弹道
领域(数学)
计算机科学
工程类
控制理论(社会学)
控制工程
车辆动力学
加速度
势场
雷达跟踪器
卡尔曼滤波器
实时计算
控制系统
作者
Dehai Zhang,Mingzhe Zhang,Ping Wang,Fengxiang Guo,Bowen Chen,Zheng Chen
出处
期刊:IEEE Transactions on Vehicular Technology
[Institute of Electrical and Electronics Engineers]
日期:2026-08-18
卷期号:75 (11): 26234-26246
标识
DOI:10.1109/tvt.2026.3724901
摘要
Accurate trajectory prediction remains challenging because existing models heavily rely on the homogeneity assumption, failing to capture how diverse driving styles dynamically reshape spatial risks and often leading to overconfident forecasts in dense traffic. This study proposes the style-adaptive graph embedding network (SAGE), a style-informed framework that modulates interaction modeling for multi-agent trajectory forecasting. Specifically, the SAGE framework operates through three sequential components. First, it employs an unsupervised driving style recognition module to capture individual behavioral preferences, providing a structural prior for feature extraction. Second, a style-modulated driver's risk field (SM-DRF) translates these abstract styles into dynamically evolving risk representations for fine-grained interaction modeling. Finally, within the multimodal decoder, a heteroscedastic uncertainty optimization mechanism aligns variance estimation with predictive uncertainty, thereby enhancing probabilistic calibration under complex multi-agent scenarios. Evaluations on the Argoverse 2 validation set under a 6 s multi-modal prediction horizon demonstrate that SAGE yields$minADE_{6}$of 0.656 m, indicating consistent improvements in trajectory accuracy over state-of-the-art methods. These results indicate that combining behavioral priors with physically grounded interaction modeling provides a robust framework for trajectory forecasting in complex urban environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI