基于迭代切距离原型学习算法的步态识别

An Iterative Gait Prototype Learning Algorithm Based on Tangent Distance

作为唯一远程生物认证技术,步态识别一方面越来越受到人们的重视,提出了很多相应的算法,另一方面,它又面临着很多挑战,其难点之一是如何从多帧步态中有效地提取步态特征.针对此问题,并基于步态能量图(GEI)在步态特征表示上的效果,提出了一种迭代切距离原型学习算法.假定各人的步态分布在不同流形上面,首先用切距离改进步态能量图的定义,进而用迭代的方法来解一个最优解问题,从而学习出步态原型图,再通过PCA对步态原形进行特征提取,最后进行识别.证明了该方法的收敛性,实验结果表明所提出的方法取得了比GEI更好的识别率,并证明了步态流形的假设的合理性.

Being the only biometry certification techniques for remote surveillance,gait recognition,on one hand,is regarded as being of important potential value,hence a lot of algorithms have been proposed,on the other hand,it has encountered a lot of challenges.Among all of the challenges gait recognition encountered,one of them is how to extract features efficiently from a sequence of gait frames.To solve this problem,and also based on the fact that gait energy image（GEI） is effective for feature representation,an iterative prototype algorithm based on tangent distance is proposed.Firstly,it is assumed that different gaits lie in different manifolds.As a result,the proposed algorithm refines the definition of gait energy image（GEI） using tangent distance.Then an iterative algorithm is proposed to learn the prototypes by solving an optimization problem.Finally,principal component analysis（PCA） is performed on the prototypes to obtain gait features for classification.The proposed method is proved converged,and experiment results show the promising results of the proposed algorithm in accuracy compared with the GEI＇s.The rationality of the assumption that gaits lie in specific manifolds is also validated through experiments.