基于序列多尺度特征融合表示的层级舞蹈动作姿态估计方法  

Hierarchical Dance Pose Estimation Algorithm Based on Sequential Multi-Scale Feature Fusion

作　　者：杨红红 王刘丽 张玉梅[1,2] 吴晓军[1,2] 党允彤[3] YANG Hong-hong;WANG Liu-li;ZHANG Yu-mei;WU Xiao-jun;DANG Yun-tong(Key Laboratory of Modern Teaching Technology,Ministry of Education,Shaanxi Normal University,Xi’an,Shaanxi 710062,China;School of Computer Science,Shaanxi Normal University,Xi’an,Shaanxi 710062,China;School of Journalism and Communication,Shaanxi Normal University,Xi’an,Shaanxi 710062,China)

机构地区：[1]陕西师范大学现代教学技术教育部重点实验室,陕西西安710062 [2]陕西师范大学计算机科学学院,陕西西安710062 [3]陕西师范大学音乐学院,陕西西安710062

出　　处：《电子学报》2021年第12期2428-2436,共9页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61907028,No.11872036,);陕西省青年科技新星项目(No.2021KJXX-91);陕西省高校科学人才托举项目(No.20200105);中央高校(No.GK202103114,No.GK202101004);陕西省自然科学基金(No.2019JQ-574)。

摘　　要：人体姿态估计是计算机视觉研究领域的热点研究问题之一,但其在传统民间舞蹈动作姿态估计方面的应用研究尚处于起步阶段.由于舞蹈图像中人体动作复杂多变、舞蹈动作连贯性强、舞蹈者存在严重遮挡不易检测等特点,传统人体姿态估计方法难以准确估计舞蹈者的动作变化,导致舞蹈动作姿态估计准确率较低.针对此问题,本文提出一种基于序列多尺度特征融合表示的层级舞蹈动作姿态估计方法,该方法针对舞蹈动作骨骼关节点尺度变化剧烈的问题,构建基于序列多尺度特征融合表示的关节点估计模型.并且,针对舞蹈姿态形变较大,遮挡严重的问题,设计基于关节点几何关系的层级姿态估计模型,提高舞蹈动作姿态估计的效果.实验结果表明,本文方法在标准人体姿态估计数据集及自建舞蹈数据集上取得较好的姿态估计结果.Human pose estimation is one of the hot research topics in the field of computer vision,but its application in traditional dance pose estimation is still in its infancy.Due to the complexity of dance pose,the strong coherence of dance movements,and difficulty in detecting of dancers'poses caused by serious occlusion in dance images,the traditional human pose estimation methods are difficult to accurately estimate the pose changes of dancers,thus resulting in low accura⁃cy in estimating dance pose.We propose a hierarchical dance pose estimation method based on sequential multi-scale fea⁃ture fusion.To address the problems of the drastic scale changes of the dancer pose,a keypoint estimation model based on sequential multi-scale feature fusion is constructed.Furthermore,aiming to solve the issues that the large deformation and serious occlusion of dance pose,a hierarchical pose estimation model based on the geometric relationship between human keypoints is designed to improve the accuracy of dance pose estimation.The experimental results show that the proposed method can achieve good pose estimation results on the standard human pose estimation dataset and the self-collected dance dataset.

关 键 词：舞蹈动作姿态估计 序列多尺度特征融合 关节点几何关系 层级姿态估计 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]