Online Quantitative Analysis of Perception Uncertainty Based on High-Definition Map.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous driving relies heavily on environmental perception for safe decision-making.
• Evaluating perception uncertainty in real-time is a critical challenge for autonomous system safety and industrial adoption.
• High-definition maps provide valuable prior information on static road elements and their relationships, aiding uncertainty assessment.

Purpose of the Study:

• To develop and present a method for online evaluation of perception uncertainty for both static and dynamic elements in autonomous driving.
• To leverage high-definition maps to improve the accuracy and reliability of perception uncertainty assessment.
• To identify real-time factors influencing perception uncertainty to optimize assessment models.

Main Methods:

• Assessed static element perception uncertainty using spatial and topological features from high-definition maps, considering existence and spatial information.
• Developed an online model to assess dynamic element perception uncertainty, inferring it from static element uncertainty.
• Constructed a model to recognize driving scenarios and weather conditions to identify real-time uncertainty triggers.

Main Results:

• The proposed method effectively evaluates real-time perception results by assessing uncertainties of both static and dynamic elements.
• Verification on the nuScenes dataset demonstrated the efficacy of the high-definition map-based uncertainty assessment.
• The system successfully identified driving scenarios and weather conditions impacting perception uncertainty.

Conclusions:

• The developed method provides a robust approach for online perception uncertainty evaluation in autonomous driving.
• Utilizing high-definition maps significantly enhances the assessment of static element uncertainty, which in turn improves dynamic element uncertainty evaluation.
• Real-time identification of influencing factors allows for adaptive optimization of perception uncertainty models, contributing to safer autonomous operations.