HULI-Track: A ROS 2 LiDAR Dataset for Human Tracking in Large Unstructured Industrial Environments

Description HULI-Track is a ROS 2 LiDAR dataset for human detection and multi-human tracking in a realistic indoor industrial environment. The dataset was acquired using a Velodyne Puck VLP-16, mounted at a height of 0.60 m at the geometric center of the reference path, and is provided in ROS 2 bag format, enabling direct integration with ROS 2-based perception pipelines for point cloud processing, 3D detection, and multi-target tracking. The recordings were collected in an industrial laboratory environment containing shelves, corridors, and partial occlusions representative of real factory-floor conditions. The dataset includes scenarios with one, two, and four participants, performing motion sequences of increasing complexity. These include single-person laps, opposite-direction crossings, same-direction motion in close proximity, overtaking events, and multiple direction changes. The floor layout is defined by marked reference paths, which can be used for trajectory-level evaluation. The reference corner coordinates (in meters) of the marked area are: A = (-2.37861, 3.47239) C = (-2.39849, -3.47688) D = (2.51301, 3.46828) F = (2.51714, -3.44547) These markings define the trajectories followed by the participants during the experiments and provide a practical reference for evaluating tracking performance. Available data Each test_N folder may contain multiple data representations associated with the same experimental run: data_bbox/ -> ROS 2 bag containing the topics /velodyne_points, /velodyne_packets, /bbox, /filtered_bbox, and /bbox_data. The topic /bbox_data contains the filtered bounding boxes in array format. filtered_bbox/ -> ROS 2 bag containing the topics /velodyne_points, /velodyne_packets, /bbox, and /filtered_bbox. Here, /bbox corresponds to the original bounding box detections, while /filtered_bbox contains the filtered detections. original recording/ -> Original ROS 2 bag recorded in real time from the LiDAR. Includes the topics /velodyne_points, /velodyne_packets, and /bbox. Test_N -> Video recording of the corresponding experimental run. testN_filtered_bbox.txt ->Text file containing exported bounding box detections. Each line corresponds to one detection in the format: [x, y, z, dx, dy, dz, score, timestamp] testN_filtered_bbox_counted.txt -> Text file containing bounding box detections grouped by timestamp, together with the number of detections for each timestamp. 3D bounding box format A 3D bounding box is represented as: (x, y, z, dx, dy, dz, yaw) where: x, y, z are the coordinates of the object center dx, dy, dz are the box dimensions yaw is the orientation in 3D space In the exported .txt files, the bounding boxes are stored as: [x, y, z, dx, dy, dz, score, timestamp] where: score is the detection confidence timestamp is the detection time ROS 2 Topics and Message Types The dataset may include recordings of the following ROS 2 topics: /velodyne_points — sensor_msgs/msg/PointCloud2-> LiDAR point cloud data. /velodyne_packets — velodyne_msgs/msg/VelodyneScan -> Raw Velodyne packet data. /bbox — vision_msgs/msg/Detection3DArray -> Original 3D bounding box detections. /filtered_bbox — vision_msgs/msg/Detection3DArray -> Filtered 3D bounding box detections. /bbox_data — custom array-format message. Filtered bounding box data stored in the format: [x, y, z, dx, dy, dz, score, timestamp]. Experimental scenarios The dataset contains 14 test scenarios of increasing complexity: test_1: 1 person, starting at B, 3 laps, anti-clockwise test_2: 1 person, starting at B, changing direction whenever reaching E, 3 laps test_3: 1 person, starting at A, changing direction midway between C and F, 3 laps test_4: 2 people, starting at B, one clockwise and one anti-clockwise, crossing at B and E, 3 laps test_5: 2 people, starting at B and E, one clockwise and one anti-clockwise, crossing midway between A-D and C-F, 3 laps test_6: 2 people, starting at A and B, both clockwise at the same speed, 3 laps test_7: 2 people, starting at A and B, both clockwise at different speeds, 3 laps test_8: same configuration as test_4 test_9: 2 people, starting at A and F, both clockwise, 3 laps test_10: 4 people, starting at A, B, C, and midpoint C-F, all anti-clockwise, 2 laps test_11: 4 people, with 2 starting at A and C and 2 additional participants joining later through A and C; 2 move clockwise and 2 anti-clockwise at constant speed, 2 laps test_12: 4 people, starting at A, B, E, and F, all anti-clockwise, different speeds with overtaking events, 2 laps test_13: 4 people, starting at B, C, E, and midpoint A-D; all initially move clockwise, with 2 participants changing direction during the run, 2 laps test_14: 2 people, starting at B and E, several direction changes at constant speed Intended use HULI-Track is intended to support research in: LiDAR-based human detection multi-target tracking trajectory estimation and analysis data association under crossing and close-proximity interactions ROS 2 perception pipelines The dataset is particularly suitable for evaluating algorithms under realistic industrial challenges such as partial occlusions, trajectory crossings, close-proximity motion, overtaking, and direction changes.