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

Abstract

To build a natural driving environment in the virtual environment for autonomous driving safety testing, a digital twin model of human driving behavior is essential. However, due to the lack of fidelity and intelligence of the driving behavior model of the background vehicle, there is an intolerable gap between the virtual simulation test environment and the real road test environment. This paper proposes the Multi-source Information Encoding Transformer (MSIE-Transformer) to model driving behaviors of background vehicles within the virtual simulation environment. This approach improves model performance through the effective encoding of multi-source features using heterogeneous encoding networks, the comprehensive integration of these features based on the multi-head self-attention mechanism, and the combination of dynamic loss functions with Bayesian optimization. The experimental results demonstrate that, benefiting from the feature extraction and integration of multi-source information, the proposed method exhibits superior performance in fidelity compared to existing approaches. It also demonstrates good performance in statistical realism and modeling heterogeneous driving behaviors. In addition, the model’s performance is validated in multi-agent control scenarios and successfully transferred to intersection scenarios.