Analysis and evaluation of optimization techniques for tracking in augmented reality applications

• Autor(a) principal: Teixeira, João Marcelo Xavier Natário
• Data de Publicação: 2013
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/12268
• topic: Realidade aumentada; Visão computacional; Alto desempenho; Augmented reality; Computer vision; High performance

Real-time computer vision applications that run on the field and make frequent use of wearable computers have a critical restriction on the amount of processing they can perform, because of the fact that most (if not all) of the application runs on the wearable platform. A balancing scheme capable of allowing the application to use more processing power is fundamental both when input scenarios present more visual restrictions regarding, for example, the object to be tracked, and also to reduce processing in order to save battery and CPU time for other applications when the captured video is better controlled (more accessible). The fact that computer vision applications may run on a variety of platforms justifies the need for defining a model that automatically adjusts the tracker being used in applications with hard performance constraints. Performance degradation in wearable platforms can be greater than expected, as desktop and mobile platforms present different levels of hardware capabilities, and consequently, different performance restrictions. This doctoral thesis addresses the object tracking problem using a decision model, in such a way that prioritizes using the least computationally intensive algorithm whenever possible. It has the following specific objectives: to investigate and implement different tracking techniques, to choose/define a reference metric that can be used to detect image interference (occlusion, image noise, etc.), to propose a decision model that allows automatic switching of different trackers in order to balance the application's performance, and to reduce the application's workload without compromising tracking quality. The effectiveness of the system will be verified by synthetic case studies that comprise different object classes that can be tracked, focusing on augmented reality applications that can run on wearable platforms. Different tracking algorithms will be part of the proposed decision model. It will be shown that by switching among these algorithms, it is possible to reach a performance improvement of a factor of three, while keeping a minimum quality defined by a reprojection error of 10 pixels when compared to the use of only the best algorithm independent of its computational cost. This work results in better performance of applications with memory and battery restrictions.