| Summary: | In the field of computer vision, developing automated systems to recognize people under unconstrained scenarios is a partially solved problem. In unconstrained sce- narios a number of common variations and complexities such as occlusion, illumi- nation, cluttered background and so on impose vast uncertainty to the recognition process. Among the various biometrics that have been emerging recently, this dissertation focus on two of them namely face and gait recognition. Firstly we address the problem of recognizing faces with major occlusions amidst other variations such as pose, scale, expression and illumination using a novel PRObabilistic Component based Interpretation Model (PROCIM) inspired by key psychophysical principles that are closely related to reasoning under uncertainty. The model basically employs Bayesian Networks to establish, learn, interpret and exploit intrinsic similarity mappings from the face domain. Then, by incorporating e cient inference strategies, robust decisions are made for successfully recognizing faces under uncertainty. PROCIM reports improved recognition rates over recent approaches. Secondly we address the newly upcoming gait recognition problem and show that PROCIM can be easily adapted to the gait domain as well. We scienti cally de ne and formulate sub-gaits and propose a novel modular training scheme to e ciently learn subtle sub-gait characteristics from the gait domain. Our results show that the proposed model is robust to several uncertainties and yields sig- ni cant recognition performance. Apart from PROCIM, nally we show how a simple component based gait reasoning can be coherently modeled using the re- cently prominent Markov Logic Networks (MLNs) by intuitively fusing imaging, logic and graphs. We have discovered that face and gait domains exhibit interesting similarity map- pings between object entities and their components. We have proposed intuitive probabilistic methods to model these mappings to perform recognition under vari- ous uncertainty elements. Extensive experimental validations justi es the robust- ness of the proposed methods over the state-of-the-art techniques. i
|
|---|
