Internships and Thesis Research

Internships & Thesis Research from our partners

Snellius wants to make it easier to connect companies and students who are looking for interns and internships respectively. Below you can find a number of interesting internship topics. If you are interested in them, send an email to snellius@tudelft.nl and we will then put you in touch with the company.

Do you want to stay up to date on internship topics? You can subscribe to the mailing list below to receive all e-mails that Snellius receives about career events/PhD positions/internships/job vacancies and more.

We wish you success with your internship!

Internship/Thesis research at CGI

For intern we offer many interesting topics in a collaboration with the educational institute, yourself and CGI. This has to be an internship or a thesis at or near the end of your study. Please reach to the CGI people around you to ask more about this or contact the person in the side bar. We work on many innovations with 3D, point clouds, object recognition, 3D data stores, geo analytics and positioning solutions. Since 2016 CGI had twelve Geo master students doing internships or master thesis projects and 4 group projects. More than 10 of those joined CGI as employee afterwards.

Afstudeerstage Geodetisch Meetproces & Automatisering

Als stagiair ga jij binnen een innovatieve werkomgeving aan de slag met het vereenvoudigen vangeodetische metingen.

Hoe ziet je afstudeerstage bij Antea Group eruit?

Binnen deze afstudeerstage werk je aan het vereenvoudigen en standaardiseren van landmeetkundige processen. De ambitie is om meetwerk efficiënter, toegankelijker en toekomstbestendig te maken door het ontwikkelen van uniforme meetprotocollen en het automatiseren van dataverwerking.

Je richt je op veelgebruikte meettechnieken zoals total stations, GNSS (GPS) en waterpassing. Door deze technieken te analyseren en te structureren, draag je bij aan een gestandaardiseerde werkwijze die breed inzetbaar is binnen projecten van Antea Group.

Daarnaast onderzoek je hoe automatisering kan worden toegepast in het meetproces, bijvoorbeeld in de koppeling tussen veldwerk en dataverwerking. Je ontwikkelt een methodiek of workflow waarmee meetresultaten sneller en consistenter kunnen worden verwerkt en gevalideerd.

Dit vind jij mooi om te doen:

• Werken aan technische vraagstukken binnen de landmeetkunde en geodesie;
• Analyseren en verbeteren van meetmethodes en werkprocessen;
• Opstellen van gestandaardiseerde meetprotocollen voor verschillende meettechnieken;
• Werken met meetinstrumenten zoals total station, GNSS en waterpasinstrumenten;
• Samenwerken met specialisten binnen geodesie en geo-informatie.

Dit bieden wij jou:

• Een afstudeerstage met directe toepassing in de praktijk van infrastructurele projecten;
• Begeleiding door ervaren specialisten in geodesie, landmeten en digitalisering;
• Ruimte voor eigen initiatief en het ontwikkelen van innovatieve oplossingen;
• Toegang tot moderne meetapparatuur en software;
• Een passende stagevergoeding;
• Een flexibele werkomgeving, met de mogelijkheid om deels vanuit huis te werken en regelmatig samen te komen op kantoor (bijv. in Oosterhout).

Dit breng jij mee:

• WO werk- en denkniveau;
• Je volgt een opleiding in Geodesie, Geo-informatie, Civiele Techniek of vergelijkbaar;
• Interesse in landmeetkunde, meetinstrumenten en data-automatisering;
• Affiniteit met programmeren of scripting (bijv. Python, Matlab of vergelijkbaar) is een pré;
• Analytisch vermogen en een gestructureerde werkwijze;
• Goede communicatieve vaardigheden en een proactieve houding.

Improving Roof Ridgeline Detection in Airborne LiDAR Point Clouds

Short description of the topic

At Geodelta, we use roof ridgelines to improve the alignment of airborne LiDAR point clouds, including data from the Actueel Hoogtebestand Nederland (AHN).

Roof ridgelines are useful because they can be determined from the intersection of two roof planes. Even when two point clouds do not measure exactly the same points, they can still capture the same roof geometry. Comparing ridgelines of stable buildings can help identify and correct differences between flight strips and acquisition years.

Geodelta has developed an automatic ridgeline detection method. We would like to understand when it works reliably, when it fails, and how it can be improved. We also want to know how precisely each ridgeline can be determined and how this precision affects the least-squares adjustment used for aligning point clouds.

