Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Creating an accurate depth map has several, valuable applications including augmented/virtual reality, autonomous navigation, indoor/outdoor mapping, object segmentation, and aerial topography. Current hardware solutions for precise 3D scanning are relatively expensive. To combat hardware costs, software alternatives based on stereoscopic images have previously been proposed. However, software solutions are less accurate than hardware solutions, such as laser scanning, and are subject to a variety of irregularities. Notably, disparity maps generated from stereo images typically fall short in cases of occlusion, near object boundaries, and on repetitive texture regions or texture-less regions. Several post-processing methods are examined in an effort to combine strong algorithm results and alleviate erroneous disparity regions. These methods include basic statistical combinations, histogram-based voting, edge detection guidance, support vector machines (SVMs), and bagged trees. Individual errors and average errors are compared between the newly introduced fusion methods and the existing disparity algorithms. Several acceptable solutions are identified to bridge the gap between 3D scanning and stereo imaging. It is shown that fusing disparity maps can result in lower error rates than individual algorithms across the dataset while maintaining a high level of robustness.
Kerr, Douglas E. Jr., "Comparative Study of Model-Based and Learning-Based Disparity Map Fusion Methods" (2020). Graduate Theses, Dissertations, and Problem Reports. 7675.