Doctor's Theses (authored and supervised):
"Design and Evaluation of Stereo Matching Techniques for Silicon Retina Cameras";
Supervisor, Reviewer: M. Gelautz, J. Scharinger;
Institut für Softwaretechnik und interaktive Systeme,
oral examination: 2016-04-29.
Nowadays, techniques for 3D reconstruction that are used in a variety of computer vision applications need to account for the 3D structure of a real-world scene. This task is often performed using a stereo vision system which consists of two digital cameras observing the same scene from two different viewing angles. A major challenge in stereo vision is the stereo matching problem, which involves finding corresponding pixels that are projections of the same scene point in the image pair. While stereo matching of images delivered by conventional cameras has been the subject of intense research for many years, this thesis focuses on the analysis of stereo data delivered by a different type of digital camera - a Silicon Retina sensor - whose stereo processing capabilities have been addressed by only few publications thus far.
The special analog pixel design of a silicon retina camera enables a high dynamic range of light and very fast pixel updates. Unlike a conventional camera, the silicon retina camera´s sensor pre-processes the information on-chip, and only transmits pixels that capture a change of light. This significantly reduces the amount of data that must be transferred and processed. However, as the process yields visual information different to a normal digital image, the data poses new challenges for solving the correspondence problem occurring in a silicon retina stereo set-up. In this thesis, we first analyze the data from a silicon retina stereo sensor and study its behavior in order to assess the impact of various algorithms on this data. Then, based on these results, we design and implement new kinds of stereo matching algorithms to overcome the imposed challenges of silicon retina data. Besides the core stereo matching algorithms,
we develop and evaluate different approaches to improve the accuracy of the stereo matching algorithms. Additionally, we design a method to generate ground truth data to better evaluate the calculated depth data; this enables meaningful discussions and interpretation of the generated stereo matching output.
Stereo Matching, Address Events, Silicon Retina Cameras
Electronic version of the publication:
Created from the Publication Database of the Vienna University of Technology.