多非线性多视角局部保持投影的步态识别

Multi-nonlinear multi-view locality preserving projections for gait recognition

中文关键词  步态识别  多视角  非线性函数  局部保持投影

英文关键词  gait recognition  multiple views  nonlinear functions  locality preserving projections

中文摘要

      已知样本与待识别样本的视角差异是影响步态识别精度的主要因素, 子空间方法将不同视角的步态投影到公共子空间, 能有效避免视角差异的影响. 但现有方法多通过学习投影矩阵对样本进行线性投影, 难以保持多视角步态数据的原始非线性结构. 针对于此, 本文提出多非线性多视角局部保持投影. 先用非线性函数族实现样本的多次非线性投影, 再基于局部结构保持原则将不同视角的样本投影到公共子空间, 最后在公共子空间中进行最近邻分类识别. 在多视角步态库CASIA(B)进行步态识别实验, 结果表明本文方法在多种视角组合下优于其它投影方法.

英文摘要

      View variation between register samples and unknown samples is the main factor affecting gait recognition accuracy. The subspace methods can effectively avoid the influence by projecting samples into a common subspace. However, most of them project samples linearly via projection matrices, which cannot maintain the original nonlinear structure of multi-view gait data. To tackle this problem, multi-nonlinear multi-view locality preserving projections is proposed in this paper. The nonlinear projection is realized by the nonlinear functions family, and then samples are projected into common subspace based on the principle of locality preserving projections. Finally, the unknown samples are identified by nearest neighbor classification in the common subspace. Experiments on the CASIA(B) gait database show that the proposed method is superior to other projection methods under multiple combinations of views.