计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (4): 150-156.DOI: 10.3778/j.issn.1002-8331.2008-0355

• 模式识别与人工智能 • 上一篇    下一篇

基于非对称双路识别网络的步态识别方法

周潇涵,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
  • 出版日期:2022-02-15 发布日期:2022-02-15

Novel Gait Recognition Method Based on Asymmetric Two-Path Network

ZHOU Xiaohan, 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:2022-02-15 Published:2022-02-15

摘要: 步态作为一种人体躯干、关节、上下肢及各肌群的周期性行为模式,是可用于身份识别过程的一种重要生物特征。针对现有的步态识别方法大都是基于步态轮廓图或者步态能量图提取的全局特征,而忽视了对细粒度步态信息的有效利用的问题,提出了一种包括全局通路和局部通路的非对称双路识别网络。其中全局通路采用三元组损失函数,用于提取步态的全局时空特征;局部通路采用交叉熵损失函数,用于识别步态中显著不同的局部特征。此外,在局部通路中加入了一个显著性特征检测器模块,用于实现有效的细粒度步态信息识别。在公开数据集CASIA-B和OU-ISIR-LP上进行了对比实验,结果表明,在跨视角和跨场景的环境下,该方法相对现有方法在步态识别的准确率方面都有显著提升。

关键词: 步态识别, 非对称双路网络, 显著性特征检测器

Abstract: As a periodic behavior pattern of the human torso, joints, lower and upper limbs, gait is an important biological feature that can be used in the identification process. In view of the fact that most of existing gait recognition methods are based on the global features extracted from gait silhouettes or gait energy images and neglect the effective use of fine-grained gait information, an asymmetric two-path identification network including the global path and the local path is proposed. The global path uses a triple loss function to extract the global spatio-temporal features of gait, while the local path uses a cross-entropy loss function to identify significantly different local features in gait. In addition, a novel module named salient local feature detector is added to the local path for recognizing fine-grained gait information effectively. Finally, comparative experiments are conducted on public datasets CASIA-B and OU-ISIR-LP, the results show that, in the cross-view and cross-scenario environment, the proposed method has a significant improvement in recognition accuracy compared to the existing methods.

Key words: gait recognition, asymmetric two-path network, salient local feature detector