GeoPointTransformer
How can 3D scans of public spaces become useful for infrastructure maintenance? GeoPointTransformer identifies objects in point clouds and extracts the geometry engineers need.
Centre of Expertise Digital Operations & Finance
Simultaneous segmentation and geometry extraction from mobile-mapping point clouds
Dutch infrastructure owners Rijkswaterstaat, ProRail, and municipalities increasingly depend on 3D digital twins built from Mobile Laser Scanning (MLS) point clouds for maintenance and inspection. Modern AI can label those clouds, but it does not output the cylinder axes, radii, or curb polylines that engineers actually use. Today, up to 90 % of MLS processing time is still spent on manual digitisation. GeoPointTransformer closes that gap with a single transformer that labels points and fits the geometry in one pass.
Goal
Develop and validate a compact transformer-based neural network that, from raw MLS point clouds, simultaneously labels every point and outputs the geometric primitives, cylinders for poles today, curbs and rails next, used by infrastructure engineers.
Target group
- Dutch surveying SMEs and engineering consultancies running MLS pipelines.
- Public asset owners - Rijkswaterstaat, ProRail, municipalities, water authorities.
- Smart-city / BIM-GIS integrators building automated inspection tooling on top of LiDAR.
- The open-source geospatial-AI community.
Intended outcomes
- An open-source prototype for automated segmentation and geometry extraction from mobile-mapping point clouds. The code will be made available on GitHub, with trained model weights published on Hugging Face.
- An accurate AI model that achieves a semantic mIoU of more than 70% on the KITTI-360 dataset. For poles, the estimated radius, height and axis deviate by no more than 10% from the ground truth.
- Validation on at least one previously unseen Dutch trajectory from the 360Geo dataset, with a qualitative assessment by Hai Performance.
- A new Dutch MLS dataset made available to the geospatial AI community, supporting further research into AI applications for the public space.
- Knowledge sharing with both practice and academia, including a presentation at GeoBuzz and a peer-reviewed scientific publication.
Duration
January 2026 – December 2026
Team
THUAS — project lead
Research group Smart Sensor Systems, Faculty TIS.
- dr. Ir. Amey Vasulkar - research lead, AI model development
- dr. John Bolte - lector, Smart Sensor Systems · technical advisor
- Jassar Hasiba - student, model development & experiments (ADS&AI)
Hai Performance BV - MLS domain partner:
- dr. Ir. Tobias Witwer - domain lead, MLS workflows and asset annotation
360Geo BV - MLS data partner:
- Jacob Muilwijk - domain lead, MLS data and visualisation
- George Boot - tech lead, MLS data
Partners
- Hai Performance BV - MLS expertise and qualitative validation.
- 360Geo BV - Dutch MLS data and demonstration corridor.
Funding
Regieorgaan SIA - KIEM HighTech 2025.
- Project page: GeoPointTransformer
Participating programmes
- Bachelor Applied Data Science & AI
- Bachelor Applied Mathematics
- Master Next Level Engineering
Contact
dr. Ir. Amey Vasulkar - [email protected]
Links and publications
Background references:
- Robert, Raguet & Landrieu (2023). Efficient 3D Semantic Segmentation with Superpoint Transformer. ICCV.
- Behley et al. (2019). SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences
Figure: The GeoPointTransformer pipeline: A raw mobile-mapping point cloud (left) is processed by the AI model (centre) and transformed into a semantically labelled 3D scene (right), where objects such as a streetlight pole are automatically identified and their geometry is modelled.