基于多尺度对抗学习的人体姿态估计  

Human Pose Estimation Based on Multi-scale Adversarial Learning

作　　者：李晓明 黄慧[1] 应毅[1] 徐建华 曾岳[2] 刘洋[3] LI Xiaoming;HUANG Hui;YING Yi;XU Jianhua;ZENG Yue;LIU Yang(School of Computer Science and Engineering,Sanjiang University,Nanjing 210012,China;School of Software Engineering,Jinling Institute of Technology,Nanjing 211169,China;Jiangsu Qingjiang Middle School,Huai’an 223001,Jiangsu,China)

机构地区：[1]三江学院计算机科学与工程学院,南京210012 [2]金陵科技学院软件工程学院,南京211169 [3]江苏省清江中学,江苏淮安223001

出　　处：《实验室研究与探索》2020年第11期35-39,共5页Research and Exploration In Laboratory

基　　金：国家自然科学基金项目(61262002);江苏省高等学校自然科学研究项目(19KJB520049,18KJB520042,17KJB520033,2019SJKY006,2018SJA0734)。

摘　　要：为解决现有基于对抗学习的人体姿态估计算法对人体尺度变化敏感问题,提出了一种基于多尺度对抗学习的人体姿态估计模型。该模型通过堆栈沙漏网络分别构建对抗网络中的生成器和判别器,再分别在生成器和判别器中引入多尺度监督损失函数,从而提高整个网络检测不同尺度人体部件的性能。为证明方法的有效性,采用了LSP、LSPE、FLIC和MPII 4个数据集进行人体姿态估计模型的训练与测试。结果表明,引入多尺度信息的人体姿态估计模型可提高0.7%的检测正确率;与现有的部分人体姿态估计算法相比,多尺度对抗学习算法具有更高的估计正确率。In order to solve the problem that the existing adversarial learning based human pose estimation algorithms are sensitive to the change of human body scale,a multi-scale adversarial learning is proposed.In this model,the generator and discriminator in the adversarial network are built separately by the stack hourglass networks.Then a multi-scale supervision loss function is introduced in the generator and discriminator respectively,it improves the performance of the whole network to detect human body parts of different scales.In order to demonstrate the effectiveness of the proposed method,the LSP,LSPE,FLIC and MPII datasets are used to train and test the model.The experimental results show that the human pose estimation model with multi-scale information can improve the detection accuracy by 0.7%;compared with some existing human pose estimation algorithms,the multi-scale adversarial learning algorithm has higher estimation accuracy.

关 键 词：人体姿态估计 对抗学习 多尺度监督 沙漏网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]