Dense Stereo Matching with Robust Cost Functions and Confidence-based Surface Prior

Kurzfassung

The goal of dense stereo matching is to estimate the distance, or depth to an imaged object in every pixel of an input image; this is achieved by finding pixel orrespondences between the source image and one or more matching images. Applications that make dense stereopsis active come from areas such as photogrammetry, remote sensing, mobile robotics, and intelligent vehicles. For instance, many remote sensing applications require depth maps to generate digital elevation models from airborne and satellite imagery. There are hundreds, if not housands, of approaches seeking to solve stereopsis. A number of factors that make computational stereopsis quite challenging become apparent once one begins to use it in real-world applications; non-Lambertian refectance, complex scene and radiometric changes, among other factors, are usually present in real-world data. Dense stereo algorithms can be categorized into two methodologies based on the way the problem is solved: Local and global. Local algorithms aim to solve the problem via a local analysis at each input-image pixel, whereas global algorithms formulate the stereopsis problem as one of finding an optimal solution to a global energy, or probability function. Developing a global stereo method considers three main factors: calculation reliable observation to measure matching similarity, formulation of energy, or probability, function using additional priors, and optimization of the global function to find the global extremum. In this dissertation two methodical novelties are contributed - the merging strategy of match costs and the confidence-based surface prior incorporating a semiglobal optimization framework. All dense stereo matching algorithms use match cost functions to measure the similarity between two pixels. In a real-world scenario, good radiometric conditions are often disrupted by complicated and dynamic lighting sources, inappropriate camera configuration, and non-Lambertian reflectance of objects. We investigate the interdependencies among matching erformance, cost functions, and observation conditions using both close-range and remote-sensing data. Our cost-merging strategy combines the advantages of different match cost functions and gives consideration to imagery configurations. In addition, a novel probabilistic surface prior is introduced incorporating a new energy optimization method, called iSGM3. Our approach builds a probabilistic surface prior over the disparity space using confidences on a set of reliably matched correspondences. Unlike many region based methods, our method defines an energy formulation over pixels, instead of regions in a segmentation; this results in a decreased sensitivity to the quality of the initial segmentation. This dissertation suggests the way to developing robust stereo methods is on the level of obtaining costs and suitable energy formulation, and not only the energy optimization. Both costs merging and the surface prior are generally applicable for almost all extended stereo methods.