多分类器信息融合的步态识别算法

Gait Recognition Based on the Fusion of Multiple Classifiers

融合运动人体整体轮廓和局部关节的特征信息,提出了一种新的步态识别算法。对每个序列进行运动轮廓抽取,从3个方向(水平、垂直、斜向)对时变的2维轮廓进行投影扫描,转换为对应的特征向量;对级联的特征向量分别采用离散正交小波变换(DWT)和核主元分析法(KPCA)提取轮廓时空变化所蕴涵的非线性步态信息,构成两个独立的全局特征分类器。对运动人体髋关节和膝关节建模,根据步态运动的准周期性,将关节角度时序信息按傅里叶级数形式展开,采用遗传算法搜索各次谐波的系数并进行尺度变换,生成局部关节时变特征向量,构成局部特征分类器。最后采用贝叶斯多分类器融合决策规则,融合整体和局部特征。在CMU步态数据库中进行实验,结果验证了算法的有效性,识别性能和验证性能都获得有效的提高。

A new approach to gait recognition based on fusion of the information of global silhouette and local joint angle is proposed. The vector data scanned from horizon, vertical and diagonal of the outer contour of binarized silhouette of a walking person are chosen as the basic image feature. Two independent global classifiers are established respectively by the decomposed feature based on the discrete wavelet transformation（DWT） and the nonlinear components of basic gait features extracted based on kernel principal component analysis （ KPCA ）. The coax and knee joint of moving body are simply modeled. The acquired joint angle information is expanded in Fourier series form in view of the periodic character of gait activity. The genetic algorithm is applied to search for the expanding coefficients, and the local feature classifier is established by the normalized eigenvector about joint angle. At last, the global and local features are fused based on different Bayesian combination rules on decision level to improve the performance of both identification and verification. This algorithm is applied to CMU database. Extensive experimental results demonstrate that the proposed algorithm performs nicer classification and verification capability.