核希尔伯特再生空间
嵌入
数学
核(代数)
图像扭曲
代表(政治)
人工智能
希尔伯特空间
算法
计算机科学
插值(计算机图形学)
参数化复杂度
再中心定理
核主成分分析
规范化(社会学)
变换几何
动态时间归整
特征学习
不变(物理)
核方法
刚性变换
形状分析(程序分析)
动作(物理)
热核特征
补语(音乐)
线性空间
几何变换
活动形状模型
模式识别(心理学)
欧几里得空间
线性地图
向量空间
理论计算机科学
分布的核嵌入
空格(标点符号)
拓扑骨架
正定矩阵
作者
Ximo Gual-Arnau,J. Monterde
摘要
Abstract. We introduce a mathematical framework for the representation and classification of 3D human actions from sequences of skeletal joint configurations. Each skeleton is modeled as a shape in Kendall’s shape space of 3D landmarks and is embedded into the space of symmetric positive semidefinite matrices with trace equal to one. This extrinsic formulation endows the representation with a linear structure that enables an embedding into a reproducing kernel Hilbert space (RKHS), thereby facilitating the direct use of kernel-based machine learning methods for action classification. Within this framework, sequences of skeletons are represented as smooth parameterized curves in the embedding space, allowing for interpolation and normalization of the number of frames while preserving the geometric structure of the shapes. The linear geometry induced by the embedding simplifies the application of modern learning tools, avoiding the computational and conceptual difficulties associated with intrinsic Riemannian approaches on Kendall’s shape space. To compare action trajectories while accounting for temporal variability, we employ Dynamic Time Warping (DTW) and the global alignment kernel (GAK), which provides a valid positive-definite kernel suitable for classifiers such as support vector machines. Experiments on benchmark skeleton-based human action datasets demonstrate that the proposed approach achieves competitive performance while offering a principled and geometrically grounded representation of 3D action dynamics.
科研通智能强力驱动
Strongly Powered by AbleSci AI