Relevant literature

Vosselman, G. (2008). Analysis of Planimetric Accuracy of Airborne Laser Scanning Surveys. https://www.researchgate.net/publication/228706705
Vosselman, G. (2012). Automated Planimetric Quality Control in High Accuracy Airborne Laser Scanning Surveys. ISPRS Journal of Photogrammetry and Remote Sensing, 74, 90-100. https://doi.org/10.1016/j.isprsjprs.2012.09.002

Objectives and challenges for this project include

Evaluation of the existing detection method: Determine under which conditions the method works well, and identify failures related to roof shapes, point density, scanning angle or obstacles.
Improving detection: Investigate and, where possible, implement improvements to the existing algorithm.
Estimating precision: Determine how accurately the position and direction of a ridgeline can be estimated from LiDAR points.
Effect on adjustment: Investigate how ridgeline errors influence the least-squares alignment of overlapping point clouds.
Observation weighting: Explore whether precision estimates can be used to give reliable ridgelines greater weight in the adjustment.

The thesis can build on existing Geodelta software and AHN datasets, allowing the student to focus on measurable improvements.

Relevant competencies

Studying Geodesy, Geomatics or a related technical field.
Knowledge of least-squares adjustment and error propagation, or interest in developing these skills.
Programming experience, preferably C#.
Interest in LiDAR processing, geometry and quantitative algorithm evaluation.

About Geodelta

At Geodelta, we believe that measuring is about more than precision. Reliability is what counts. Are the measurements correct, and have the right procedures been followed?

We are an independent engineering consultancy specializing in geodesy, photogrammetry and laser scanning. We develop user-friendly software and provide expert technical advice to help our clients trust their measurements. We are innovative, independent and pragmatic, and above all, we share a passion for geodesy.

Our office is located in Delft, in the historic former Faculty of Geodesy building at TU Delft, close to the Faculty of Architecture.

Interested?

If you are interested in this topic, or if you have an interesting research idea of your own, we would love to hear from you! We are happy to meet with interested students to discuss the research and find a project that fits their interests and academic background.

Joint Least-Squares Adjustment of Terrestrial Laser Scans Using Geometric Features

Short description of the topic

In traditional surveying, a total station can observe the same point from different instrument positions. Combining these observations in a least-squares adjustment improves estimates of both the point locations and the instrument positions and orientations. Can we develop a similar approach for stationary laser scanning?

A laser scanner measures millions of different points, but repeated scans generally do not sample the exact same surface points. Instead, they measure common surfaces such as walls or floors. Geodelta wants to investigate whether these shared geometric features can support a joint adjustment of scanner positions, orientations and potentially individual measured points.

One idea is to construct virtual reference points from intersections: two lines in 2D or three planes in 3D can define a point that need not have been captured by the scanners. If the same intersection can be measured independently with a total station, it can also place the point cloud in real-world coordinates.

The project starts in 2D, which is mathematically more manageable and relevant for horizontal facade mapping, for example for Kadaster applications. Extension to 3D is optional.

Relevant literature

Grant, D., Bethel, J. & Crawford, M. (2012). Rigorous Point-to-Plane Registration of Terrestrial Laser Scans. ISPRS Archives, XXXIX-B5, 181-186. https://doi.org/10.5194/isprsarchives-XXXIX-B5-181-2012
Schnabel, R., Wahl, R. & Klein, R. (2007). Efficient RANSAC for Point-Cloud Shape Detection. Computer Graphics Forum, 26(2), 214-226. https://doi.org/10.1111/j.1467-8659.2007.01016.x

Objectives and challenges for this project include

Representing lines and planes: Investigate which parameter descriptions are most suitable for adjustment, such as a normal vector and distance.
Developing a joint adjustment: Formulate a method using common features to improve stationary scanner positions and orientations, and investigate possible corrections to the measurements themselves.
Using virtual reference points: Study the precision of intersection points and how independently surveyed intersections can connect relative scans to a global coordinate system.
Understanding measurement precision: Investigate how RANSAC plane or line detection influences the precision and reliability of the full solution.
Prototype and evaluation: Develop and test a mathematical model in 2D with simulated and/or measured data; extend to 3D only if time permits.

Moving scanners (terrestrial or airborne) are a possible longer-term research direction, not a required part of this thesis.

Relevant competencies

Studying Geodesy, Geomatics or a related field.
Interest in linear algebra, adjustment theory and measurement accuracy.
Programming experience, ideally in C#.
Interest in terrestrial laser scanning, surveying and mathematical modelling.

