Extended hough methodology using bayesian theory
Jihun Cha, Rufus H. Cofer · 2002
Feature detection and extraction plays a key role in modern computer vision systems; however, few advances have been made to date demonstrating their full potential with respect to real-world complex imagery. As they are intermediate-level computer vision processes, feature detection and extraction are situated between yet lower-level and even higher-level vision processes. The ultimate goal of feature detection and extraction is thus to extract features based on the evidence provided by the lower-level vision processes and to pass the results on to the higher-level vision processes forming scene content hypotheses. Because lineal features are a principle constituent of man-made objects, they are considered as one of the prime cartographic features within overhead imagery. Thus this dissertation focuses on lineal feature extraction via optimization of a related Hough algorithmic concept. With evident potential for parallel computation and robustness to noise/occlusion, the Hough Transform has been previously investigated by many researchers in attempts to overcome its main deficiencies—that of high computation and storage requirements. This dissertation presents a new and novel technique that largely solves these problems while insuring higher accuracy of the extended features. The specifically novel approach of this work lies in the construction of two separate Hough spaces to overcome the Hough Transform's classical inability to detect vertical lines of its slope-intercept representation and to allow more efficient implementation. This is done by adopting a third parameter, that of x-coordinate of an image pixel, along with the usual two parameters, slope and y-intercept of the more classical Hough algorithm. This third parameter also provides the new ability of adapting to the length of the lineal feature to allow even higher accuracy and faster feature search capabilities. By accommodating the additional dimension, the Hough transform also can acquire a higher level of regularity in 3D Hough space. This allows clear prediction of the quantity and the location of required information in 3D Hough space needed to search for the next piece of evidence of the object. Not only does this provide a faster sequential search but also a high adaptability to hardware implementation.