A comparative study of fast dense stereo vision algorithms

With recent hardware advances, real-time dense stereo vision becomes increasingly feasible for general-purpose processors. This has important benefits for the intelligent vehicles domain, alleviating object segmentation problems when sensing complex, cluttered traffic scenes. In this paper, we presents a framework of real-time dense stereo vision algorithms that all based on a SIMD architecture. We distinguish different methodical components and examine their performance-speed trade-off. We furthermore compare the resulting algorithmic variations with an existing public source dynamic programming implementation from OpenCV and with the stereo methods discussed in Sharstein and Szeliski's survey. Unlike the previous, we evaluate all stereo vision algorithms using realistically looking simulated data as well as real data, from complex urban traffic scenes.