AbstractThis thesis is concerned with the extraction of dense three-dimensional depth maps from sequences of two-dimensional greyscale images using correlation based matching. In particular the thesis is focused on the noise processes that occur in the depth map and the removal of that noise using nonlinear filters based on fuzzy systems.
The depth from stereo algorithm is reviewed and a widely used correlation based matcher, the Sum Squared Difference (SSD) matcher, is introduced together with an established method of measuring sub-pixel disparities in stereo pairs of images. The noise in the disparity map associated with this matcher is investigated. The conjecture is made that a fuzzy inferencing system can be trained to perform a nonlinear filtering process which is more effective than conventional filters at removing the mixed impulsive and Gaussian-like noise present in the depth map.
Six methods of training fuzzy systems of the Sugeno type based on the simulated annealing algorithm are proposed and tested by training fuzzy systems to approximate a simple function of two variables
The thesis reviews existing fuzzy logic based filters and proposes a taxonomy for such systems. This distinguishes between direct and indirect acting fuzzy filters. An indirect acting fuzzy filter is applied to the task of smoothing a disparity map. The first order Sugeno fuzzy system is then proposed as an architecture that would be suitable as the basis for a direct acting fuzzy filter. This architecture is then applied to the task of smoothing depth maps derived from real and simulated data.
The main contributions of the thesis are the identification of the Sugeno fuzzy system as a form of filter, the proposed training techniques, and the application of fuzzy filters to depth map smoothing.
|Date of Award||1999|
- Fuzzy filters
- Fuzzy Systems