About Geodelta

At Geodelta, we believe that measuring is about more than precision. Reliability is what counts. Are the measurements correct, and have the right procedures been followed?

Our office is located in Delft, in the historic former Faculty of Geodesy building at TU Delft, close to the Faculty of Architecture.

Interested?

Developing AI-Based Mapping Methods for Oblique Aerial Imagery

Short description of the topic

Deep-learning models can already extract buildings, roads and other mapped features from aerial photographs. Some models produce vector geometries, such as building polygons or road outlines, rather than just pixel masks. Most have been trained on orthophotos or near-vertical images, so they may not work as well on oblique photographs.

Oblique aerial images show roofs and facades from different angles, but perspective, occlusion and changing scale make automatic mapping difficult. At Geodelta, we want to adapt promising existing AI models to these images. The student will prepare an oblique-image dataset, retrain or fine-tune selected models, and develop and test improvements based on the results. The goal is to achieve more accurate geometric mapping from oblique imagery, not only to assess pretrained models.

Possible starting points include LDPoly (road polygons), Pix2Poly (building polygons) and BONAI (building footprints from off-nadir images). MapTRv2, which generates vectorized maps from vehicle-camera images, could provide a separate direction using Geodelta panoramic imagery. Students are not expected to design an entire deep-learning network from scratch, but training and adapting an existing model are central parts of the project.

Relevant literature

Jiao, W. et al. (2025). LDPoly: Latent Diffusion for Polygonal Road Outline Extraction in Large-Scale Topographic Mapping. https://arxiv.org/abs/2504.20645
Adimoolam, Y. K., Poullis, C. & Averkiou, M. (2025). Pix2Poly: A Sequence Prediction Method for End-to-end Polygonal Building Footprint Extraction from Remote Sensing Imagery. https://arxiv.org/abs/2412.07899
Wang, J. et al. (2022). Learning to Extract Building Footprints from Off-Nadir Aerial Images (BONAI). https://arxiv.org/abs/2204.13637
Liao, B. et al. MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction. https://arxiv.org/abs/2308.05736

Objectives and challenges for this project include

Investigating AI models: Study promising architectures and choose one or two methods that could be adapted for the desired mapping task (buildings or roads).
Preparing a dataset: Prepare Geodelta oblique images and suitable reference annotations, considering the differences between roof outlines, visible silhouettes and ground-level footprints.
Testing and retraining on oblique imagery: Apply one or two selected models with existing implementations and investigate the value of retraining or fine-tuning on oblique images.
Evaluating detection quality: Measure geometric accuracy and identify failures related to viewing angle, occlusion, building height and object type.
Identifying next steps: Recommend which methods or model adaptations could support practical mapping at Geodelta.

Relevant competencies

Studying Geomatics, Computer Science, AI or a related field.
Experience with Python and basic machine learning/deep learning.
Interest in computer vision, image processing and photogrammetry.
Willingness to prepare datasets and evaluate experiments critically.

About Geodelta

At Geodelta, we believe that measuring is about more than precision. Reliability is what counts. Are the measurements correct, and have the right procedures been followed?

Our office is located in Delft, in the historic former Faculty of Geodesy building at TU Delft, close to the Faculty of Architecture.

Interested?

Intuitive Navigation Through Indoor Point Clouds

Short description of the topic

Laser scanners capture detailed three-dimensional point clouds of indoor spaces, including walls,
floors, doors, staircases and furniture. However, exploring these point clouds in a conventional viewer
can be surprisingly difficult. Users may accidentally move through walls, leave the building or pass
through floors and ceilings.

Geodelta wants to investigate a more natural way to explore indoor point clouds. The viewer could use
the observed building geometry to keep the camera inside accessible space, move through doorways
and corridors, and move between floors.

A major challenge is that a point cloud is not a solid building model: it does not explicitly say which
spaces are accessible, and gaps in the measurements do not necessarily represent openings. The aim is
to identify navigable space directly from the point cloud, preferably without first constructing a complete
BIM or 3D model.

