A Dependable Trajectory Prediction Pipeline for Autonomous Driving: Integrating Computer Vision and Bayesian Networks for Highway Safety

Safe autonomous driving requires an accurate and robust perception system capable of detecting, localizing, and tracking surrounding vehicles and road features. The trajectory prediction task in autonomous driving enables the vehicle to anticipate the future movements of other road users. However, trajectory prediction remains challenging due to sensory disturbances, environmental uncertainties, and road users' unpredictable behavior. This contribution demonstrates how the robustness of a machine-learning-based trajectory prediction can be evaluated using Bayesian Networks. A reliability score was obtained by systematically considering vehicle and road features. This reliability score provides an additional layer of safety awareness, enabling the system to make risk-aware decisions based on the predicted trajectory confidence. The evaluation pipeline is trained and evaluated using the KITTY dataset.