作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (8): 251-259,270. doi: 10.19678/j.issn.1000-3428.0058544

• 图形图像处理 • 上一篇    下一篇

基于改进特征金字塔网络的人体姿态估计

王柳程, 欧阳城添, 梁文   

  1. 江西理工大学 信息工程学院, 江西 赣州 341000
  • 收稿日期:2020-06-04 修回日期:2020-07-17 发布日期:2020-07-23
  • 作者简介:王柳程(1995-),男,硕士,主研方向为人体姿态估计;欧阳城添,副教授、博士;梁文,硕士研究生。
  • 基金资助:
    江西省教育厅科学技术研究项目(GJJ170518);江西省研究生创新专项资金(YC2018-S331)。

Human Pose Estimation Based on Improved Pyramid Feature Network

WANG Liucheng, OUYANG Chengtian, LIANG Wen   

  1. School of Information and Engineering, Jiangxi University of Science and Technology, Ganzhou, Jiangxi 341000, China
  • Received:2020-06-04 Revised:2020-07-17 Published:2020-07-23

摘要: 为在人体姿态估计过程中有效获取多尺度特征和关键点坐标,建立一种基于改进特征金字塔网络(IPFN)的人体姿态跟踪模型。在原特征金字塔上采用新的检测器扩大感受野得到新特征金字塔,通过引入多尺度卷积生成高斯热点图,同时搜索和定位关键点,使坐标转换层将高斯热点图转为坐标,实现端到端训练过程。实验结果表明,相比FPN, IPFN模型在MPⅡ数据集的PCKh和COCO数据集的AP上分别提高了2.05和3.20个百分比;在踝、膝、腕和肘4个难检测部位上的PCKh分别提高了3.95、2.80、2.52和2.05个百分点。

关键词: 人体姿态估计, 特征金字塔, 坐标变换, 多尺度卷积, 高斯热点图

Abstract: To accurately acquire multi-scale features and keypoint coordinates in human pose estimation, a human pose tracking model based on Improved Pyramid Feature Network(IPFN) is established. On the basis of the original feature pyramid, a new detector is used to expand the receptive field for the new feature pyramid. Then gaussian heatmaps are generated by using multi-scale convolutions to search and locate the keypoints. The heatmaps are converted into coordinates by the coordinate transformation layer to realize end-to-end learning. Experimental results show that compared with traditional FPN, the IPFN model improves PCKh by 2.05 percentage point on the MPⅡ data set and AP by 3.2 percentage point on the COCO data set. The proposed model also improves PCKh by 3.95 percentage point for the ankle part, 2.80 percentage point for the knee part, 2.52 percentage point for the wrist part, and 2.05 percentage point for the elbow part.

Key words: human pose estimation, feature pyramid, coordinate transformation, multi-scale convolution, Gaussian heatmap

中图分类号: