Gait Recognition Research Based On SVM

Recent years,with a growing need for a full range of visual surveillance and monitoring systems in security-sensitive environments such as banks,airports,human identification at a distance has gained increasing interest from computer vision researchers. Gait recognition is only a recognizing technology can be detected and measured at a distance. It overcomes the shortcomings which need physical contact or close contact such as fingerprint recongition, face recognition by its inherent biological characteristics: non-invasive,non-contact,different to hide and to be disguised,far-distance recognition. The researches of gait recognition and its key technologies have important academic significance and practical value. At present,the researches of gait recognitoin are based on small samples. Support Vector Machine(SVM)is a novel powerful machine learning method developed in the framework of Statistical Learning Theory(SLT). It solves practical problems such as small samples,nonlinearity,over learning and high dimension,which exist in most of learning methods, and has high generalization. So SVM is suitable for gait recognition. Focusing on this topic,this dissertation mainly includes the following issues:(1) According to the different feature between body silhouette and gait variance, a new gait recognition approach based on region variance feature and SVM called RFSBGR(Region Feature and SVM Based Gait Recognition) was presented. First,the ratio of width and height of body was obtained. Then gait silhouettes were divided into several regions for each sequence and feature vectors were acquired. Finally,gait classification and recognition were performed by SVM. The experiment result shows that not only can the approach overcome the information lost which result from a few feature vectors but also get exciting recognition performance.(2) According to outer silhouette variance with walking ,a new gait recognition method based on kernel principal component analysis and SVM called KSBGR was explored. First, The outer silhouette was projected along its four directions up,down,left and right, every projection vector was represented by one 1-D signal, gait information was composed by four 1-D signals. Then KPCA, a nonlinear machine learning method, was performed to extract gait features from silhouettes for individual recognition. Finally, gait classification and recognition were carried out by SVM.The methods above were applied to two data-sets(UCSD and CASIA).Experimental results demonstrate that the approaches get exciting recognition performance with relatively lower computational cost. they are effective methods of gait feature extraction and recognition.