基于上下文注意力机制的人体姿态估计网络  

Human Pose Estimation Network Based on Context Attention Mechanism

作　　者：吴樾 梁桥康[1,2] 孙炜 张柯毅[4] WU Yue;LIANG Qiaokang;SUN Wei;ZHANG Keyi(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;Hunan Key Laboratory of Intelligent Robot Technology in Electronic Manufacturing,Changsha 410082,China;Shenzhen Research Institute of Hunan University,Shenzhen 518055,China;Sichuan University-Pittsburgh Institute,Sichuan University,Chengdu 610207,China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082 [2]电子制造业智能机器人技术湖南省重点实验室,长沙410082 [3]湖南大学深圳研究院,深圳518055 [4]四川大学匹兹堡学院,成都610207

出　　处：《无人系统技术》2022年第6期74-85,共12页Unmanned Systems Technology

基　　金：国家重点研发计划(2022YFB4703103,2021YFC1910402);国家自然科学基金(NSFC.62073129,U21A20490);湖南省自然科学基金(2022JJ10020)。

摘　　要：传统基于深度神经网络的人体姿态估计方法主要集中于网络模型的设计,对关节点、人体结构信息的建模关注较少。从人体结构建模的角度来看,现有的人体姿态估计方法仍然存在不足。首先,为了提升网络对人体结构信息的感知能力,提出了一种基于上下文注意力机制的人体姿态估计网络,通过注意力机制对人体关节点之间的相对位置关系进行建模。其次,为了增强训练数据集的规模和质量,提出了基于语义分割的数据增强方法。实验结果表明,对关节点相对位置的建模可以提升神经网络的感知能力。使用数据增强策略能够生成大量难样本,增强了网络的泛化能力。所提出的方法在COCO验证集上的人体姿态估计精度为79.5%,在COCO test-dev数据集上的精度为76.7%,高于其他具有相近复杂度网络的精度。Previous research on human pose estimation methods based on deep neural networks has mainly focused on the design of network models, while less attention has been paid to the modelling of keypoints and body structure information. From the perspective of human structure modeling, the existing human pose estimation methods still have shortcomings. First, in order to enhance the network’s ability to perceive the human body structure information, a human pose estimation network based on the context attention mechanism is proposed. The relative position relationships between keypoints are modelled by the attention mechanism. Second, in order to enhance the size and quality of the training data set, a data augmentation method based on semantic segmentation is proposed. The experimental results show that modeling the relative positions of keypoints can enhance the perceptual ability of the neural network. The data augmentation strategy can generate a large number of difficult samples, which enhances the generalization ability of the network. The accuracy of the proposed method is 79.5% for human pose estimation on the COCO validation set and 76.7% on the COCO test-dev dataset, which are higher than the accuracy of other networks with similar complexity.

关 键 词：人体姿态估计 注意力机制 上下文信息 HRNet 特征融合 数据增强 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]