Beddhu Murali, PhD -- Associate Professor

Recent years have seen several new methods that compute disparity values for stereo image pairs in an accurate manner. These are typically area-based methods that enforce smoothness across scan lines. The biggest drawback of these methods is the amount of time it takes to compute disparity values for even moderately sized images. Scan-line-based algorithms, on the other hand, are much faster methods and can be easily implemented in hardware. However, they tend to produce noisy results. The goal of Dr. Murali's efforts is to design a scan-line-based algorithm that is as accurate as other state-of-the-art methods. His min-max SAD (Sum of Absolute Differences) matching algorithm falls under the general category of feature-based matching techniques. According to the level set method, shocks and expansion waves (i.e., discontinuities in the speed function) form at the extrema of the level set function. In the in-between region between two consecutive extrema, the speed function can be assumed to be smooth. It is well-known that the stereo vision problem can be treated as a 1D optical flow problem (i.e., as a 1D inverse level set problem). Thus, in the present algorithm, the region between consecutive extrema is called a front and a scan line is broken into a set of fronts. The strength of a front is defined as the absolute difference between the min and max intensities (or a color metric). The fronts from the right and left scan lines are sorted based on their strengths. Then, using the minimum SAD score of the extremal pixels as the match criterion, the matching process matches the strongest front first and continues until all the visible fronts are matched. A front is visible if it turns to be the match of its match.

Contiguity-Preserving Min-Max Matching Technique for Stereo Vision

This technique employs the idea that if two consecutive fronts in one scan line have visible matches on the other scan line, then in the majority of cases the matches will also be consecutive. So, when the matches have gaps (i.e., are non-contiguous) then it might be due to disocclusion or due to mismatches. These ambiguities are resolved using additional information in the vicinity of the original fronts.

An Upwind Method for Optical Flow Velocity Estimation

Since optical flow constraint is a hyperbolic equation, central differencing-based schemes (without explicit viscous dissipation) for estimating optical flow velocity components would either be unstable or produce spurious oscillations, especially near motion boundaries. However, for developing an upwind scheme, one faces the question: How to upwind the flux computations when the velocity components are the unknowns? Using a level set formulation of the optical flow constraint, this approach uses the local time derivative to upwind the flux computations.

Shoreline Extraction from Topographic and Bathymetric LIDAR Data

Given a seamlessly integrated dataset of topographic and bathymetric data this algorithm extracts the MHW (Mean High Water) shoreline. In order to do this, ENCs (Electronic Nautical Charts) are used to obtain an initial MHW shoreline and are adjusted based on LIDAR data to obtain a final MHW shoreline.

Image Segmentation using Feature-Fronts

Under this project the idea of feature-fronts as defined in the stereo vision algorithm is being used in order to identify object boundaries within an image.


Page URL: http://www.orca.st.usm.edu/murali.php