Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (3): 266-273.DOI: 10.3778/j.issn.1002-8331.2009-0332

• Graphics and Image Processing • Previous Articles     Next Articles

Gait Recognition Combined with Convolutional Neural Network with Attention and Part-Level Features

HU Shaohui, WANG Xiuhui   

  1. Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, College of Information Engineering, China Jiliang University, Hangzhou 310018, China
  • Online:2022-02-01 Published:2022-01-28

结合注意力卷积网络与分块特征的步态识别

胡少晖,王修晖   

  1. 中国计量大学 信息工程学院 浙江省电磁波信息技术与计量检测重点实验室,杭州 310018

Abstract: Currently, though deep learning algorithms have been widely used in the field of gait recognition, most existing methods extract global gait feature by convolutional neural networks, which ignores many local features with key gait information and to a certain extent, weakens the accuracy and improvement potential of gait recognition. To solve the above problems, a cross-view gait recognition method combined with convolutional neural network with attention and part-level features is proposed. By using gait silouettes as input, each gait frame passes through the same convolutional neural network with attention and forms the overall information, adding an effective attention mechanism CBAM into the network can show the importance of modeling the spaces and channels and increase the weight of significant regional features. Then the whole information is horizontally divided into two parts for training and gait recognition, the extracted gait local features are more suitable for fine gait classification. Finally, cross-view gait recognition experiments on the gait datasets CASIA-B and OU-ISIR-MVLP are carried out. The results show that the proposed method has a significant improvement effect compared with the existing methods under the condition of reorganization and insufficiency of the training dataset.

Key words: gait recognition, convolutional neural network, attention mechanism, part-level feature

摘要: 目前深度学习算法已经广泛应用于步态识别领域,但是大多数现有方法通过卷积神经网络提取步态全局特征时,忽略了许多包含关键步态信息的局部特征,在一定程度上削弱了步态识别的精度和提升潜力。针对上述问题,提出了一种结合注意力卷积神经网络与分块特征的跨视角步态识别方法,该方法以步态轮廓图序列为输入,每帧图片分别经过相同结构的注意力卷积神经网络融合成整体特征,在网络中加入有效的注意力机制CBAM能显式地建模各空间及通道的重要程度,增大显著区域特征的权重;整体特征被水平分成两块进行训练和步态识别,提取的步态局部特征更适合精细的步态分类。在步态数据集CASIA-B和OU-ISIR-MVLP上进行跨视角步态识别实验,结果证明在训练数据集充足与不足的条件下,该方法在识别精度上均优于现有方法。

关键词: 步态识别, 卷积神经网络, 注意力机制, 分块特征