This PhD aims to develop uncertainty-aware hybrid AI models for robust vehicle state estimation and trajectory prediction in autonomous vehicles. Current localization systems face critical failure during GNSS outages, such as in tunnels and urban canyons, where the absence of satellite correction leads to rapid state estimation degradation. To address this, this research integrates machine learning with vehicle motion models and statistical filters to ensure continuous and reliable state estimation. Multiple uncertainty modelling strategies for hybrid AI systems are systematically explored, with their robustness evaluated under demanding real-world conditions.
The second phase focuses on multi-modal trajectory prediction, systematically embedding prior knowledge of driving scenarios, such as road topology, traffic rules, and agent interactions, into prediction algorithms to enhance their accuracy and physical plausibility. Anchoring data-driven models with structured scene understanding ensures that predicted trajectories remain contextually aware, realistic and aligned with automotive safety requirements.
