Toward Trustworthy Autonomous Driving Through the Quantification and Propagation of Uncertainty From HD Maps to Trajectory Prediction

The lack of trustworthiness of autonomous driving systems is one of the bottlenecks hindering their large-scale implementation and promotion in intelligent transportation systems. Current research predominantly relies on deterministic assumptions, and it remains unclear how to quantify the uncertainty of upstream modules effectively or whether incorporating such uncertainty into downstream modules can contribute to enhancing overall system performance. To address this gap, this paper constructs a trustworthy autonomous driving framework tailored for high-definition (HD) map estimation and trajectory prediction, with an emphasis on uncertainty awareness. An HD map estimation model adapted for uncertainty quantification is validated, enabling a trustworthiness evaluation of map estimation via an ensemble approach. Through a propagation-driven design, this uncertainty information is effectively leveraged in the trajectory prediction task. Experimental results on two classic downstream trajectory prediction models, HiVT and DenseTNT, reveal that incorporating the uncertainty of upstream map estimation can markedly enhance the accuracy and robustness of downstream trajectory prediction in complex, dynamic, and uncertain environments. Moreover, this research reveals a non-proportional relationship between map estimation accuracy and trajectory prediction performance, suggesting that greater emphasis should be placed on overall system performance rather than isolated task-specific metrics. These essential findings shed light on the development of trustworthy autonomous driving systems.