Real-Time Semi-Global Matching on the CPU
Abstract
Among the top-performing stereo algorithms on the Middlebury Stereo Database, Semi-Global Matching (SGM) is commonly regarded as the most efficient algorithm. Consequently, real-time implementations of the algorithm for graphics hardware (GPU) and reconfigurable hardware (FPGA) exist. However, the computation time on general purpose PCs is still more than a second. In this paper, a real-time SGM implementation on a general purpose PC is introduced. Parallelization and image subsampling is used while ensuring the full disparity resolution for small disparities. This approach is especially beneficial for robotic and driver assistance systems. The system is able to compute 640x320 image pairs at more than 14Hz, leaving also computational resources for subsequent processing such as free space computation or object detection. The algorithmic approach is portable to multi-core embedded processors.
Cite
Text
Gehrig and Rabe. "Real-Time Semi-Global Matching on the CPU." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2010. doi:10.1109/CVPRW.2010.5543779Markdown
[Gehrig and Rabe. "Real-Time Semi-Global Matching on the CPU." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2010.](https://mlanthology.org/cvprw/2010/gehrig2010cvprw-realtime/) doi:10.1109/CVPRW.2010.5543779BibTeX
@inproceedings{gehrig2010cvprw-realtime,
title = {{Real-Time Semi-Global Matching on the CPU}},
author = {Gehrig, Stefan K. and Rabe, Clemens},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
year = {2010},
pages = {85-92},
doi = {10.1109/CVPRW.2010.5543779},
url = {https://mlanthology.org/cvprw/2010/gehrig2010cvprw-realtime/}
}