Modeling auditory and visual spatial perception: A unified computational approach

This thesis aims to develop algorithms for artificial perception tasks by studying the perception of three dimensional attributes of auditory and visual signals in human sensory systems. The common underlying representation principles considered lead to a unified model of sensory function for both audition and vision. These principles include: extraction of invariant or quasi-invariant parameters; efficient storage and retrieval; graceful degradation in performance; generalization and learning capabilities; and fuzzy connectionist models of the sensory perceptual processes. The computational models are based on observed biological systems and thus offer plausible models of biological signal processing in the auditory and visual system.; In auditory perception, the localization phenomenon is studied and modeled. Principles of auditory localization using source invariant spectral cues from local time-frequency processing mid-to-high frequency range, based on using directional frequency cues that are imposed by the pinnae, head and torso on the IID at the two ears. The results correspond well with reported acuity of human localization.; In visual perception, we consider the interpretation of 3-D face shape and view invariant recognition of the appearances of faces. A novel approach called shape from recognition and learning is developed to estimate 3-D shape from 2-D images. This approach is based on using meaningful parts of faces, which reduces the dimensionality of the shape-from problem making it tractable. A novel representation for three dimensional objects is developed, that encapsulates both the 3-D structure and its appearance from a continuum of views. This representation is demonstrated for pose estimation and pose-invariant recognition of human faces.; This thesis investigates an area which is important for designing computational algorithms for spatial perception as well as for understanding human perception. It presents a theory for uniform representation and processing of auditory and visual information and identifies the paradigms used to achieve this. It presents several novel algorithms for spatial understanding and perception in auditory localization and 3-D vision.