| dc.description.abstract | Hand motion provides a natural way of interaction that allows humans to interact not
only with the environment, but also with each other. The effectiveness and accuracy of
hand-tracking is fundamental to the recognition of sign language. Any inconsistencies
in hand-tracking result in a breakdown in sign language communication. Hands are
articulated objects, which complicates the tracking thereof. In sign language communication the tracking of hands is often challenged by the occlusion of the other hand, other body parts and the environment in which they are being tracked. The thesis investigates whether a single framework can be developed to track the hands independently of an individual from a single 2D camera in constrained and unconstrained environments without the need for any special device. The framework consists of a three-phase strategy, namely, detection, tracking and learning phases. The detection phase validates whether the object being tracked is a hand, using extended local binary patterns and random forests. The tracking phase tracks the hands independently by extending a novel data-association technique. The learning phase exploits contextual features, using the scale-invariant features transform (SIFT) algorithm and the fast library for approximate nearest neighbours (FLANN) algorithm to assist tracking and the recovering of hands from any form of tracking failure. The framework was evaluated on South African sign language phrases that use a single hand, both hands without occlusion, and both hands with occlusion. These phrases were performed by 20 individuals in constrained and unconstrained environments. The experiments revealed that integrating all three phases to form a single framework is suitable for tracking hands in both constrained and unconstrained environments, where a high average accuracy of 82,08% and 79,83% was achieved respectively. | en_US |
