Font Size: a A A

Researches On Gait Recognition Based On Feature Description Of Silhouette And System

Posted on:2013-08-24Degree:MasterType:Thesis
Country:ChinaCandidate:J XiangFull Text:PDF
GTID:2248330362973546Subject:Communication and Information System
Abstract/Summary:PDF Full Text Request
Among various identification algorithms which are based on biologycharacter, gait recognition technology has been extensively applied to many fieldsdue to its unique superiority compared with others. As the key issues in gaitrecognition, feature extraction and expression have received considerableattention in recent years. Generally speaking, potential sources for gait biometricscome from two aspects: features based both on shape and dynamics. However,in-depth research shows that gait dynamics is vulnerable to changing of variablessuch as walking surface and walking velocity. Whether the person carriessomething or not can even exert an influence on the feature’s validity. Therefore,in the field gait recognition, the most effective and stable feature should comefrom shape information. On the other hand, it is noticed that consistency anddifferences among individuals can be better described by the statistical property offeature. Compared with those template matching based algorithms, the algorithmsbased on feature statistics have more robustness and low computationalcomplexity, when considering noise and non-consistency of image information.This dissertation gives systematic and deep investigation to feature extractionand expression based on silhouette information. The main contributions of thisthesis are given below.1As the important basis, silhouette expression based gait recognition hasbeen addressed, including detection, feature extraction, feature space transformand design of classifier. A new feature extraction has been proposed, whichcombines spatial-temporal transformation with moment invariants. By computingmoment invariants of the1D distance signals, the problem of dimensionnormalization for silhouette sequence, which is necessary in spatial-temporalmethod, and the problem of high computation cost in traditional momentinvariants method have been overcome.2Based on time-variant shape information, a new representation scheme forfeature description is proposed which utilizes non-stationary property in thedistribution of feature relationships. Generally speaking, We take the relativedirection of8-Neighborhoods adjacent edge pixels as one of the attributescharacterizing relationship, and label distance from the current edge pixel tocenter point of shape as the other attribute. In this way, the joint probabilitydistribution has been obtained to describe time-variant silhouette. PCA is adopted for feature reduction. Finally, the nearest-neighbor classifier is adopted forclassification. The experiment result demonstrates the efficient of the proposedmethod.3As a feature extraction method having nonlinear data representation,Locality Preserving Projections(LPP) has not utilized the information in the sameclass as well as information between different classes. The paper has presented animproved LPP algorithm by introducing a revised factor and thus discriminateinformation can be used. The improved LPP algorithm has been applied totwo-dimensional feature matrix, and the experiment result demonstrates that thelost of flow information can be avoided.4A new classifier is proposed by introducing the concept of binary detectionto classification. The detector takes the form of randomized forest with the idea toturn gait recognition problems into tasks of training classifier for the samecategory. Feature matrix has been described by2bit Binary Patterns, and therecognition results of detector are obtained by maximum a posteriori criterion.Owing to on-line learning strategy, all possible gait sample information thatsomeone may appears in the video sequence can be properly recorded. Theproposed approach has shown better performance, considering variouscomplicated variations such as occlusion and self-occlusion of object, appearancechanges etc.5We also constructed an identification system based on gait recognition,which has friendly human-computer interface and can be operated simply. Thesystem can not only realize algorithms proposed in this paper, but also provide avaluable platform for further research on gait recognition.
Keywords/Search Tags:gait recognition, feature relationship, feature description, principal component analysis, Locality PreservingProjections, randomized forest, classifier, Identificationsystem
PDF Full Text Request
Related items