View-based models for visual tracking and recognition

Abstract

The thesis aims to develop efficient view-based models for tracking and classifying target objects in images. It first proposes a kernel-based method for tracking objects under affine transformation. The basis of the method is an affine matching technique, which, by precisely characterizing each object's spatial and spectral features, is capable of distinguishing similar objects in cluttered scenes while providing important posture information for motion understanding and recognition. The affine matching technique leads to an efficient and analytical mathematic solution for tracking. Extensive experiments have been conducted, yielding positive results. The thesis also presents a neural network model called kernel autoassociators that takes advantage of kernel feature space to learn nonlinearity in samples. In addition, the thesis extends kernel autoassociators to a Gabor wavelet associative memory model, which inherits advantages of Gabor wavelet networks for facial image representation. A high-performance face recognition system is built based on the model.