A Very Fast Census-Based Stereo Matching Implementation on a Graphics Processing Unit

摘要

In this paper a very fast graphics processing unit implementation of a local, census-correlation-based stereo matching algorithm is presented. In comparison to absolute or squared difference correlation techniques, the census transform is computational more expensive which led to the motivation of a GPU-based implementation. Due to the parallel architecture of modern graphics cards, complex algorithms can be executed very efficiently. Thus, this work deals with the question how to use a GPU for high speed and high quality stereo matching. The proposed implementation achieves 75.7 fps at an image resolution of 640 × 480 and a disparity search range of 50 on a NVIDIA GeForce GTX 280 graphics card.