| KEYWORD |
Generating 3D synthetic data to train AI algorithms for intelligent vehicle applications
Thesis in external company
keywords ARTIFICIAL INTELLIGENCE, COMPUTER ANIMATION, COMPUTER GRAPHICS, COMPUTER VISION, DEEP LEARNING, GENERATIVE AI, NEURAL NETWORKS
Reference persons FABRIZIO LAMBERTI
External reference persons FEDERICO BOSCOLO
Research Groups DAUIN - GR-09 - GRAphics and INtelligent Systems - GRAINS
Thesis type THESIS IN COLLAB. W/ A COMPANY, THESIS WITH A COMPANY
Description The recognition of the vehicle’s owner or the person authorized to operate a given vehicle will be a challenge for future intelligent vehicle applications. Furthermore, the person to be recognized stands outside the car, in different and complex conditions (recognition in the “wild”): highly variable lighting, presence of other subjects, etc. Apart from the difficulty and complexity of the recognition algorithm, one of the fundamental problems becomes the data needed to train required AI models, which cannot be available yet.
In this case, so-called synthetic data need to be employed. Synthetic data, generated through some simulation environments, can mimic operational conditions, by also enabling to keep edge cases into account. In fact, real-world datasets often contain imbalances, because edge cases, which do not happen frequently in real life, are not sufficiently represented. Finally, with synthetic data, other issues, e.g., related to privacy and GDPR compliance, can be easily overcome.
The focus of this thesis, developed in collaboration with Centro Ricerche Fiat (CRF) and Stellantis, will be the use of computer graphics simulation (e.g., with tools like Unreal Engine's MetaHuman) or generative methods for the the production of synthetic data that could be later used to train AI algorithms for the recognition of vehicle owners that are robust to variations in terms of, e.g., illumination, presence of other human and non-human subjects, etc.
See also http://grains.polito.it/work.php
Deadline 15/01/2025
PROPONI LA TUA CANDIDATURA