东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (11): 1529-1536.DOI: 10.12068/j.issn.1005-3026.2024.11.002

• 信息与控制 • 上一篇    下一篇

基于图卷积和卷积的行人轨迹预测算法

冯昂1, 宫俊1(), 王念2, 王景龙3   

  1. 1.东北大学 信息科学与工程学院,辽宁 沈阳 110819
    2.东风汽车集团有限公司 技术中心,湖北 武汉 430000
    3.燕山大学 电气工程学院,河北 秦皇岛 066004
  • 收稿日期:2023-06-05 出版日期:2024-11-15 发布日期:2025-02-24
  • 通讯作者: 宫俊
  • 作者简介:冯 昂(1999-),男,海南三亚人,东北大学硕士研究生
    宫 俊(1972-),男,江苏徐州人,东北大学副教授,硕士生导师.
  • 基金资助:
    国家自然科学基金资助项目(61871106)

Pedestrian Trajectory Prediction Algorithm Based on Graph Convolution and Convolution

Ang FENG1, Jun GONG1(), Nian WANG2, Jing-long WANG3   

  1. 1.School of Information Science & Engineering,Northeastern University,Shenyang 110819,China
    2.Technology Center,Dongfeng Motor Group Co. ,Ltd. ,Wuhan 430000,China
    3.College of Electrical Engineering,Yanshan University,Qinhuangdao 066004,China.
  • Received:2023-06-05 Online:2024-11-15 Published:2025-02-24
  • Contact: Jun GONG
  • About author:GONG Jun, E-mail: gongjun@ise.neu.edu.cn

摘要:

行人轨迹预测取得了重要的进展,但现有的方法大多会受限于有限的车载计算资源,如何在自动驾驶车辆上实现高效的行人轨迹预测仍然存在着不足.针对该问题,提出了一种轻量化的行人轨迹预测算法,使用卷积神经网络(convolutional neural network,CNN)和图卷积神经网络(graph convolutional neural network,GCN)来处理和融合多模态信息.首先基于CNN设计了多尺度特征处理模块,使用多个卷积模块捕获行人轨迹和场景信息在不同时间和空间尺度上的特征;然后基于GCN构造特征融合模块,用于高效地建立轨迹和场景特征之间的时空关系并获得多个预测表示,最后融合多个预测表示以获得行人轨迹预测结果.在PIE和JAAD数据集上的实验表明,所提方法在仅用最少网络参数的情况下取得了最佳的预测性能,验证了所提方法的有效性;对比先前最轻量化的方法,参数优化了73%.

关键词: 自动驾驶, 行人轨迹预测, 图卷积, 多尺度, 轻量化模型

Abstract:

Significant progress has been made in pedestrian trajectory prediction, but most of the existing methods are constrained by the limited on?board computing resources. How to achieve efficient pedestrian trajectory prediction in autonomous vehicles is still insufficient. To solve this problem, a lightweight pedestrian trajectory prediction algorithm is proposed, which uses convolutional neural network (CNN) and graph convolutional neural network (GCN) to process and integrate multimodal information. Firstly, a multi?scale feature processing module is designed based on CNN. Multiple convolution modules are used to capture the features of pedestrian tracks and scene information at different time and spatial scales. Then, a feature integration module is constructed based on GCN, which is used to efficiently integrate the spatial?temporal relationship between trajectory and scene features and obtain multiple prediction representations. Finally, multiple prediction representations are integrated to obtain pedestrian trajectory prediction results. Experiments on PIE and JAAD datasets show that the proposed method achieves competitive and optimal prediction performance with the least network parameters, respectively, which verifies the effectiveness of the proposed method. Compared with the previous lightest method, the parameters are optimized by 73%.

Key words: autonomous driving, pedestrian trajectory prediction, graph convolution, multi?scale features, lightweight model

中图分类号: