Natural and Semi-Urban Dataset

Weekly Multi-Temporal Dataset for Short-Term Localization and Environmental Change Analysis

Dataset website: raoufdannaoui1.github.io/Natural_and_Semi-Urban_Dataset/

Overview

The dataset is a short-term, high-resolution, multi-modal dataset focused on understanding how real-world changes—such as vegetation growth, trimming, and object displacement—affect 3D LiDAR-based localization in dynamic outdoor environments. The data was collected weekly from February to April 2025, across two contrasting outdoor scenarios:

Natural Environment: INRAE Forest, Clermont-Ferrand
Semi-Urban Environment: Around Université Clermont Auvergne

This dataset provides a unique opportunity to analyze and benchmark localization robustness over short time intervals, making it ideal for applications in re-localization, change detection, SLAM, and dynamic mapping.

Dataset Description

Each weekly session contains:

Two Environments:
- Track 1-2 (bidirectional): Semi-urban university campus (~850m)
- Track 3-4 (bidirectional): Natural forest path (~680m)

Data Structure (per week):

weekXX_hhmm-DD-MM-YYYY/
│── assets/
│   └── track_trajectories
│
│── images_360/
│   ├── SemiUrban_track1-2/
│   └── Urban_track3-4/
│
└── point_clouds/
    ├── SemiUrban_track1-2/
    └── Urban_track3-4/

Contents:
- Colorized & classified point clouds
  - Provided per semantic class (ground, vegetation, vehicles, etc.)
  - ~100 million points for a Semi-Urban map and ~150 million points for a Natural map
  - Average size: ~4 GB for a Semi-Urban map and ~5.5 GB for a Natural map (LAS format files)
- Panoramic images
  - 360° images captured every 3 meters along each track
  - Resolution: 7040 × 3520 pixels
  - Average size: ~2–3 MB per image (JPEG)
  - Includes GNSS-tagged camera trajectory
- GNSS/IMU trajectories
  - Provides full pose: position (x, y, z) and orientation (quaternion: qw, qx, qy, qz)

Applications

This dataset supports research in:

Re-localization and map alignment
Short-term environmental dynamics
ICP evaluation (Point-to-Point, Point-to-Plane, etc.)
Scene segmentation and change detection
Structure-aware risk prediction in navigation

Acquisition Platform

All data is recorded using a Leica Pegasus TRK100 mounted on a Zoe electric vehicle, the Leica is equipped with:

Two 16-beam LiDARs
360° panoramic camera
GNSS and IMU for centimeter-level pose accuracy

Data Availability

The full dataset files (point clouds, images, and trajectories) is publicly available directly on Hugging Face, or via raoufdannaoui1.github.io/Natural_and_Semi-Urban_Dataset for selective download.

Citation

If you use this dataset in your research, please cite our paper.

 @inproceedings{ardannaoui_2025_icp_analysis,
   title={When and Where Localization Fails: An Analysis of the Iterative Closest Point in Evolving Environments},
   author={Dannaoui, Abdelraouf and Laconte, Johann and Debain, Christophe and Pomerleau, Fran{\c{c}}ois and Checchin, Paul},
   booktitle={2025 European Conference on Mobile Robots (ECMR)},
   year={2025},
   organization={IEEE},
   note={Accepted for publication}
 }

Contact

For questions or collaborations, feel free to contact:

Abdel-Raouf Dannaoui
Ph.D. Candidate in Robotics -- INRAE
dannaoui.abdelraouf@gmail.com

image imagewidth (px) 7.04k 7.04k