This repository provides the CARLA simulation dataset used in the following paper:
VLM-Enhanced Vehicle-Infrastructure Collaborative Framework for Incremental HD Map Updating
Shaoting Qiu, Dongzhe Su, Runzhi Hu, Weisong Wen, Tacitus Hui, Stella Zhu, Feng Huang
HD maps are essential for autonomous driving but frequently become outdated due to urban construction and infrastructure changes. This dataset supports research on incremental HD map updating using vehicle-infrastructure collaborative perception.
The dataset contains two components:
| Component | Source | Description |
|---|---|---|
| CARLA Simulation | CARLA Simulator + HKSTP map | Two construction change scenarios (this repo) |
| Real-World | UrbanV2X | HKSTP construction site, Feb–Dec 2025 |
Each scenario uses two recordings — one with clear road conditions and one with construction obstacles — shared across both scenes with swapped roles.
HD_MAP_UPDATE_DATASETS/
├── east_clear/ # Vehicle-side · clear road
│ ├── town03_lidar_test_clear.bag # ROS bag: vehicle LiDAR + camera, 2.21 GB
│ ├── gt_global.txt # Global ground truth trajectory
│ ├── route1_gt_global.txt # Route-level ground truth
│ ├── town03_gt.m # MATLAB ground truth helper
│ ├── route_vehicle_status.csv # Vehicle status log (per route)
│ └── vehicle_status.csv # Vehicle status log (full session)
├── east_construction/ # Vehicle-side · construction scenario
│ ├── hkstp_east_construction.bag # ROS bag: vehicle LiDAR + camera, 2.21 GB
│ ├── gt_global.txt
│ ├── town03_gt.m
│ └── vehicle_status.csv
├── roadside_0923/
│ ├── data_clear/ # Roadside · clear road
│ │ ├── lidar_test_data_with_seg_clear.bag # ROS bag: roadside LiDAR, 6.48 GB
│ │ └── config/ # Sensor and world config files
│ └── data_with_obs/ # Roadside · construction scenario
│ ├── lidar_test_data_with_seg.bag # ROS bag: roadside LiDAR, 6.49 GB
│ ├── actor_settings_hksp_with_infr.json
│ ├── sensor_config_infrastructure.json
│ ├── sensor_config_template_32line.json
│ └── world_config_town03_revise_hkstp.json
└── README.md
How the folders map to each scenario:
| Folder | S1: Construction | S2: Restoration |
|---|---|---|
east_clear |
Vehicle baseline | Vehicle new scan |
east_construction |
Vehicle new scan | Vehicle baseline |
roadside_0923/data_clear |
Roadside baseline | Roadside new scan |
roadside_0923/data_with_obs |
Roadside new scan | Roadside baseline |
The simulation replicates the UrbanV2X sensor platform deployed at the Hong Kong Science and Technology Park (HKSTP).
| Sensor | Specification |
|---|---|
| LiDAR | 360° spinning, mounted at 6 m height |
| Camera | RGB surround-view cameras |
| GNSS | Fixed reference position |
| Sensor | Specification |
|---|---|
| LiDAR | 64-beam spinning LiDAR |
| Camera | Surround-view RGB cameras |
| GNSS | RTK-GPS |
Roadside Construction View Roadside view with temporary construction-related changes. |
Roadside Clean View Roadside view under clean road conditions. |
Side-by-side visualization of roadside observations under construction and clean conditions.
View construction video | View clean video | View side-by-side demo page
- Change type: Additions only
- Description: Construction barriers and equipment are introduced into a road segment at HKSTP
- Task: Detect newly added obstacles and modified lane boundaries
- Result: 3D mean Euclidean distance of 4.07 cm
- Change type: Bidirectional (additions + deletions)
- Description: Construction elements are removed and original road markings are restored
- Task: Simultaneously detect deleted construction-period boundaries and newly restored markings
- Result: 3D mean Euclidean distance of 15.54 cm
Real-world experiments in this paper use the publicly available UrbanV2X dataset collected at HKSTP, Hong Kong.
- Baseline map: Captured in February 2025 (construction barriers present)
- New scan: Collected in December 2025 (construction completed, original road restored)
Please refer to the official UrbanV2X repository for download and usage instructions:
Qin, Q., Zhang, Z., Zhong, Y., Huang, F., Liu, X., Hu, R., Chen, H., Hu, W., Su, D., Zhang, J., Ng, H.-F., & Wen, W. (2025).
UrbanV2X: A Multisensory Vehicle-Infrastructure Dataset for Cooperative Navigation in Urban Areas.
Accepted by IEEE ITSC 2025.
🔗 https://polyu-taslab.github.io/UrbanV2X/
| Scenario | |ΔX| (cm) | |ΔY| (cm) | |ΔZ| (cm) | 3D Mean (cm) |
|---|---|---|---|---|
| S1: Construction | 2.61 ± 2.36 | 2.86 ± 3.69 | 1.51 ± 2.12 | 4.07 |
| S2: Restoration | 10.81 ± 7.39 | 5.18 ± 3.62 | 10.07 ± 13.31 | 15.54 |
# Python 3.8+, ROS Noetic recommended
pip install open3d numpyROS bag files are hosted on Dropbox. Scene 1 data is currently available; Scene 2 will be released in a future update.
Scene 1 — Construction Work
| Folder | Link |
|---|---|
east_clear (vehicle baseline) |
Download |
east_construction (vehicle new scan) |
Download |
roadside_0923/data_clear (roadside baseline) |
Download |
roadside_0923/data_with_obs (roadside new scan) |
Download |
Scene 2 — Post-Construction Restoration: coming soon.
Play back ROS bags:
# Vehicle-side
rosbag play east_clear/town03_lidar_test_clear.bag
rosbag play east_construction/hkstp_east_construction.bag
# Roadside
rosbag play roadside_0923/data_clear/lidar_test_data_with_seg_clear.bag
rosbag play roadside_0923/data_with_obs/lidar_test_data_with_seg.bagGround truth trajectories are provided in gt_global.txt (format: timestamp x y z qx qy qz qw).
This dataset is released for academic research and non-commercial use.
Unless otherwise specified, the dataset metadata, configuration files, and documentation in this repository are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).
You are free to use, share, and adapt the dataset for non-commercial research and educational purposes, provided that proper credit is given to the authors and the related paper is cited.
Commercial use, redistribution for commercial purposes, or deployment in commercial autonomous driving systems is not permitted without prior written permission from PolyU TAS Lab.
For commercial inquiries or permission requests, please contact: welson.wen@polyu.edu.hk
This work was supported by the Innovation and Technology Fund under the projects "Safety-Certified Multi-Source Fusion Positioning for Autonomous Vehicles in Complex Scenarios (ZPE8)" and "Advanced Smart Mobility Road-Side and Edge System (ART/369CP)".
The authors thank Ziqi Zhang and Qijun Qin for their generous support in providing experimental data and guidance for the vehicle-infrastructure collaborative mapping experiments.
The HD vector map generation workflow in this project was developed with reference to the open-source HD map construction work by Runzhi Hu:
- Runzhi Hu, HDMap: HD Vector Map Builder
https://github.com/ebhrz/HDMap
We sincerely thank Runzhi Hu for making the HDMap project publicly available and for providing valuable methodological references for HD vector map generation and urban mapping experiments.
The real-world experiments are related to the UrbanV2X dataset collected at the Hong Kong Science and Technology Park. Please refer to the official UrbanV2X website for dataset details and citation information.