Robustness Evaluation of Multi-modal Perception Systems in Autonomous Driving

This thesis examines a key safety risk in self-driving cars: their sensors can fail in everyday conditions such as rain, glare or electrical faults. When this happens, the car may not interpret its surroundings correctly. Current testing focuses on ideal conditions and relies on costly human-labelled data, which does not reflect real-world driving. This thesis develops a new way to measure how stable a vehicle remains when its sensors are disturbed. It tests the car’s decision-making under simulated hazards, offering a practical and low-cost method to improve the safety of autonomous vehicles before they are deployed.