摘要
为了提升机器视觉中特定舞蹈动作识别的性能,设计了基于全局上下文的特定舞蹈动作识别方法。该方法基于Hourglass结构,通过连接高低分辨率的特征图,将具备全局信息的深层特征图上采样与浅层特征融合,使得每一个阶段的高分辨率特征图均具有低分辨率的特征图表示,从而得到信息更丰富的高分辨率特征图表示,最终回归人体姿态热力图。在Balletto舞蹈视频数据库中的测试结果表明,相比基于CPN和基于Hourglass的算法,所提算法的AP值提高2.4%,AR提升了1.6%。
To improve the performance of specific dance action recognition in machine vision,a specific dance action recognition method based on global context is designed.This method is based on Hourglass structure.By connecting the high-resolution and low-resolution feature maps,the deep feature map with global information is sampled and fused with the shallow feature map,so that the high-resolution feature map of each stage has the low-resolution feature map representation,so as to obtain the high-resolution feature map representation with more information,and finally return to the human posture thermal map.The test results in Balletto dataset show that compared with the algorithms based on CPN or Hourglass,the AP score and AR score of the proposed algorithm are increased by 2.4%and 1.6%,respectively.
作者
毕雪超
BI Xuechao(Youth League Committee,Xi’an Vocational and Technical College of Aeronautics and Astronautics,Xi’an 710089,China)
出处
《微型电脑应用》
2020年第10期21-23,共3页
Microcomputer Applications
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)。
