Research Of Gait Recognition Algorithm Based On Tensor Discriminant Analysis

Biometrics is the technology of establishing, verifying and recognizing the identityof an individual based on the physical or behavioral attributes of the person. In recentyears, there has been increasing attention on human biometrics. Biometric technologyhas achieved rapid development in many fields, such as public security, informationsecurity, financial security. As a new technology of biometrics, gait recognition’s uniqueadvantages are its unobtrusiveness, offering potential for recognition at a distance or atlow resolution when the human subject occupies too few image pixels, so gaitrecognition has attracted a great deal of interest in computer vision and patternrecognition community, and has become one of the popular research directions.However, many factors negatively influence gait image sequences. These factorsinclude view, shooting environment, appearance and occlusions. Most of such factorsimpose challenges to gait recognition so that current gait recognition methods are farfrom practical application requirements. Therefore, building a robust featurerepresentation and efficient discriminant algorithm become an important solution toimprove gait recognition performance. The content of this dissertation are as follows.(1) Technology of gait silhouette detection and extraction was studied. Firstly, weintroduced human detection method based on background subtraction. Secondly,incremental learning for robust visual tracking algorithm was applied to human trackingof gait recognition, which combines incremental learning online for principalcomponent analysis algorithm and particle filter. This approach can quickly andefficiently extract human regions of gait image sequences. Thirdly, Gaussian mixturemodel and expectation maximization algorithm was studied to extract accurate gaitsilhouettes.(2) For periodic, long distance and low distortion characteristics of gait imagesequences, two gait cycle detection methods was firstly studied: the first one is based onthe binary image of gait silhouette, and the second one is based on gait outer contour.Then three effective gait feature representations were studied and proposed: Gait EnergyImage which reflects major shapes of silhouettes and their changes over the gait cycle wasdescribed in detail; Mutichannel GEI which preserves more temporal information wasproposed; Three different features using Gabor functions based on Mutichannel GEIwere developed, and these features describe as much as local and global information ofgait silhouette images based on the multi-scale and multi-orientation characteristics ofGabor functions.(3) High dimensional features and undersample problem of gait recognition weresolved. Firstly, the classical Linear Discriminant Analysis algorithm was briefiyintroduced; Secondly, Graph Embedding, its general framework and MarginalDiscriminant Analysis were studied in depth. Graph Embedding is able to effectivelysolve the problem of high dimensional features, and MDA can effectively overcome thelimitations of the traditional LDA algorithm due to data distribution assumptions andavailable projection directions. Finally, Marginal Tensor Discriminant Analysis algorithm was proposed, and we tested its performance. Experimental results show thaton the one hand, the performance of MTDA algorithm is superior when the the numberof the training samples is limited, that is, it significantly reduces the effects ofundersampling on classification. On the other hand, the alternating projectionoptimization procedure of MTDA converges rapidly. The new MTDA algorithm basedon the developed Gabor features achieve better recognition rate than those of previousalgorithms.