Abstract
This dissertation explores various problems for recovering the trajectory of a mobile platform, and reconstructing the 3-D models of objects of interest from stereo data. It is primarily targeted to develop a robust, efficient and unified solution for autonomous navigation and 3-D reconstruction for applications in uncontrolled environments, including the mapping of benthic objects in underwater applications. To achieve these goals, the topics of research become considerably broader in scope, encompassing feature tracking, stereo matching, stereo fusion, recursive motion and structure estimation, and 3-D reconstruction. The unknown, unstructured and uncontrollable nature of environmental conditions dictate relaxing many restrictive assumptions of most earlier techniques, predominantly developed for operations within "more friendly" terrestrial environments. The critical issue is how to improve robustness while enhancing generality. While robust statistical measures have been introduced to improve the performance in some existing methods, the computational complexity and requirements prohibit real-time performance. To meet these objectives, a framework is envisioned that enables effective use of online real-time recursive estimates for offline dense optimal reconstruction based on global solutions. More precisely, some of our techniques are targeted directly to address real-time performance needs and computational constraints. Yet, their solutions provide estimates of certain parameters that are instrumental for the application of offline global reconstruction techniques.The epiflow framework has been proposed as the central foundation to regulate the infamous ill-posed problem of establishing image correspondence through motion and stereo cooperation, by fully exploiting the geometric and photometric constraints embedded in two pairs of stereo views. The epiflow framework has been successfully applied to stereo feature tracking and stereo fusion. The dual recursive estimation based on EKF aims to address the computation efficiency needs in uncontrolled environments. These provide solutions for online processing, resulting in an accurate estimate of the platform trajectory, scene structure in terms of the 3-D positions of certain prominent features, as well as dense local reconstructions by the application of well-known match propagation. Two improvements to match propagation enables robust dense stereo disparity computation in near real time. Our global method comprises stereo matching by improved belief propagation, where our revised data term provides a more solution to combat local brightness variations due to shading effects that are quite common particularly in underwater video but generally ignored by previous techniques. Employing these new developments, we have formalized an integrated automatic in situ 3-D reconstruction solution, to build 3-D target model from stereo sequences robustly and efficiently in the presence of significant visual artifacts and disturbances. Additionally, other immediate benefits include scalability, flexibility and adaptivity. Extensive experiments with underwater data demonstrate the efficacy of the proposed system.