MSIE-Transformer: A novel driving behavior modeling approach for virtual simulation test environment.

Area of Science:

• Artificial Intelligence
• Autonomous Driving Systems
• Virtual Environment Simulation

Background:

• Accurate simulation of human driving behavior is crucial for autonomous driving safety testing.
• Existing virtual environments lack fidelity and intelligence in modeling background vehicle behaviors, creating a gap with real-world testing.

Purpose of the Study:

• To propose the Multi-source Information Encoding Transformer (MSIE-Transformer) for modeling background vehicle driving behaviors in virtual simulation environments.
• To enhance the fidelity and intelligence of driving behavior models for more realistic autonomous driving safety testing.

Main Methods:

• Utilizing heterogeneous encoding networks for effective encoding of multi-source features.
• Employing a multi-head self-attention mechanism for comprehensive feature integration.
• Combining dynamic loss functions with Bayesian optimization for improved model performance.

Main Results:

• The MSIE-Transformer demonstrates superior performance in fidelity compared to existing approaches.
• The model shows good performance in statistical realism and modeling heterogeneous driving behaviors.
• Validated performance in multi-agent control and intersection scenarios.

Conclusions:

• The proposed MSIE-Transformer effectively models background vehicle driving behaviors, significantly improving virtual simulation fidelity.
• This approach enhances the realism of virtual driving environments for autonomous vehicle testing.
• The model's adaptability to different scenarios shows its potential for broad application in autonomous driving research.