基于多核稀疏编码的三维人体姿态估计  

3D Human Pose Estimation Based on Multi-kernel Sparse Coding

作　　者：余家林[1] 孙季丰[1] 李万益[1] 

机构地区：[1]华南理工大学电子与信息学院,广东广州510641

出　　处：《电子学报》2016年第8期1899-1908,共10页Acta Electronica Sinica

基　　金：国家自然科学基金青年科学基金(No.61202292);广东省自然科学基金(No.9151064101000037)

摘　　要：为了准确有效的重构多视角图像中的三维人体姿态,该文提出一种基于多核稀疏编码的人体姿态估计算法.首先,针对连续帧姿态估计的歧义问题,该文设计了一种用于表达多视角图像的HA-SIFT描述子,其中,人体局部拓扑、肢体相对位置及外观信息被同时编码;然后,在多核学习框架下建立同时考虑特征空间内在流形结构与姿态空间几何信息的目标函数,并在希尔伯特空间优化目标函数以更新稀疏编码、过完备字典与多核权值;最后,利用姿态字典原子的线性组合来估计对应未知输入的三维人体姿态.实验结果表明,与核稀疏编码、Laplace稀疏编码及Bayesian稀疏编码相比,文本方法具有更高的估计精度.In order to reconstruct 3D human pose from multi-view images accurately and effectively,a novel human pose estimation algorithm based on multi-kernel sparse coding is proposed.First,for the ambiguity of human pose estimation between the consecutive frames,we describe multi-view images using a special HA-SIFT descriptor,in which the human body local topology,relative coordinates and appearance information are encoded simultaneously;then,an objective function is established within the framework of multi-kernel learning,it takes both intrinsic manifold structure of the feature space and geometrical information of the pose space into consideration.The sparse coding,over-complete dictionary and multi-kernel weight are updated by optimizing the objective function iteratively in the Hilbert space;finally,the corresponding 3 D human pose of the unknown input image is estimated by a linear combination of the bases of the human pose dictionary.The experi-mental results show that our proposed method provides higher accuracy of human pose estimation compared with kernel sparse coding,Laplace sparse coding and Bayesian sparse coding.

关 键 词：人体姿态估计 多视角图像 多核学习 稀疏编码 字典学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]