基于注意力机制改进的多人姿态估计网络研究  

Research on Multi Person Pose Estimation Network Based on Improved Attention Mechanism

作　　者：底家治 王奔 DI Jiazhi;WANG Ben(School of Information Science and Technology,Hangzhou Normal University,Hangzhou 311121,China)

机构地区：[1]杭州师范大学信息科学与技术学院,浙江杭州311121

出　　处：《杭州师范大学学报（自然科学版）》2024年第3期248-254,264,共8页Journal of Hangzhou Normal University(Natural Science Edition)

基　　金：浙江省基础公益研究计划项目(LGF19F020011).

摘　　要：为解决多人姿态估计中小尺度关节点定位准确率低的问题,采用自顶向下的方法,结合人体目标检测模型YOLOv4-tiny,提出一种基于堆叠沙漏网络改进的多人姿态估计网络.该网络包含人体目标检测器和人体姿态估计算法,通过在沙漏网络原始残差模块中融入坐标注意力机制进行特征增强,抑制无用特征的同时增强有用特征,从而提高对人体中小尺度关节点的识别准确率.实验结果表明,该模型在COCO数据集上获得了64.9%的平均准确率,在MPII数据集上正确关键点的比例达88.8%,验证了网络的有效性.To solve the problem of low accuracy in locating small and medium-sized joint points in multi person pose estimation,an improved multi person pose estimation network based on a stacked hourglass network was proposed using a top-down approach combined with the human object detection model YOLOv4-tiny.This network included a modified YOLOv4-tiny(MYT)and a coordinate-stacked hourglass networks(COD-SHN)algorithm.The features were enhanced by incorporating coordinate attention mechanism into the original residual module of the hourglass network,which suppressed useless features while enhancing useful ones.Therefore,the recognition accuracy of small and medium-sized human joints was improved.The experimental results showed that the model achieved an average precision of 64.9%on COCO dataset,and the percentage of correct keypoints on MPII dataset reached 88.8%,indicating the effectiveness of the network.

关 键 词：人体姿态估计 人体目标检测 注意力机制 堆叠沙漏网络 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]