基于2D姿势估计的高动态舞蹈动作识别方法  

High dynamic dance motion recognition method based on 2D pose estimation

作　　者：毕雪超[1] BI Xue-chao(Xi’an Vocational and Technical College of Aeronautics and Astronautics,Xi’an 710089,China)

机构地区：[1]西安航空职业技术学院,西安710089

出　　处：《信息技术》2020年第4期60-64,共5页Information Technology

基　　金：陕西省高等教育工作委员会研究课题(2017FKT03);西航职院2018年度科研计划项目(18XHGZ-011)。

摘　　要：针对舞蹈视频图像中动作复杂多变、连贯性强、遮挡问题严重等问题,文中结合先进的舞蹈动作识别技术发展方向及其应用场景,同时考虑到彩色图像处理中计算机处理负载过重的问题,设计了一种基于2D姿势估计的高动态舞蹈动作识别方法。该方法主要分为模板建立与姿势估计两个部分,主要涉及的处理操作有图像预处理、模板特征提取和模板匹配这三种。验证测试结果表明,训练集图像经过灰度化与阈值化后,即可获得图像中前景舞者的图形,再利用Kinect人体模型提取动作特征信息。由于考虑到拍摄角度等原因导致的特征差异,将描述同一动作的多张训练图的特征信息保存在同一信息矩阵中,可进一步提高动作识别的准确性。Aiming at the problems of complex and changeable movements,strong coherence and serious occlusion in dance video images,a high dynamic dance motion recognition method based on 2D posture estimation is designed,which combines the development direction of advanced dance motion recognition technology and its application scenarios,and considers the problem of computer processing overload in color image processing.This method is mainly divided into two parts:template building and pose estimation.The main processing operations involved are image preprocessing,template feature extraction and template matching.The validation test results show that the foreground dancer’s figure can be obtained from the training set image after graying and thresholding,and then the Kinect human body model is used to extract the action feature information.Considering the difference of feature caused by shooting angle,the accuracy of action recognition can be further improved by storing the feature information of multiple training maps describing the same action in the same information matrix.

关 键 词：舞蹈动作识别 KINECT 灰度化 阈值化 姿势估计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]