一种结合轻量级注意力机制的人体姿态估计算法
A HUMAN POSE ESTIMATION ALGORITHM COMBINED WITH LIGHTWEIGHT ATTENTION MECHANISM
李文星 1喻明毫 2王子牛 1高建瓴 2林宁3
作者信息
- 1. 贵州大学计算机科学与技术学院 贵州贵阳 550025
- 2. 贵州大学大数据与信息工程学院 贵州贵阳 550025
- 3. 中国科学院计算技术研究所计算机体系结构国家重点实验室 北京 100190
- 折叠
摘要
针对现有的人体姿态估计模型存在的模型参数量和计算量大、冗余度高、耗时长等问题,提出一种基于轻量级注意力机制的网络框架.使用轻量级网络MobilenetV3替代了原OpenPose的主干网络VGG-19;对OpenPose的二分支多阶段的卷积神经网络框架进行压缩;引入空间和通道相结合的注意力机制模块CBAM对模型的速度和精度进行权衡.实验结果表明,该方法下的网络模型大小和浮点计算量分别为10.51 MB和22.65 GFlops,相对于原OpenPose减少了 79.91％和83.35％;在COCO2017测试集下,能够在保持较高的检测精度和召回率的基础上显著提升检测速度.
Abstract
Aimed at the problems of the large amount of model parameters and calculation,high redundancy,and long time-consuming in existing human pose estimation models,a network framework of lightweight attention mechanism is proposed.The lightweight network MobilenetV3 was used to replace the original OpenPose backbone network VGG-1.The two-branch multi-stage convolution neural network framework of OpenPose was compressed.The attention mechanism module CBAM that combined space and channel was introduced,and the speed and accuracy of the model were weighed.Experimental results show that the network model size and floating-point calculation amount under this method are 10.51 MB and 22.65 GFlops,respectively,which are reduced by 79.91％and 83.35％compared with the original OpenPose.Under the COCO2017 test set,this algorithm significantly improves the detection speed on the basis of maintaining high detection accuracy and recall rate.
关键词
人体姿态估计/计算机视觉/OpenPose/轻量级网络/注意力机制Key words
Human pose estimation/Computer vision/OpenPose/Lightweight network/Attention mechanism引用本文复制引用
基金项目
国家自然科学基金(62063002)
国家自然科学基金(61532017)
出版年
2024
国家科技期刊平台
NSTL
万方数据