汽车工程 ›› 2026, Vol. 48 ›› Issue (3): 627-637.doi: 10.19562/j.chinasae.qcgc.2026.03.013

• • 上一篇    

基于贝叶斯网络的行人穿越不确定性量化研究

杨彪1,杜梦沄1,朱俊瑞1,王海2,蔡英凤3()   

  1. 1.常州大学王诤微电子学院集成电路产业学院,常州 213159
    2.江苏大学汽车与交通工程学院,镇江 212013
    3.江苏大学汽车工程研究院,镇江 212013
  • 收稿日期:2025-04-08 修回日期:2025-05-07 出版日期:2026-03-25 发布日期:2026-03-19
  • 通讯作者: 蔡英凤 E-mail:1000004192@ujs.edu.cn
  • 基金资助:
    国家自然科学基金面上项目(62576052)、江苏省科技厅基金(BK20250969)、常州市应用基础项目(CJ20240039)和常州市领军型创新人才引进培育项目(CQ20250044)资助。

Pedestrian Crossing Intention Prediction Based on Bayesian Neural Network Quantifying Uncertainty

Biao Yang1,Mengyun Du1,Junrui Zhu1,Hai Wang2,Yingfeng Cai3()   

  1. 1.School of Wang Zheng Microelectronics and Integrated Circuit Industry,Changzhou University,Changzhou 213159
    2.School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang 212013
    3.Institute of Automotive Engineering,Jiangsu University,Zhenjiang 212013
  • Received:2025-04-08 Revised:2025-05-07 Online:2026-03-25 Published:2026-03-19
  • Contact: Yingfeng Cai E-mail:1000004192@ujs.edu.cn

摘要:

随着自动驾驶技术的发展,行人穿越意图预测已成为减少人车冲突的重要手段。然而,传统意图预测方法无法估计预测结果的不确定性,导致车辆在复杂交通环境中的决策行为缺乏可信度。针对上述问题,本文提出了一种基于多模态输入的不确定性行人穿越预测网络(UN-PCPNet),将行人姿态、边界框与车速3种输入分别提取特征后送入特征融合模块进行多模态融合,进而通过贝叶斯多层感知模块输出穿越意图预测结果,并通过结果的方差分析量化预测过程的不确定性。本文方法在JAAD和PIE公共数据集上可以实现92%和89%的AUC得分,并且可以在不影响预测准确性的前提下获得可靠的不确定性估计。实车实验也验证了本文方法在实际交通场景中的有效性,可以为提高自动驾驶安全性提供支持。

关键词: 贝叶斯神经网络, 行人穿越意图预测, 多模态融合, 不确定性估计

Abstract:

With the development of autonomous driving technology, pedestrian crossing intention prediction has become an important approach to reducing pedestrian-vehicle conflict. However, traditional intention prediction methods fail to estimate the uncertainty of the prediction results, which leads to a lack of reliability in vehicle decision-making in complex traffic environment. To address this issue, this paper proposes an uncertainty pedestrian crossing prediction network (UN-PCPNet) with multimodal input. The network extracts features from the three types of input of pedestrian pose, bounding box, and vehicle speed, which are then sent to a feature fusion module for multimodal integration. The Bayesian multilayer perceptron module outputs the crossing intention prediction result, and the uncertainty of the prediction process is quantified through the variance analysis of the results. The proposed method achieves AUC scores of 92% and 89% on the JAAD and PIE public datasets, respectively, and provides reliable uncertainty estimation without compromising prediction accuracy. The real-world experiments also validate the effectiveness of this method in practical traffic scenarios, which can support the enhancement of autonomous driving safety.

Key words: Bayesian neural networks, pedestrian crossing intention prediction, multi-modal fusion, uncertainty estimation