计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (24): 198-203.DOI: 10.3778/j.issn.1002-8331.1709-0045

• 图形图像处理 • 上一篇    下一篇

基于改进CNN和加权SVDD算法的人体姿态估计

韩贵金   

  1. 西京学院 信息工程学院,西安 710123
  • 出版日期:2018-12-15 发布日期:2018-12-14

Human pose estimation based on improved CNN and weighted SVDD algorithm

HAN Guijin   

  1. College of Information Engineering, Xijing University, Xi’an 710123, China
  • Online:2018-12-15 Published:2018-12-14

摘要: 卷积神经网络是人体姿态估计中应用最成功的深度学习模型,但仍存在着提取图像特征时不同图像区域和提取出的不同抽象特征被平等对待的缺陷。为此,提出了一种基于改进卷积神经网络和加权支持向量数据描述算法的关节外观模型,并用其设计了一种人体姿态估计算法。卷积神经网络卷积层中不同图像区域的卷积操作被赋以不同的权值系数以体现其不同作用;采用加权支持向量数据描述算法对每一种抽象特征都构造关节子外观模型,将所有关节子外观模型按不同权值进行线性组合建立了新的关节外观模型,权值的不同体现了不同抽象特征的不同作用。仿真实验表明,与现有基于卷积神经网络的人体姿态估计算法相比,所设计的人体姿态估计算法具有更高的估计准确度。

关键词: 人体姿态估计, 深度学习, 卷积神经网络, 加权支持向量数据描述, 线性组合

Abstract: The Convolutional Neural Network(CNN) is the most successful deep learning model for human pose estimation, but there are still have some deficiencies such as different image regions when extracting image features and different abstract features extracted by CNN are treated equally. For overcoming these two deficiencies, a joint appearance model based on improved CNN and weighted Support Vector Data Description(SVDD) algorithm is proposed, and a human pose estimation algorithm is designed. The convolutional operations for different image regions are assigned to different weight coefficients, which reflects their different effects. The weighted SVDD algorithm is used to build joint appearance sub-model for each abstract feature, and then a new joint appearance model is built by linear combination of all sub-models with different weights, which reflects different characteristics of different abstract features. The simulation results show that the proposed algorithm has higher estimation accuracy than the human pose estimation algorithms based on the traditional CNN.

Key words: human pose estimation, deep learning, Convolutional Neural Network(CNN), weighted support vector data description, linear combination