Doctor of Philosophy
Electrical and Computer Engineering
Dr. Jagath Samarabandu
Integrating images taken at street-level with satellite imagery is becoming increasingly valuable in the decision-making processes not only for individuals, but also in business and governmental sectors. To perform this integration, images taken at street-level need to be accurately georeferenced. This georeference information can be derived from a global positioning system (GPS). However, GPS data is prone to errors up to 15 meters, and needs to be corrected for the purpose of geo-referencing. In this thesis, an automatic method is proposed for correcting the georeference information obtained from the GPS data, based on image registration techniques. The proposed method uses an optimization technique to find local optimal solutions by matching high-level features and their relative locations. A global optimization method is then employed over all of the local solutions by applying a geometric constraint.
The main contribution of this thesis is introducing a new direction for correcting the GPS data which is more economical and more consistent compared to existing manual method. Other than high cost (labor and management), the main concern with manual correction is the low degree of consistency between different human operators. Our proposed automatic software-based method is a solution for these drawbacks.
Other contributions can be listed as (1) modified Chamfer matching (CM) cost function which improves the accuracy of standard CM for images with various misleading/disturbing edges; (2) Monte-Carlo-inspired statistical analysis which made it possible to quantify the overall performance of the proposed algorithm; (3) Novel similarity measure for applying normalized cross correlation (NCC) technique on multi-level thresholded images, which is used to compare multi-modal images more accurately as compared to standard application of NCC on raw images. (4) Casting the problem of selecting an optimal global solution among set of local minima into a problem of finding an optimal path in a graph using Dijkstra's algorithm.
We used our algorithm for correcting the georeference information of 20 chains containing more than 7000 fisheye images and our experimental results show that the proposed algorithm can achieve an average error of 2 meters, which is acceptable for most of applications.
Zouqi, Mehrnaz, "Improving the Geotagging Accuracy of Street-level Images" (2013). Electronic Thesis and Dissertation Repository. 1202.