Zheng, Liyong (Tongji University), Rao, Hongyu (Hangzhou Hikvision Digital Technology Co. Ltd), Zhu, Jiang (Hangzhou Hikvision Digital Technology Co. Ltd), Jiang, Weihao (Hangzhou Hikvision Digital Technology Co. Ltd), Hao, Yonggang (Hangzhou Hikvision Digital Technology Co. Ltd), Zhao, Wei (Hangzhou Hikvision Digital Technology Co. Ltd), Shao, Jianxuan (Hangzhou Hikvision Digital Technology Co. Ltd), Su, Bin (Hangzhou Hikvision Digital Technology Co. Ltd), Li, Wenjing (Hangzhou Hikvision Digital Technology Co. Ltd), Xie, Yijue (Hangzhou Hikvision Digital Technology Co. Ltd), Ni, Wenke (Hangzhou Hikvision Digital Technology Co. Ltd), Sheng, Boheng (Hangzhou Hikvision Digital Technology Co. Ltd)

Vehicle Trajectory Imputation at Intersection Based on Ensemble Learning

Scheduled for presentation during the Regular Session "S08c-Intelligent Modeling and Prediction of Traffic Dynamics" (WE-LA-T8), Wednesday, November 19, 2025, 16:40−17:00, Coolangata 2

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 19, 2025

Abstract

Vehicle trajectory-based applications in intelligent transportation systems face data missing problems. This issue particularly affects traffic state estimation and traffic signal control. In recent years, the radar-and-video-fused perception technology has much more facilitated the vehicle trajectory data collection. But the data collecting is often hampered by radar's insensitivity to stationary targets and the occlusion by large vehicles, especially at intersections. To address the data missing issue, most of the existing studies adopted either knowledge-driven or data-driven methods, both of which had advantages and disadvantages. This paper proposed a novel vehicle trajectory imputation algorithm integrating the knowledge-driven Full Velocity Difference model (FVD) and the data-driven Transformer model (TF) based on ensemble learning (EL) (EL-based TF-FVD). The EL-based TF-FVD model leveraged the knowledge distillation to complete model training and integration. A traffic signal encoding module was added to account for the traffic rule constraints on vehicle movement. Experimental results in a radar-video-fused trajectory dataset showed the introduction of the FVD model and the traffic signal encoding module led to accuracy improvements by 16.1% and 11.4% respectively. In the SinD dataset, EL-based TF-FVD model achieved a 34.7% accuracy improvement compared to other data-driven algorithms, and a 12.59% improvement compared to the Physics-Informed Deep Learning based TF-FVD model. Also, a 9.14% decrease on the error of travel delay distribution calculated from the imputed trajectories, which implies its value in following applications.