摘要
针对目前复杂的人体姿态估计网络参数量大、计算成本高以及读者对算法的理解分析难度大等问题,提出了一种基于注意力机制的人体姿态估计网络。该网络使用结构简单的编码器、解码器来预测人体关节点热图。为了使编码层提取的特征更具代表性,在编码层中加入了通道注意力机制,该操作在降低模型复杂度的同时保证了其预测精度。解码层采用多个反卷积模块得到最终的预测结果。算法模型在两个数据集(MPⅡ和COCO)上进行验证,在MPⅡ数据集上PCKh@0.5达到89%;在COCO数据集上,与CBA对比,虽然AP略低0.2个百分点,但单张图片推理速度提升了10.1 ms。实验结果表明,所提方法能够有效检测出人体关节点,并且优于各种先进的姿态估计方法。
At present,the complex human pose estimation network has many problems,such as large parameters,high calculation cost and difficulty for readers to understand and analyze the algorithm.To solve these problems,a human pose estimation network integrating attention mechanism is proposed.The network uses a simple encoder and decoder to predict the human joint point heat map.In order to make the features extracted in the coding layer more representative,the channel attention mechanism is added in the coding layer,which reduces the complexity of the model and ensures its prediction accuracy.The decoding layer uses several deconvolution modules to get the final prediction result.The algorithm model is verified on two data sets(MPⅡand COCO)PCKh@0.5 Up to 89%.On COCO dataset,compared with CBA,although AP is slightly lower by 0.2 percentage points,the reasoning speed of a single picture is increased by 10.1 ms.Experimental results show that the proposed method can effectively detect human joints,and is superior to most advanced pose estimation methods.
作者
方芹
缪宁杰
张如宏
刘晓泽
王佳敏
罗文东
周霖
Fang Qin;Miao Ningjie;Zhang Ruhong;Liu Xiaoze;Wang Jiamin;Luo Wendong;Zhou Lin(Innovation and Entrepreneurship Center of State Grid Zhejiang Electric Power Co.,Ltd.,Zhejiang Hangzhou,310051,China;State Grid Hangzhou Power Supply Company,Zhejiang Hangzhou,310009,China;Zhejiang Genius-pros Intelligent Technology Co.,Ltd.,Zhejiang Hangzhou,311100,China;Hangzhou Zhicheng Electronic Technology Co.,Ltd.,Zhejiang Hangzhou,310051,China;Beijing Dadaohechuang Technology Co.,Ltd.,Beijing 100085,China)
出处
《机械设计与制造工程》
2022年第3期117-122,共6页
Machine Design and Manufacturing Engineering
基金
国家电网有限公司总部管理双创孵化培育基金资助项目(SGZJSC00XMJS2000031)。
关键词
姿态估计
深度学习
注意力
pose estimation
deep learning
attention
