一种新的步态特征提取方法的研究

提出一种新的步态特征提取方法，首先利用坎尼算子提取前端腿部的边缘线，并对底部可能产生的误差点进行剔除;接着计算每两帧之间腿部边缘线每个像素点前进的位移，作为原始步态特征;然后，对步态特征进行远近归一化处理，消除被拍摄对象与拍摄镜头之间因距离不同所产生的影响;最后运用主成分分析，将特征空间维度由60维降到3～4维。在识别阶段，用归一化欧式距离计算样本之间的相似程度。提出的这种新的步态特征提取算法在3个人每人4个序列的小样本纯数据库上用最近标本分类器验证所提算法的性能，正确分类率为83.33%;在5个人每人4个序列的小样本混合数据库上，正确分类率为55%。

Research of novel gait feature extraction method

Proposes a new gait feature extraction arithmetic.Firstly，we use “Canny” to get the contour of the front legs and delete the error pixel in the bottom.Then the displacement in every two pixels of the front leg’s silhouette in different frames is computed，which will be the primitive gait signature.Furthermore，the distance unitary of the signature is processed in order to eliminate the influence caused by different distances between objects and camera.At last，we apply the Principle Component Analysis（PCA），which successfully reduce 60 dimensions to 3 or 4 in the eigenvector space.In the recognized stage，applies the normalized euclidean distance to compute the resemblance between samples.Recognizes rate of the arithmetic proposed can reach to 83.33% when using ENN in the small pure database，containing 3 people，for each 4 sequences.The rate can also reach 55% in the mixed database，containing 5 people，for each 4 sequences.