基于空间交叉卷积的轻量级人体姿态估计算法  

Lightweight Human Pose Estimation algorithm Based on Spatial Cross Convolution

作　　者：方益 石守东[1] 方靖森 叶永芳 蓝艇[1] FANG Yi;SHI Shoudong;FANG Jingsen;YE Yongfang;LAN Ting(Faculty of Electrical Engineering and Computer Science,Ninbo University,Ningbo Zhejiang 315211,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《传感技术学报》2024年第3期439-445,共7页Chinese Journal of Sensors and Actuators

基　　金：浙江省公益技术应用研究项目(LGF22F020029);中国创新挑战赛(宁波)项目(2022T001)。

摘　　要：针对改进轻量级OpenPose网络在预测阶段仍有较大参数量会降低模型推理速度,不利于在边缘设备部署的问题,提出一种基于改进卷积方法的人体姿态估计网络,使用空间交叉卷积来代替部分标准卷积,减少网络预测阶段的参数量。网络的输入为单目摄像头捕获的RGB图像,以MobileNetV3-Large为主干网络,并在其中加入了CBAM注意力模块,提取不同重要程度的空间和通道特征。获取图像特征后,送入两个分支中分别预测关键点位置和关键点组合关系。以空间交叉卷积代替两个分支中的部分标准卷积核,相对标准卷积能够减少80%的参数量。实验结果表明,相较于原方法,所提方法在精度下降较小的情况下,总参数量降低了22%,部署在CPU端的测试结果显示,速度能够达到6 FPS,提升了4倍。To address the problem that the number of parameters in the prediction phase of the lightweight OpenPose network are still large,and this can slow down model inference and is not conducive to deployment in edge devices,a human pose estimation network based on an improved convolution approach is proposed,using spatial cross-convolution to replace some of the standard convolutions and reduce the number of parameters in the prediction phase of the network.The input of the network is RGB images captured by a monocular camera.MobileNetV3-Large is used as the backbone network,and the CBAM attention module is added to the network to extract spatial and channel features of different importance.After obtaining the image features,the images are fed into two branches to predict the position and combination relationship of key points.Spatial cross-convolution is used to replace some standard convolution kernels in the two branches,which can reduce the number of parameters by 80% compared with traditional convolution.The experimental results show that,compared with the original method,the total number of parameters of the proposed method is reduced by 22% with only a small decrease in accuracy.The test results of the deployment on the CPU side show that the speed can reach 6 FPS,which is nearly 4 times higher.

关 键 词：人体姿态估计 轻量级网络 空间交叉卷积 OpenPose 边缘设备 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]