Off-campus UMass Amherst users: To download dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users, please click the view more button below to purchase a copy of this dissertation from Proquest.

(Some titles may also be available free of charge in our Open Access Dissertation Collection, so please check there first.)

Model reconstruction and refinement using multiple aerial images

Xiaoguang Wang, University of Massachusetts Amherst


Site modeling from aerial imagery is an important research area in computer vision. A key issue in site modeling problems is how to make use of the available multiple images for improving the quality of modeling. This thesis presents a set of multi-image algorithms to increase the amount of information extracted from the input images and to improve the correctness and accuracy of site modeling. In combination with the UMass Ascender and Terrest systems, these algorithms constitute a prototype system which addresses site modeling of both urban sites and natural terrain from aerial imagery. The proposed prototype system is formed with two major processing stages: model reconstruction and model refinement. In the reconstruction stage, increasing the number of participating images is critical to the quality of the reconstructed 3-D feature points. By solving the general 3-D rigid motion problem, the thesis proposes a multi-image feature point correspondence approach that combines isolated image sequences into the reconstruction process. The proposed algorithm is especially good at removing outliers and severely noisy instances from the input data, both of which are major causes of 3-D reconstruction failures. For refinement of urban site models, where a large variety of culturally important structures are of interest, the thesis proposes the concept of “surface microstructure” and establishes a subsystem for extracting such structures. The subsystem is based on an architecture that effectively collects texture information from multiple images and extracts microstructures with a generic appearance. Experimental results show that the subsystem is capable of recovering a wide range of detailed man-made objects in urban sites. For refinement of natural terrain models, where various textural types of ground cover are of concern, classification of the ground surface is the major goal. The thesis proposes the concept of “3-D world texture,” which takes into account the 3-D information in the ground surface texture. A set of 3-D textural features is defined using two-view stereo analysis. It is shown that the proposed 3-D features improve the accuracy of terrain classification over the traditional 2-D image features.

Subject Area

Computer science

Recommended Citation

Wang, Xiaoguang, "Model reconstruction and refinement using multiple aerial images" (2001). Doctoral Dissertations Available from Proquest. AAI3000354.