gantry-mounted radars, IDs 51 and 52). Each file contains: Time-ordered frames (timestamps, range–Doppler maps, sensor ID). A flat table of detections (range, Doppler, azimuth, elevation, magnitude). Static sensor metadata (pose on the gantry, coordinate frame). Radar configuration parameters (resolutions, frequency, etc.). Each sensor file is approximately 900 MB. Note: The dataset is not manually annotated. Labels can be obtained using the external ground-truth system. HDF5 File Structure Each file shares the same internal structure: Root group / Attributes descriptionDescription of the dataset ("Infrastructure radar frames with range–Doppler maps and detections"). rd_map_layoutText description of how the range–Doppler map indices map to range and Doppler bins. Groups / datasets /frames – per-frame data /detections – all point detections (single table) /sensor – static sensor mounting metadata /radar_params – radar configuration parameters /frames (per-frame data) timestamp 1D dataset, shape (N,), type int64. Reception time of each radar frame in the acquisition system. sensor_id 1D dataset, shape (N,), type int32. Identifier of the radar sensor (e.g., 51 or 52). rd_map 3D dataset, shape (N, 256, 128) and (N, 128, 256), corresponding to sensors 51 and 52, respectively. type uint16. rd_map[i, :, :] is the range–Doppler map for frame i: First dimension: range bins. Second dimension: Doppler bins. Values encode radar return intensity. A frame is obtained by indexing timestamp[i], sensor_id[i], and rd_map[i, :, :]. /detections (all point-cloud detections) 1D dataset, shape (M,), compound type with fields: frame_index (int32)Index into /frames/timestamp and /frames/rd_map (0-based). timestamp (int64)Same time as the corresponding frame. range (int32)Range bin index (scaling defined by /radar_params /range_resolution_m). doppler (int32)Doppler bin index (scaling defined by /radar_params /doppler_resolution_mps). azimuth (int32)Azimuth angle in degrees. elevation (int32)Elevation angle in degrees. magnitude (int32)Detection intensity amplitude. radar_id (int32)Radar sensor identifier. Frames may have from 0 up to roughly 40 detections.Detections belonging to frame i are all rows with frame_index == i. /sensor (static sensor mounting) Attributes position_xyz_m[x, y, z] position of the radar on the gantry in meters (in the local coordinates of the test field). mounting_height_mVertical height above the road surface (meters). orientation_yaw_deg, orientation_pitch_deg, orientation_roll_degStatic orientation of the sensor in degrees. coordinate_frameText description of the coordinate convention (e.g., "x-east, y-north, z-altitude"). /radar_params (radar configuration) Attributes (sensor-native quantities; detailed meaning in the paper, Section III-B) range_resolution_mRange resolution per bin (meters). doppler_resolution_mpsDoppler (radial velocity) resolution per bin (m/s). max_range_mMaximum unambiguous range (meters). max_doppler_mpsMaximum unambiguous radial velocity (m/s). num_range_binsNumber of range bins. num_doppler_binsNumber of Doppler bins. Using the Data A Jupyter Notebook of minimal examples for Python radar_dataset_examples.ipynb is included in Radar_dataset.zip along with the radars' data. Requirements: pip install h5py numpy matplotlibFurther stand-alone radar data, inlcuding three additional sensors on the south-side of the intersection, can be found at VIDETEC-2 | SENTIRE RADAR. Along this data comes the SenTool, a useful tool to visualize the raw radar data. CAMERA The multi-camera recordings collected on two separate days are provided as two compressed archives: image_data_Oct_08.tar.zip and image_data_Oct_09.tar.zip. Each archive, once extracted, includes the following components: image_data_downscaled: This directory contains all recorded image data from eight cameras, organized according to bridge location and camera viewing direction. Specifically, the dataset includes two bridges (s110 and s120), and the suffixes o, w, s, and n indicate the east, west, south, and north viewing directions, respectively. For instance, s110_o_cam_8.tar.gz corresponds to images captured by the east-facing camera installed on bridge s110. runs_vru: This directory includes multiple scenario segments of approximately five minutes each (9 segments for Oct 08 and 6 segments for Oct 09).Each segment contains compressed files for all eight cameras. After extraction, each camera folder includes: A directory storing image data, linked via symbolic links to the image_data_downscaled directory Corresponding object detection results generated by YOLOv9, stored in .csv format A video file of the respective camera recording for the given scenario segment tracker_data: This directory contains tracking_results.tar.gz. After extraction, it provides tracking results derived from multi-camera detection outputs across all eight cameras. GNSS The GNSS-based position of the moving VRU recorded with a ublox ZED F9R GNSS receiver at 10Hz can be found in gnss.zip as seven rosbag folders rosbag2_[YY-MM-DD]-[HH-mm-ss]. Along the metadata.yaml and the rosbag .db3 files a .csv export can be found. This CSV file contains the UTC timstamp in nanoseconds, the decimal degree position in lattitude and longitude and the fix type: Fix type Meaning 1 Stand-alone 2 SBAS 3 DGNSS 4 RTK Fix 5 RTK Float For convenience, all rosbags CSV of both days have been added in on file: vru-rtk-track.csv Time synchronization and coordinate systems All sensors are expressed in a shared local Cartesian coordinate frame, allowing data from the different modalities to be overlaid directly. The origin of the Cartesian coordinate frame is set to UTM Easting 695310.500 m and Northing 5347376.094 m. The X-axis pointing to the East, the Y-axis to the North, the Z-axis Up against Gravity. Time synchronization was achieved via NTP using a local server synchronized to an external Stratum 1 source, resulting in sub-millisecond clock alignment across the sensor systems. Timestamps are expressed accross teh dataset in UTC Linux System Epoch time. Ground truth / annotations The dataset includes RTK-GNSS ground truth for the equipped test VRU. In addition, camera-based outputs provide 2D detections, tracked objects, 3D bounding box attributes, and object trajectories in the local coordinate frame. Radar data itself is not manually annotated, but labels can be derived using the synchronized camera-based object information. How to load and use the data The dataset is intended for cross-modal analysis as well as subsystem-specific studies. Users can process the UWB CIR recordings for sensing and localization, analyze radar detections and range-Doppler maps, use the camera outputs as visual reference or semantic labels, and compare results against the RTK-GNSS reference trajectory of the test VRU. Shared timestamps and a common coordinate frame enable straightforward synchronization and fusion across modalities. Citation and References If you use this dataset, please cite the associated publication describing the test site, sensor deployment, synchronization, calibration, subsystem file structure, and example measurements. de Ponte Müller, Fabian and Raslan, Mouhamed Aghiad and Liu, Mingyu and Oikonomopoulos-Zachos, Christos and Schmidhammer, Martin and Rashdan, Ibrahim and Merk, Benedikt and Kulke, Reinhard and Strand, Leah and Lakshminarasimhan, Venkatnarayanan and Uhlich, Tobias and Becker, Andreas and Sand, Stephan and Knoll, Alois (2026) A Multi-Modal UWB-Radar-Camera Dataset for Vulnerable-Road-User Sensing at Urban Intersections. In: 2026 IEEE 103rd Vehicular Technology Conference: VTC2026-Spring. 2026 IEEE 103rd Vehicular Technology Conference: VTC2026-Spring, 2026-06-09 - 2026-06-12, Nice, France. For further technical details on the measurement setup, metadata fields, and intended use cases, please refer to the accompanying paper. [IEEExplore Link] electronic library - A Multi-Modal UWB-Radar-Camera Dataset for Vulnerable-Road-User Sensing at Urban Intersections Further information on the VIDETEC-2 Project can be found https://www.videtec-projekt.de/