Algorithmic Considerations for Real-Time Stereo Vision Applications
Kristian Ambrosch, Christian Zinner, Wilfried Kubinger · 2009
Real-time stereo vision is a very resource intensive application, requiring a high computational performance. Therefore, we analyze the well known Census Transform not only for an increase in accuracy, but also for a reduction in complexity. We propose a novel approach, using the Modified Census Transform on the intensity as well as the gradient images, that can be efficiently combined with a sparse computation. Our evaluation of this approach on the images of the Middlebury stereo ranking shows that it allows scaling the algorithm’s complexity down by a factor of 5.8, while still being more accurate than the original transform. 1