Recognition Based On Kfda's Gait Identity

In this paper the technology of human identification based on gait is mainly concerned.We analyze some gait recognition methods that we have known and probed into the gait features extracting method especially.And a new gait recognition method based on kernel-based Fisher Discrimination Analysis is proposed.For extracting the gait features,we use Least Median of Squares(LMedS) method to carry on background modeling,and the background image is gotten,then carry out difference to subtract the background,and process the subtracted image by counting the sobel edges and mathematics morphology operation,and get two value gait subtraction.Finally,we carry on the aftertreatment, getting rid of the noise and packing the cavity in the outline. Draw the gait outline of people subtraction we received,and calculate the outline centroid.Then we obtain n dots at the outline , calculating the normalized distances of the dots we obtained with the centroid,and the distances constitute the input vector.In training process,we use kernel-based Fisher Discrimination Analysis(KFDA) method to train the input sample vectors.The method has been used in face recognition and has been demonstrated better recognition capability than other methods(PCA,KPCA,SVM).We calculate the optimal subspace Wopt and project the sample gait sequences to Wopt ,then get the tracks of the sequences,calculate the track centroid and calculate the Exemplar Projection Centroid of the sequences in the same class,and the Exemplar Projection Centroid represents the class template.To test the class of a gait sequence,we also project the test sequence to the eigenspace,and calculate the track centroid,then calculate the Euclidean Distance of the test sequence tracking centroid with the sample sequences'Exemplar Projection Centroids.And the class which the test sequence belongs to is the one that the sample sequence which the Euclidean Distance is shortest belongs to.Different kernel functions are tested and their results are compared in this paper. After preliminary experiment,the method that we adopt has reached good recognition result in case of little sample.This demonstrates the feasibility of the method we adopt.