Lian, Bangan (Tongji University), Wang, Shihan (Tongji University), Zheng, Liyong (Tongji University), Zhu, Jiang (Hangzhou Hikvision Digital Technology Co. Ltd), Sun, Jian (Tongji University), Sun, Jie (Tongji University)

Decision-Making Modeling at Complex Mixed-Flow Intersections: A Hybrid CVAE-Transformer Framework

Scheduled for presentation during the Regular Session "S36c-Behavior Modeling and Decision-Making in Traffic Systems" (FR-LA-T36), Friday, November 21, 2025, 16:40−17:00, Surfers Paradise 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Human-like decision-making behaviors are critical for the rapid deployment of autonomous vehicles (AVs), where left-turning behavior under unprotected phase at mixed-flow intersection is one of the most challenging tasks. Given the non-strict priority, heterogeneous interaction participants, and temporal dependent decision-making process, existing models fail to adequately consider the characteristics. Furthermore, data-driven models usually lack interpretability of the decision-making processes. Therefore, this study aims to propose a decision-making model―DeCFormer at complex mixed-flow intersections, integrating conditional variational autoencoder (CVAE) and Transformer. Specifically, we first abstract the complex interactive environment using velocity and risk field models, which provide unified and effective representations for efficiency and safety. Then, we employ CVAE to embed the efficiency-safety input features, integrating with Transformer to model the spatiotemporal dependencies and generate coherent decision sequences. Moreover, we use the predicted features from CVAE to interpret the decision-making processes. Experiments on the Xianxia-Jianhe intersection data demonstrate that DeCFormer achieves superior performance compared to baseline models with decision accuracy of 85.66%, BLEU-4 score of 0.787, and ROUGE-L score of 0.85.