基于Gabor小波和局部二值模式的步态识别

利用步态对个人身份进行识别已经受到越来越多生物识别技术研究者的重视。步态能量图(GEI-Gait Energy Image)是一种有效的步态表征方法,Gabor小波能提取不同方向、不同尺度空间频率特征,因此,首先利用Gabor小波提取步态能量图不同方向、不同尺度的信息,得到其幅值谱图,再利用LBP来提取Gabor幅值谱图的局部信息,相对于LBP直接作用于步态能量图,能提取步态能量图更多方向、更多尺度的局部特征。最后,利用具有良好降维和辨识能力的辨识共同向量(DCV-Discriminant Common Vector)对提取的LBP特征进行维数约减和特征选择,只需利用简单的最近邻分类器就能取得较好的识别效果。该算法在中科院自动化所的CASIA数据库上面进行试验取得了较高的正确识别率。还针对步态识别中的小样本问题提出了一种样本扩充方法,解决了步态识别中的小样本问题,并提高了算法的识别率。

Gait Recognition Based on Gabor Wavelet and Local Binary Pattern

Recently, gait recognition for individual identification has been attracting increasing attention from biometrics researchers. It is well known that Gait Energy Image (GEI) is an efficient representation for gait, and Local Binary Pattern (LBP) can extract the local information efficiently, but the information lack of the orientation and scale characteristic, Gabor wavelet can extract the feature of different orientation and scales. First, using Gabor wavelet to extract the different orientation and scales-information of GEI, the magnitude spectral image is obtained. Second, LBP is used to extract the local information from magnitude spectral image, it can extract more local orientation and scale feature than the method of directly use LBP on GEI. At last, as the dimension of the LBP feature is usually very high, this paper employs a popular method called Discriminative Common Vectors (DCV) for further dimensionality reduction, which minimizes the within-class distance and maximizes the between-class distance as much as possible. Finally, for simplicity consideration, the nearest neighbor classifier to classification is used. Experimental results on CASIA databases show that our algorithm is effective and obtains high recognition rates. Further, a sample expand method is proposed for the small sample problem in gait recognition, the method increase the recognition rates.