《计算机应用》唯一官方网站 ›› 2024, Vol. 44 ›› Issue (6): 1965-1971.DOI: 10.11772/j.issn.1001-9081.2023060897

• 前沿与综合应用 • 上一篇    

基于双支路卷积网络的步态识别方法

王晓路, 千王菲()   

  1. 西安科技大学 通信与信息工程学院,西安 710054
  • 收稿日期:2023-07-11 修回日期:2023-08-25 接受日期:2023-08-31 发布日期:2023-09-14 出版日期:2024-06-10
  • 通讯作者: 千王菲
  • 作者简介:王晓路(1977—),男,四川广安人,副教授,博士,主要研究方向:物联网、人工智能;
  • 基金资助:
    西安市科技计划项目(2020KJRC0070)

Gait recognition method based on two-branch convolutional network

Xiaolu WANG, Wangfei QIAN()   

  1. College of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an Shaanxi 710054,China
  • Received:2023-07-11 Revised:2023-08-25 Accepted:2023-08-31 Online:2023-09-14 Published:2024-06-10
  • Contact: Wangfei QIAN
  • About author:WANG Xiaolu, born in 1977, Ph. D., associate professor. His research interests include internet of things, artificial intelligence.
  • Supported by:
    Xi’an Science and Technology Plan Project(2020KJRC0070)

摘要:

针对步态识别易受拍摄视角、外观变化等影响的问题,提出一种基于双支路卷积网络的步态识别方法。首先,提出随机裁剪随机遮挡的数据增强方法RRDA(Restricted Random Data Augmentation),以扩展外观变化的数据样本,提高模型遮挡的鲁棒性;其次,采用结合注意力机制的两路复合卷积层(C-Conv)提取步态特征,一个分支通过水平金字塔映射(HPM)提取行人外观全局和最具辨识度的信息;另一分支通过多个并行的微动作捕捉模块(MCM)提取短时间的步态时空信息;最后,将两个分支的特征信息相加融合,再通过全连接层实现步态识别。基于平衡样本特征的区分能力和模型的收敛性构造联合损失函数,以加速模型的收敛。在CASIA-B步态数据集上进行实验,所提方法在3种行走状态下的平均识别率分别达到97.40%、93.67%和81.19%,均高于GaitSet方法、CapsNet方法、双流步态方法和GaitPart方法;在正常行走状态下比GaitSet方法的识别准确率提升了1.30个百分点,在携带背包状态下提升了2.87个百分点,在穿着外套状态下提升了10.89个百分点。实验结果表明,所提方法是可行、有效的。

关键词: 步态识别, 双支路卷积网络, 注意力机制, 金字塔映射, 深度学习

Abstract:

Aiming at the problem that gait recognition is easily affected by changes in shooting angle and appearance, a gait recognition method based on a two-branch convolutional network was proposed. Firstly, a data augmentation method of random cropping and random occlusion, named RRDA(Restricted Random Data Augmentation), was proposed to expand the data samples of appearance changes and improve the robustness of model occlusion. Secondly, the attention mechanism was used to form a two-branch Composite-Convolutional (C-Conv) layer to extract gait features. One branch network extracted the global and most recognizable information of pedestrian appearance through Horizontal Pyramid Mapping (HPM); the other branch used multiple parallel Micro-Motion Capture Modules (MCMs) to extract short-term gait spatio-temporal information. Finally, the feature information of the two branches was added and fused, and then the gait recognition was achieved through a fully connected layer. A joint loss function was constructed based on the discriminative ability of balanced sample features and the convergence of the model to accelerate the convergence of the model. Experiments were conducted on the gait recognition dataset CASIA-B, the recognition accuracies of the proposed method in three states of walking are 97.40%, 93.67% and 81.19%, which are higher than those of GaitSet method, CapsNet method, two-stream gait method and GaitPart method; compared to GaitSet method, the recognition accuracy of the proposed method is 1.30 percentage points higher in the state of normal walking, 2.87 percentage points higher on carrying backpack, and 10.89 percentage points higher on wearing jacket. Experimental results show that the proposed method is feasible and effective.

Key words: gait recognition, two-branch convolutional network, attention mechanism, pyramid mapping, deep learning

中图分类号: