结合二阶注意力机制的多尺度人体姿态估计  

Multi-scale Human Pose Estimation Combined with Second-order Attention Mechanism

作　　者：张云绚 董绵绵[1] 王鹏[1] 李晓艳[1] 吕志刚[1] 邸若海 毋宁 ZHANG Yun-xuan;DONG Mian-mian;WANG Peng;LI Xiao-yan;Lü Zhi-gang;DI Ruo-hai;WU Ning(School of Electronic and Information Engineering,Xi'an Technological Lniversity,Xi'an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院,西安710021

出　　处：《科学技术与工程》2022年第32期14321-14327,共7页Science Technology and Engineering

基　　金：国家自然科学基金(62171360);陕西省科技厅重点研发计划(2022GY-110);西安工业大学校长基金面上培育项目(XGPY200217);2022年度陕西高校青年创新团队项目。

摘　　要：为解决人体姿态估计任务中存在的不同视角下人体实例尺度变化、遮挡问题导致的人体关键点定位不准确问题,提出融入二阶注意力机制的多尺度人体姿态估计网络模型GOS-HRNet。首先,在特征提取阶段为了获得高质量的特征图,通过在多分辨率网络结构中使用Octave卷积,保留更多的图像空间特征信息以提高关键点定位准确率;然后,为有效的利用图像上下文信息,融入二阶注意力模块使网络能更好地学习各分辨率表征的空间信息;最后,为了应对尺度变换对关键点定位的影响采用尺度增强训练方法,提高模型对尺度变化的鲁棒性。所提模型在MS COCO 2017数据集上进行实验,结果表明:所提出的GOS-HRNet模型平均检测精度比HRNet模型提升了2.2%,能够更加准确地利用上下文信息、丰富空间特征信息以提高对关键点定位的准确性。In order to solve the problem of inaccurate positioning of the key points of the human body caused by the change of the human body instance scale and the occlusion problem in the human body pose estimation task in different perspectives,a multi-scale human body pose estimation network model(GOS-HRNet)incorporating a second-order attention mechanism was proposed.First of all,in the feature extraction stage,for obtaining high-quality feature maps,Octave convolution was used in the multi-resolution network structure to retain more image spatial feature information to improve the accuracy of key point positioning.Then,the second-order attention module was incorporated so that the network can better learn the spatial information represented by each resolution,and effectively use the image context information.Finally,so as to cope with the impact of scale transformation on the positioning of key points,a multi-scale training method was adopted to improve the generalization of the model to scale changes.A proposed model to conduct experiments on the MS COCO 2017 data set wap.The results show that the average detection accuracy of the proposed GOS-HRNet model is improved by 2.2%compared with the HRNet model.It can more accurately use context information and enrich spatial feature information to improve the key accuracy of point positioning.

关 键 词：多尺度 高质量特征图 姿态估计 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]