计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (8): 96-103.DOI: 10.3778/j.issn.1002-8331.1910-0386

• 模式识别与人工智能 • 上一篇    下一篇

利用改进生成对抗网络进行人体姿态识别

吴春梅,胡军浩,尹江华   

  1. 1.广西科技师范学院 数学与计算机科学学院,广西 来宾 546199
    2.中南民族大学 数学与统计学学院,武汉 430074
    3.内蒙古大学 数学科学学院,呼和浩特 010021
  • 出版日期:2020-04-15 发布日期:2020-04-14

Using Improved Generative Adversarial Network for Human Pose Estimation

WU Chunmei, HU Junhao, YIN Jianghua   

  1. 1.School of Mathematics and Computer Science, Guangxi Science & Technology Normal University, Laibin, Guangxi 546199, China
    2.College of Mathematics and Statistics, South Central University for Nationalities, Wuhan 430074, China
    3.School of Mathematical Sciences, Inner Mongolia University, Hohhot 010021, China
  • Online:2020-04-15 Published:2020-04-14

摘要:

针对人体模型中某些重要关节点准确定位的问题,提出了一种新型深度卷积生成对抗网络以进行静态图像中人体姿态的估计的方法。该方法采用了深度卷积的堆叠沙漏网络来准确提取图像上关键关节点的位置,该网络的生成和辨别部分被设计用于编码第一层次结构(亲本)与第二层次结构(子本)中的空间关系,并且展示了人体部位的空间层次。生成器和判别器在网络中被设计为两部分,并按照顺序连接在一起用来编码外观可能的关系,同时为人体部位存在的可能性以及身体的每个部分与其亲本部分之间的关系进行编码。在静态图像中,可以较准确地识别人体模型关键节点以及大致人体姿态。该方法在不同的数据集上进行了实验,在大部分情况下,提出的方法获得的结果优于其他几种对比方法。

关键词: 卷积神经网络, 生成对抗网络(GAN), 人体姿态识别, 堆叠沙漏网络, 层次感知

Abstract:

Aiming at the problem of accurately locating some important joint points in human body modeling, a novel deep convolution generative adversarial network is proposed to estimate human poses from still images. Specifically, a stacked hourglass network with deep convolution network is used to accurately locate the key joint points in the images, the generative and the discriminative portions of the proposed adversarial deep neural network are designed to encode the spatial relationship between the parts in the first stage of the hierarchy(parents) and the parts in the second stage of the hierarchy(children). Each of the generator and the discriminator networks is designed as two components, which are sequentially connected together to infer rich appearance potentials and to encode not only the likelihood of the part’s existence but also the relationships between each body part and its parent. Finally, in the static images, the key nodes of the human body model and the approximate human poses can be estimated accurately. The method is tested on different data sets, in most cases, the results of the proposed method are better than those of other methods.

Key words: convolutional neural network, Generative Adversarial Network(GAN), human pose estimation, stacked hourglass network, hierarchy-aware