Date of Award


Degree Name

Master of Science


Computer Science


Dr. Patricia J. Teller

Second Advisor

Dr. Sarala Arunagiri


Stereo vision or stereopsis is a biological process in which the impression of the depth of a scene viewed by the two eyes is perceived. Stereo vision helps us to see where objects are in relation to our bodies. Computational stereo describes the process of synthesizing the mechanics of binocular vision; and stereo matching is an important process in the problem of computational stereo.

Stereo matching is used for Digital Elevation Modeling (DEM), which is the method of extracting 3D information from digital images obtained by cameras and, therefore, is a crucial process for representing a 3D view of surfaces like terrains, planets, and asteroids. There are many applications of stereo matching including extraction of weak targets in a cluster, automatic target detection, and vision-based obstacle detection and avoidance for unmanned aerial vehicles (UAVs). Stereo matching is a computationally complex problem and, therefore, has been one of the most heavily investigated topics in computer vision.

Scene conditions have a considerable influence on the performance of stereo matching algorithms. Global optimization approaches are more affected by this than other approaches. This is because the recognition of contour edges for object segmentation is crucial for stereo matching. In most cases, the object contours are not perfectly recovered in the disparity map. Thus, enhancing the output quality of stereo matching by reducing the errors at the edges of the generated image, as compared to the original image, i.e., the ground truth, is the first issue addressed in this thesis. The second issue addressed is the well-understood problem of performance tuning global stereo matching algorithms. In particular, for two global cost functions based on normalized cross correlation, NCC and MS1-NCC, which have three parameters, i.e., the scaling factor, lambda, and weight, we demonstrate the effectiveness of the use of an input-dependent scaling factor in contrast to the use of a fixed scaling factor.

For the three minimization algorithms, simulated annealing, graph cut, and graph cut one iteration, using input-dependent scaling factors in the stereo matching process produces better quality disparity maps than fixed scaling factors. For all the image pairs considered, the generated images have less error both with respect to RMS as well as pixel counts. Also, for all the image pairs considered, compared to the original NCC cost function, the modified cost function MS1-NCC, which includes edge information, enhances the image quality and reduces the RMS errors as well as the total pixel mismatch counts. In conclusion, applying the two proposed modifications to the stereo matching process enabled us to get well defined crisper images.




Received from ProQuest

File Size

89 pages

File Format


Rights Holder

Aritra DattaGupta