改进的基于锚点的三维手部姿态估计网络  

Improved 3D hand pose estimation network based on anchor

作　　者：危德健 王文明[1] 王全玉[1] 任好盼 高彦彦 王志 WEI Dejian;WANG Wenming;WANG Quanyu;REN Haopan;GAO Yanyan;WANG Zhi(School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学计算机学院,北京100081

出　　处：《计算机应用》2022年第3期953-959,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(71834001)。

摘　　要：近年来基于锚点的三维手部姿态估计方法比较流行,A2J(Anchor-to-Joint)是比较有代表性的方法之一。A2J在深度图上密集地设置锚点,利用神经网络预测锚点到关键点的偏差以及每个锚点的权重。A2J使用预测的偏差和权重,以加权求和的方式计算关键点的坐标,降低了网络回归结果中的噪声。虽然A2J简单高效,但是不恰当的网络结构和损失函数影响了网络的准确度,因此提出改进的网络HigherA2J。首先,使用一个分支预测锚点到关键点的XYZ偏差,更好地利用深度图的3D特性;其次,简化A2J的网络分支结构从而降低网络参数量;最后,设计关键点估计损失函数,结合关键点估计损失和偏差估计损失,有效提高估计准确度。在三个数据集NYU、ICVL和HANDS 2017上的实验结果显示,手部姿态估计的平均误差比A2J都有所降低,分别降低了0.32 mm,0.35 mm和0.10 mm。In recent years,anchor-based 3D hand pose estimation methods are becoming popular,and Anchor-to-Joint(A2J)is one of the more representative methods.In A2J,anchor points are densely set on depth map,and neural network is used to predict offsets between anchor points and key points together with weights of anchor points;predicted offsets and weights are used to calculate the coordinates of key points in a weighted summation mode to reduce noise in network regression results.A2J methods are simple and effective,but they are sensitive to ill-suited network structure and prone to inaccurate regression due to loss function.Therefore,an improved network HigherA2J was proposed.Firstly,a single branch jointly predicted X,Y and Z offsets between anchors and key points to better utilize 3D characteristics of depth map;secondly,network branch structure was simplified to reduce network parameters;finally,the loss function for key point estimation was designed,combined with offset estimation loss,which improved the overall estimation accuracy effectively.Experimental results show the reductions in average hand pose estimation error of 0.32 mm,0.35 mm and 0.10 mm compared to conventional A2J on three datasets NYU,ICVL and HANDS 2017 respectively.

关 键 词：三维手部姿态估计 深度学习 卷积神经网络 锚点 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]