Stereoscopic depth estimation for online vision systems

The human visual perception heavily depends on stereoscopic vision. By fusing the two views that our eyes provide, a 3-D sensation of the surrounding is generated. It is therefore natural to assume that machine vision systems also benefit from a comparable sense. A lot of work has been done in the area of machine stereo vision, but a severe drawback of today’s algorithms is that they either achieve high accuracy and robustness by sacrificing real-time speed or they are real-time capable but with major deficiencies in quality. The goal of this thesis is to tackle the problem of efficient, real-world, stereoscopic depth processing. In particular, the processing should be lightweight enough to be processed on mobile platforms. To this end, two new methods are introduced that have a very good balance between speed and accuracy. First, the summed normalized cross-correlation (SNCC) is proposed which constitutes a new cost function for block-matching stereo processing. In contrast to most standard cost functions it hardly suffers from the fattening effect while being computationally very efficient. Evaluations on standard benchmarks show that SNCC brings the standard local block-matching stereo very close to the performance of global optimization methods based on graph cut or belief propagation. As an additional benefit for real-time implementation, a multi-core parallel processing scheme for block-matching stereo is discussed which allows for a runtime gain that is linear in the number of processing cores. Second, the direct surface fitting, a new algorithm for fitting parametric surface models to stereo images is introduced. This algorithm is inspired by the homography-constrained gradient descent methods which are frequently used to estimate the orientation and depth of planar surfaces. By replacing the gradient descent search with the direct search method of Hooke-Jeeves the fitting of any parametric surface model with arbitrary cost functions becomes feasible. A comparison on standard benchmarks shows that the direct surface fitting has a depth