Human ear recognition in three dimensions

Biometrics deal with recognition of individuals based on their physiological or behavioral characteristics. Ear, a viable new class of biometrics, has certain advantages over other biometrics (e.g. face and fingerprint). An ear can be imaged in 3D using a range sensor and it contains shape information, which makes it possible to develop a robust 3D ear biometrics. In this dissertation, we propose a human recognition system using 3D ear biometrics. These algorithms include: (1) Precise localization of 3D ears: We propose a two-step approach that uses a single reference 3D shape model for precise localization of ear in the 3D range image and the registered color images. In the first step color and range images are fused to locate regions-of-interest (ROIs). In the second step, the reference 3D shape model is adapted to the image by following a global-to-local registration procedure. The local deformation drives the initial global registration towards the target with the topology of the shape model preserved. (2) 3D ear recognition: We propose two representations for ear recognition. These include the ear helix/anti-helix representation obtained from the detection algorithm and the local surface patch (LSP) representation. A local surface descriptor is characterized by a centroid, a local surface type and a 2D histogram. Both shape representations are used to estimate the initial rigid transformation and then an Iterative Closest Point (ICP) algorithm is run for the verification. (3) Rapid 3D ear recognition: We propose a novel method for rapid 3D ear recognition which combines the feature embedding and the machine learning techniques. The FastMap algorithm embeds LSPs into a low-dimensional space with distance relationships preserved. By searching the nearest neighbors in low dimensions, the similarity between a model-test pair is computed. The similarities for all model-test pairs are ranked using the Support Vector Machines rank learning algorithm to generate a short list of candidate models for verification.; The experimental results on the UCR dataset of 155 subjects with 902 images under pose variations and the University of Notre Dame dataset of 302 subjects with time-lapse gallery-probe pairs are presented to demonstrate the effectiveness of the proposed algorithms and the system.