Research On Binocular Stereo Matching Fast Algorithms

Binocular stereo vision is an important branch of computer vision,and has a wide range of practical applications,such as 3D stereo display,automatic driving,robot navigation and intelligent factory,etc.High-quality real-time stereo matching has the potential to enable various computer vision applications.This paper introduces the basic steps and classification methods of the stereo matching algorithm,and then researches on the related algorithms according to the three important steps of stereo matching: matching cost computation,cost(support)aggregation,disparity optimization.The paper presents a Census-based fast stereo matching algoritm by making experiments and analysis of ASW,SNCC,ADCensus,CNN and IDR algorithm.It aggregates with two-pass approximation of adaptive support-weight and updates in a low-complexity iterative disparity refinement technique to produce the disparity results.It has good performance in both accuracy and speed listed on the latest Middlebury stereo benchmark 3.There are two contributions of this algorithm.Firstly,it introduces a Threshod into the traditional census transform,which improves the ability of anti-noise.Sencondly,it combines the random sampling(Brief)measure and the traditional census transform,which enlarges the window without more computation.Experimentally,our cost function is better than others in matching results and robustness.We have finished the binocular stereo vision system in embedded system Jetson TK1 with OpenCV stereo calibration and OpenGL 3D reconstruction.And the stereo matching algorithm has been implemented on parallel high-performance graphics hardware(GPU)provided with Jetson TK1.We also have made sone experimental evaluation on the system;it is proved that the system has practical value in speed and precision.Therefore,the system lays the foundation for the practical application based on binocular stereo vision.