Assessing autonomous driving performance and environmental influencing factors using real-world operational...

Area of Science:

• Transportation Engineering
• Human-Computer Interaction
• Robotics

Background:

• Commercial deployment of autonomous vehicles (AVs) is increasing, yet large-scale empirical data comparing AVs with human driving is scarce.
• Understanding AV performance and human-machine interaction is crucial for safe integration into public roadways.

Purpose of the Study:

• To empirically compare autonomous vehicle (AV) and human driving performance using naturalistic data.
• To analyze longitudinal dynamics, ride comfort, and human-machine interaction across different operational regimes.
• To investigate the characteristics and environmental associations of AV takeover events.

Main Methods:

• Analysis of over 100 million trajectory records from 444 vehicles across Beijing and Shanghai demonstration zones.
• Hierarchical segmentation pipeline to process AV/manual mode switches and identify stable driving segments.
• Classification of in-motion takeover events (Smooth, Braking, Accelerating) and exposure-adjusted SHAP analysis for environmental associations.

Main Results:

• Routine jerk levels are comparable between AVs and human drivers; AV jerk is significantly lower during active maneuvers in Shanghai.
• Takeover event types differ by city: Braking dominates Shanghai, while Smooth transitions are more common in Beijing.
• Environmental factors like road curvature and intersection proximity are distinctly associated with different takeover types.

Conclusions:

• Total disengagement counts obscure crucial behavioral differences in AV interventions.
• Analyzing takeover events by post-takeover dynamics and operational regime is essential for a nuanced understanding of AV performance.
• Findings provide insights for improving AV safety and human-machine interaction in diverse driving environments.