Quantifying learning algorithm uncertainties in autonomous driving systems: Enhancing safety through Polynomial Chaos Expansion and High Definition maps

Abstract

Autonomous driving systems (ADS), leveraging advancements in learning algorithms, have the potential to significantly enhance traffic safety by reducing human errors. However, a major challenge in evaluating ADS safety is quantifying the performance uncertainties inherent in these black box algorithms, especially in dynamic and complex service environments. Addressing this challenge is crucial for maintaining public trust and promoting widespread ADS adoption. In this work, we propose a Polynomial Chaos Expansion (PCE) approach, utilizing High Definition (HD) maps to quantify positional uncertainties from an ADS object detection algorithm. The PCE-based approach also offers the flexibility for online self-updating, accommodating data shifts due to changing operational conditions. Tested in both simulation and real-world experiments, the PCE method demonstrates more accurate uncertainty quantification than baseline models. Additionally, the results highlight the importance and effectiveness of the self-updating capability, particularly when encountering weather changes.