Stereo vision based on compressed feature correlation and graph cut

Abstract

This dissertation has developed a fast and robust algorithm to solve the dense correspondence problem with a good performance in untextured regions by merging Sparse Array Correlation from the computational fluids community into graph cut from the computer vision community. The proposed methodology consists of two independent modules. The first module is named Compressed Feature Correlation which is originated from Particle Image Velocimetry (PIV). The algorithm uses an image compression scheme that retains pixel values in high-intensity gradient areas while eliminating pixels with little correlation information in smooth surface regions resulting in a highly reduced image datasets. In addition, by utilizing an error correlation function, pixel comparisons are made through single integer calculations eliminating time consuming multiplication and floating point arithmetic. Unlike the traditional fixed window sorting scheme, adaptive correlation window positioning is implemented by dynamically placing strong features at the center of each correlation window. A confidence measure is developed to validate correlation outputs. The sparse depth map generated by this ultra-fast Compressed Feature Correlation may either serve as inputs to global methods or be interpolated into dense depth map when object boundaries are clearly defined. The second module enables a modified graph cut algorithm with an improved energy model that accepts prior information by fixing data energy penalties. The image pixels with known disparity values stabilize and speed up global optimization. As a result less iterations are necessary and sensitivity to parameters is reduced.

(cont.) An efficient hybrid approach is implemented based on the above two modules. By coupling a simpler and much less expensive algorithm, Compressed Feature Correlation, with a more expensive algorithm, graph cut, the computational expense of the hybrid calculation is one third of performing the entire calculation using the more expensive of the two algorithms, while accuracy and robustness are improved at the same time. Qualitative and quantitative results on both simulated disparities and real stereo images are presented.