Coupled Metric Learning And Its Applications On Gait Recognition

With the increase of computer storage capability and computing ability, biometric identification and recognition based on computer vision techniques develop rapidly, and some biometric recognition techniques under controlled scenarios such as face recognition, fingerprint recognition and iris recognition have been relatively mature and applied to many fields, e.g. business, public security, medical science, and even entertainment. However, the query biological features are usually required to be captured under non-contact and long-distance condition in real scenes by different types of sensors with low quality, which poses significant challenge to traditional biometric identification systems.Gait feature has many advantages over other biometrics, such as remote detection, non-contact collection, difficulty of camouflage and imitate, little environment affection, and little memory occupation, etc. Thus, it is recognized as one of the most potential biological features and possesses extensive application prospect and economic values. But view change can cause great difficulties for gait identification because it will alter available features for recognizing substantially. To solve the cross-view gait recognition problem caused by view change, the project proposes cross-view gait recognition methods based on coupled metric learning. In all, the main contribution of this dissertation is listed as follows:● Research background and significance are studied and research status on gait recognition across views as well as coupled metric learning is reviewed.● Coupled metric learning (CML) algorithms based on manifold alignment on vector space are discussed, including Coupled Distance Metric Learning (CDML), Coupled Marginal Fisher Analysis (CMFA) and Simultaneous Discriminant Analysis (SDA). Specially, we propose coupled marginal discriminant analysis (CMDA) which aims to reduce data gap of cross-domain problems and enhance coupled performance between heterogeneous data. In addition, we apply these CML methods on vector space to cross-view gait recognition task for the first time.● We propose a coupled metric learning skeleton based on manifold alignment on matrix space and develop several two-dimensional (2D) algorithms, including 2D-CDML,2D-CMFA,2DCoupled Linear Discriminant Analysis (2D-CLDA),2D-SDA and 2D-CMDA. Moreover, we validate the effectiveness of proposed algorithms under the skeleton by 2D gait energy images (GEI) of CASIA (B) dataset by Institution of Automation of Chinese Academy of Science.● We propose a coupled metric learning skeleton based on manifold alignment on tensor space (TS) which extends coupled metric learning methods on vector and matrix space to expression on high-dimensional space, aiming to solve high-dimensional heterogeneous recognition problems. Based on the skeleton, this dissertation proposes TCDML, TCMFA, TCMDA, TCLDA and TSDA algorithms on tensor space. Furthermore, this thesis studies gait tensor features based on Gabor representation and verifies the priority of these algorithms on high-dimensional heterogeneous data.