A VARIATIONAL MODEL WITH ADAPTIVE REGULARIZATION BASED DENSE STEREO MATCHING

摘要

An adaptive regularization based variational model is presented in this work, for obtaining dense disparity map of non-rectified stereo images. To overcome the problems such as to estimate accurate disparities near object boundaries, in repetitive texture regions or textureless regions and in occluded areas, we estimate disparity map by minimizing the global energy functional consists of data and regularizer terms, using variational model with coarse-to-fine pyramidal approach. The pyramidal approach is used to handle large disparities. To optimize the regularizer in the energy fuctional, we use spatially varying regularization parameter instead of a fixed value for the entire image which is common to any variational framework but unsuitable for remotely sensed stereo images because of various image characteristices such as texture. In this approach, we use the pixel wise image gradient and the estimated intermediate disparity gradient to initialize and update the regularization parameter at each pixel location. The initialization consists of K-means clustering in the image gradient space and assignment of a per-class value of regularization. This has impact on the required regularization factor for a group of pixels. Step wise updation is involved at all levels in the pyramid by calculating the disparity in scale space followed by computing the derivative of the disparity map.The proposed method is found to be effective in dealing with the limitation of fixed regularization of the core variational method for increasing the accuracy while estimating the dense disparity map. We evaluate the estimated disparity map quantitatively using bad pixel error with various threshold values comparing with the ground truth. Bad pixel error is calculated considering all the pixels of the input image as well as only nonoccluded pixels.