电子学报 ›› 2022, Vol. 50 ›› Issue (7): 1537-1547.DOI: 10.12263/DZXB.20210762

• 学术论文 •    下一篇

基于Transformer动态场景信息生成对抗网络的行人轨迹预测方法

裴炤1,2, 邱文涛2, 王淼3(), 马苗2, 张艳宁4,5   

  1. 1.陕西师范大学现代教学技术教育部重点实验室, 陕西 西安 710119
    2.陕西师范大学计算机科学学院, 陕西 西安 710119
    3.上海交通大学航空航天学院, 上海 200240
    4.空天地海一体化大数据应用技术国家工程实验室, 陕西 西安 710129
    5.西北工业大学计算机学院, 陕西 西安 710129
  • 收稿日期:2021-06-16 修回日期:2021-11-07 出版日期:2022-07-25
    • 通讯作者:
    • 王淼
    • 作者简介:
    • 裴 炤 男,1983年2月生,陕西西安人,博士、教授、博士生导师.主要从事计算机视觉与人工智能、图像处理与模式识别、机器学习的相关研究.E-mail:zpei@snnu.edu.cn
      邱文涛(1984-)男,1996年8月生,山东枣庄人.陕西师范大学计算机科学学院研究生,主要研究方向为计算机视觉和行人轨迹预测.E-mail:qiuwentao@snnu.edu.cn
      王淼男,1981年7月生,河南义马人,博士,上海交通大学航空航天学院助理研究员,主要研究方向为智能信息处理、数据挖掘、计算机视觉.E-mail:miaowang@sjtu.edu.cn
    • 基金资助:
    • 国家自然科学基金 (61971273); 陕西省重点研发计划 (2021GY-032); 中央高校基本科研业务 (GK202003077); 上海市自然科学基金 (20ZR1427800)

Pedestrian Trajectory Prediction Method Using Dynamic Scene Information Based Transformer Generative Adversarial Network

PEI Zhao1,2, QIU Wen-tao2, WANG Miao3(), MA Miao2, ZHANG Yan-ning4,5   

  1. 1.Key Laboratory of Modern Teaching Technology(Ministry of Education), Shaanxi Normal University, Xi’an, Shaanxi 710119, China
    2.School of Computer Science, Shaanxi Normal University, Xi’an, Shaanxi 710119, China
    3.School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China
    4.National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, Xi’an, Shaanxi 710129, China
    5.School of Computer Science, Northwestern Polytechnical University, Xi’an, Shaanxi 710129, China
  • Received:2021-06-16 Revised:2021-11-07 Online:2022-07-25 Published:2022-07-30
    • Corresponding author:
    • WANG Miao
    • Supported by:
    • National Natural Science Foundation of China (61971273); Key R&D Project of Shaanxi Province (2021GY-032); Fundamental Research Funds for the Central Universities (GK202003077); National Natural Science Foundation of Shanghai Municipality, China (20ZR1427800)

摘要:

行人轨迹预测是视频监控的重要组成部分,因现有方法未充分利用场景特征信息造成其预测轨迹不符合生活常识,导致行人轨迹预测精度较低出现明显偏离真实轨迹的情况.针对上述不足本文提出一种基于Transformer动态场景信息生成对抗网络(Generative Adversarial Network,GAN)的行人轨迹预测方法.该方法利用动态场景特征提取模块的卷积神经网络(Convolutional Neural Networks,CNN)模型对目标行人的动态场景信息进行特征提取,同时生成器网络中的编码器利用Transformer对行人的社会交互信息特征以及轨迹信息特征进行建模.在ETH和UCY数据集上的实验结果表明,与Social GAN模型相比,本文方法在多个场景下的平均位移误差准确率提高了25.61%,最终位移误差准确率提高了38.44%.

关键词: 行人轨迹预测, 生成对抗网络, 转换器, 深度学习, 长短期记忆网络

Abstract:

Pedestrian trajectory prediction is an important part of video surveillance. The current methods are not accurate and sometimes violate common senses because scene information is not fully used. To eliminate the above shortcomings, this paper proposes a transformer generated adversarial network(GAN) algorithm which combines dynamic scene information with pedestrian social interaction information. The convolution neural network model of the dynamic scene extraction module is utilized to extract the dynamic scene information features of the target pedestrian, and the encoder in the generator network uses transformer to model the features of social interaction information and trajectory information of pedestrians. Experimental results on ETH and UCY datasets show that, compared with social GAN model, our method improves the accuracy of average displacement error by 25.61% and the accuracy of average final displacement error by 38.44% in multiple scenarios.

Key words: pedestrian trajectory prediction, generative adversarial networks, transformer, deep learning, long short-term memory

中图分类号: