Quantifying learning algorithm uncertainties in autonomous driving systems: Enhancing safety through Polynomial Chaos...

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Autonomous driving systems (ADS) promise enhanced traffic safety by minimizing human error.
• Quantifying performance uncertainties in complex, dynamic environments is a critical challenge for ADS safety validation.
• Black box nature of learning algorithms in ADS complicates safety evaluation and public trust.

Purpose of the Study:

• To introduce a novel Polynomial Chaos Expansion (PCE) approach for quantifying positional uncertainties in ADS object detection.
• To enable online self-updating of uncertainty quantification to adapt to changing operational conditions.
• To enhance the reliability and trustworthiness of autonomous driving systems.

Main Methods:

• Utilized Polynomial Chaos Expansion (PCE) to model and quantify positional uncertainties from ADS object detection.
• Integrated High Definition (HD) maps to provide accurate spatial context for uncertainty analysis.
• Developed an online self-updating mechanism within the PCE framework to handle data shifts.
• Validated the PCE approach through both simulation and real-world driving experiments.

Main Results:

• The PCE method demonstrated superior accuracy in uncertainty quantification compared to baseline models.
• The self-updating capability of the PCE approach proved effective in adapting to changing environmental conditions, such as weather.
• Accurate uncertainty quantification is vital for robust ADS safety assessment.

Conclusions:

• The proposed PCE approach offers a robust solution for quantifying uncertainties in ADS object detection.
• The self-updating feature enhances the adaptability and reliability of ADS in real-world, dynamic environments.
• This work contributes to building public trust and facilitating the widespread adoption of autonomous driving technology.