Relevant literature

Díaz-Vilariño, L. et al. (2016). Indoor Navigation from Point Clouds: 3D Modelling and Obstacle Detection. ISPRS Archives, XLI-B4, 275-281.
https://doi.org/10.5194/isprs-archives-XLI-B4-275-2016
Balado, J. et al. (2019). Point Clouds to Direct Indoor Pedestrian Pathfinding. ISPRS Archives, XLII-2/W13, 753-759. https://doi.org/10.5194/isprs-archives-XLII-2-W13-753-2019
Díaz-Vilariño, L. et al. (2019). Obstacle-Aware Indoor Pathfinding Using Point Clouds. ISPRS International Journal of Geo-Information, 8(5), 233.
https://doi.org/10.3390/ijgi8050233

Objectives and challenges for this project include

Identifying navigable space: Detect floors, walls and openings from indoor point clouds.
Navigation constraints: Use detected geometry to prevent the camera from moving through solid
obstacles while allowing natural movement.
Connecting rooms and floors: Identify possible passages and transitions between rooms, corridors
and levels.
Handling incomplete data: Distinguish genuinely free space from regions that were simply not
observed.
Prototype and evaluation: Build a navigation demonstration and compare its usability with
unrestricted free-flight controls.

Relevant competencies

Studying Geomatics, Computer Science, Geo-Information Science or a related field.
Programming experience, preferably Python or C#.
Interest in point-cloud processing, 3D geometry and spatial data structures.
Interest in interactive visualization and usability.

About Geodelta

At Geodelta, we believe that measuring is about more than precision. Reliability is what counts. Are the measurements correct, and have the right procedures been followed?

Our office is located in Delft, in the historic former Faculty of Geodesy building at TU Delft, close to the Faculty of Architecture.

Interested?

Automatic Detection of Low-Flying Bird Flocks in Airborne LiDAR Data

Short description of the topic

The Actueel Hoogtebestand Nederland (AHN) is a high-resolution elevation dataset built using airborne
laser scanning. The resulting point clouds contain measurements of the ground and permanent objects,
but sometimes also of birds flying through the survey area.

High-flying birds are often easily identified as outliers. Flocks flying only a few metres above the ground
are harder to recognize and may escape existing filters. They can introduce false elevations into the data.
Geodelta performs AHN quality control, and a successful method could contribute to checking new
datasets scheduled for delivery in 2027.

Processing efficiency is important: AHN point clouds are huge. Ideally, suspicious bird patterns can be
found directly in the much smaller raster products, avoiding another pass through all the LiDAR points.
The scientific challenge is whether enough information about birds remains after rasterization, and how
to distinguish them from genuine objects.

Relevant literature

Griffioen, S. (2017). 3D Mathematical Morphology on Voxelized Point Clouds for Outlier Detection. TU
Delft. https://repository.tudelft.nl/file/File_f0c7c79e-8362-4520-b7a7-95e9ce97c64a
Carrilho, A. C., Galo, M. & Santos, R. C. (2018). Statistical Outlier Detection Method for Airborne LiDAR Data. ISPRS Archives, XLII-1, 87-92.
https://doi.org/10.5194/isprs-archives-XLII-1-87-2018
Jiang, G. et al. (2024). A Maximum Entropy Based Outlier Removal for Airborne LiDAR Point Clouds. IEEE JSTARS, 17, 19130-19145.
https://doi.org/10.1109/JSTARS.2024.3478069

Objectives and challenges for this project include

Understanding outliers: Examine how low-flying bird flocks appear in point clouds and elevation
rasters, and why existing methods miss them.
Investigating raster methods: Study approaches from image processing, statistics or machine
learning for detecting unusual height patterns.
Building a detection workflow: Test promising methods in raster data, with targeted point-cloud
checks if necessary.
Validating reliability: Create a reference dataset and quantify missed flocks and false detections,
including confusion with trees or structures.
Assessing efficiency: Measure the trade-off between detection accuracy and processing time to
assess use in operational AHN quality control.

Geodelta can provide relevant datasets and support algorithm development. The central MSc
contribution is a rigorous comparison and understanding of which information and methods are needed
for reliable detection.

Relevant competencies

Studying Geomatics, Geodesy, Computer Science or a related field.
Programming experience, preferably Python or C#.
Interest in LiDAR, raster analysis, statistical filters or machine learning.
Ability to design careful tests and quantify data quality.

About Geodelta

At Geodelta, we believe that measuring is about more than precision. Reliability is what counts. Are the measurements correct, and have the right procedures been followed?

Our office is located in Delft, in the historic former Faculty of Geodesy building at TU Delft, close to the Faculty of Architecture.