On 30 May 2025 at 10:15 Alireza Akhavi Zadegan will defend his doctoral thesis "A Multimodal approach for refining Mapping and Localization by Integrating Generative AI and Pedestrian-Centric Data" to obtain the degree of Doctor of Philosophy (in Computer Science).
Supervisor
Assoc. Prof. Amnir Hadachi, University of Tartu
Opponents
Assoc. Prof. Alain Kibangou, University Grenoble Alpes (France)
Assist. Prof. Salvatore Flavio Pileggi, University of Technology Sydney (Australia)
Summary
Cities are dynamic ecosystems where sidewalks, bike lanes, roads, and crosswalks seamlessly merge into shared spaces. With the increasing presence of delivery robots, and e-scooters alongside pedestrians, the demand for accurate, up-to-date digital maps has never been greater. Yet, traditional mapping focuses primarily on car roads, leaving pedestrian zones—here some of the most transformative mobility innovations are unfolding—generally underexplored.
This research utilizes multimodal and multisensory data to improve mapping and localization, making maps more adaptable. Instead of relying solely on satellite imagery or slow manual updates, this approach leverages AI to generate new map perspectives, connect different viewpoints, and streamline the mapping process.
To collect real-world urban data, the thesis' author designed a custom electric scooter equipped with an advanced suite of sensors, including stereo cameras, LiDAR, GPS, and an audio recorder. This mobile platform captures detailed multimodal data from sidewalks, crosswalks, and shared spaces. Cameras record colors and textures, LiDAR provides depth measurements, and audio recordings capture environmental sounds such as traffic and construction. Together, these sensors form the DELTA dataset, a collection specifically focused on pedestrian pathways.
To improve map creation and localization, the author developed Street2Sat, a framework that transforms street-level images into satellite views. By connecting these perspectives and identifying key landmarks, the system enhances mapping and localization.
To make this data more useful for urban planners and navigation systems, the thesis' author introduced Street2GIS. This framework converts processed images into standard Geographic Information System (GIS) formats, such as shapefiles, outlining roads, sidewalks, buildings, and vegetation. It achieves this by integrating depth estimation for understanding distances and semantic segmentation for labeling different features within a streamlined processing pipeline.
By improving the way we map pedestrian spaces, this research contributes to a more accessible and responsive urban environment. With better data on sidewalks, crossings, and shared spaces, city planners can design infrastructure that meets the needs of all residents, from pedestrians and cyclists to individuals with mobility challenges.
Beyond urban planning, these AI-driven maps could support emerging technologies like autonomous delivery robots and smart mobility solutions, helping them navigate complex city landscapes more effectively. As cities grow and change, having a flexible, data-driven approach to mapping will be key to creating safer, more inclusive public spaces for everyone.
The defence will be held also in Zoom (meeting ID: 670 504 9543, passcode: ati).
