A Near Real-Time Color Stereo Matching Method for GPU
Naiyu Zhang, Hongjian Wang, Jean-Charles Créput, Julien Moreau, Yassine Ruichek · 2013
Abstract—This paper presents a near real-time stereo matching method with acceptable matching results. This method consists of three important steps: SAD-ALD cost measure, cost aggregation in adaptive window in cross-based support regions and a refine-ment step. These three steps are well organized to be adopted by the GPU’s parallel architecture. The parallelism brought by GPU and CUDA implementations provides significant acceleration in running time. This method is tested on six pairs of images from Middlebury dataset, each possibly declined within different sizes. For each pair of images it can generate acceptable matching results in roughly less than 100 milliseconds. The method is also compared with three GPU-based methods and one CPU-based method on increasing size image pairs.