Vehicle Lane-Changing scenario generation using time-series generative adversarial networks with an Adaptative...

Area of Science:

• Artificial Intelligence
• Autonomous Systems
• Transportation Engineering

Background:

• Connected and automated vehicles (CAVs) promise improved safety and efficiency.
• Rigorous testing across diverse scenarios is crucial for CAV deployment.
• Existing methods face challenges in scenario diversity and comprehensive evaluation.

Purpose of the Study:

• To propose a novel method for generating diverse vehicle lane-changing scenarios for CAV testing.
• To address challenges in test scenario diversity and evaluation comprehensiveness.
• To enhance the safety and reliability of autonomous driving technologies.

Main Methods:

• Utilized a time-series generative adversarial network (TimeGAN) with an adaptive parameter optimization strategy (APOS).
• Trained TimeGAN with a limited subset of parameter combinations (13.3%).
• Evaluated generated scenarios for diversity, fidelity, and utility; employed a Lane-Changing Risk Index (LCRI) to identify adversarial cases.

Main Results:

• Successfully trained a TimeGAN to generate a substantial number of lane-changing scenarios.
• Generated scenarios effectively captured a wide range of driving situations.
• The method generated 27 times more adversarial cases with 1.8 times higher average risk compared to real scenarios.

Conclusions:

• The proposed TimeGAN-based method enhances CAV test scenario generation.
• The approach is effective in uncovering critical safety vulnerabilities.
• This contributes to more comprehensive and reliable autonomous driving testing and technology.