A general subspace ensemble learning framework via totally-corrective boosting and tensor-based and local patch-based extensions for gait recognition

Subspace learning methods have played an important role on handling the high-dimensional gait template for human identification. Particularly, linear discriminant analysis (LDA) method has been widely applied to find one discriminant low-dimensional subspace for gait recognition. However, when gait templates changed by clothing, view angle, load carrying and road surface variants, only learning one single subspace is prone to dropping into local optimum. In this paper, a subspace ensemble learning using totally-corrective boosting (SEL_TCB) framework and its tensor-based and local patch-based extensions are proposed for gait recognition. In this framework, multiple discriminant subspaces are iteratively learned using totally-corrective boosting technology to preserve the proximity relationships described by instance triplets. Meanwhile, by constructing different triplet set, the presented framework can deal with complex application environments. Further, we extend SEL_TCB framework to tensor SEL_TCB (TSEL_TCB) framework which effectively preserves the structural information of the gait template. Meanwhile, compared with the holistic appearance-based SEL_TCB framework, a local patch-based SEL_TCB (LPSEL_TCB) framework is proposed, which iteratively learns multiple discriminant subspaces corresponding to several local patches selected from the gait template. The proposed method is compared with the recently published gait recognition approaches on USF HumanID database and CASIA Gait database. Experimental results indicate that the proposed method achieves highly competitive performance against state-of-the-art gait recognition approaches.