To contribute a novel approach in the field of image registration and panorama creation, this algorithm foregoes any scene knowledge, requiring only modest scene overlap and an acceptable amount of entropy within each overlapping view. The weighted and filtered mutual information (WFMI) algorithm has been developed for multiple stationary, color, surveillance video camera views and relies on color gradients for feature correspondence. This is a novel extension of well-established maximization of mutual information (MMI) algorithms. Where MMI algorithms are typically applied to high altitude photography and medical imaging (scenes with relatively simple shapes and affine relationships between views), the WFMI algorithm has been designed for scenes with occluded objects and significant parallax variation between non-affine related views. Despite these typically non-affine surveillance scenarios, searching in the affine space for a homography is a practical assumption that provides computational efficiency and accurate results, even with complex scene views. The WFMI algorithm can perfectly register affine views, performs exceptionally well with near-affine related views, and in complex scene views (well beyond affine constraints) the WFMI algorithm provides an accurate estimate of the overlap regions between the views. The WFMI algorithm uses simple calculations (vector field color gradient, Laplacian filtering, and feature histograms) to generate the WFMI metric and provide the optimal affine relationship. This algorithm is unique when compared to typical MMI algorithms and modern registration algorithms because it avoids almost all a priori knowledge and calculations, while still providing an accurate or useful estimate for realistic scenes. With mutual information weighting and the Laplacian filtering operation, the WFMI algorithm overcomes the failures of typical MMI algorithms in scenes where complex or occluded shapes do not provide sufficiently large peaks in the mutual information maps to determine the overlap region. This work has currently been applied to individual video frames and it will be shown that future work could easily extend the algorithm into utilizing motion information or temporal frame registrations to enhance scenes with smaller overlap regions, lower entropy, or even more significant parallax and occlusion variations between views.
Library of Congress Subject Headings
Video recordings--Data processing; Image registration; Image processing--Digital techniques; Video surveillance--Data processing
Department, Program, or Center
Microelectronic Engineering (KGCOE)
Keane, Thomas, "Weighted and filtered mutual information: A Metric for the automated creation of panoramas from views of complex scenes" (2011). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus