Research On Binocular Vision Stereo Matching Technology

Binocular stereo vision technology is an important area for the research of computer vision. By imitating the perception of the human eye, we can obtain three-dimensional structure of the object to be observed, determining the accurate position in the real world. The acquisition of location in the 3D world makes it possible to recognize objects, reconstruct 3D objects, and navigate robotics. It is also a great impetus to the development of artificial intelligence. Our major research content and the achievements .are as follows:1)From the aspects of theoretical derivation and experimental data, we have studied the robust of different costs for passive binocular stereo methods with respect to radiometric variations of the input images, including exposure differences, varying lighting and vignetting. The study includes two categories: matching cost computing based on parametric transform or non-parametric transform, such as brightness, gradient, Census transform,normalized cross correlation and the derivations. The global and local cost aggregation algorithms are used to smooth the cost value, which are based on guided filter, box filter, minimum spanning tree and fast global smoother constructed with weighted least squares. Meanwhile, we use the simulated and real radioactive differences to evaluate the matching cost computing algorithms, and finally the same experimental results are achieved.2) We presents a state-of-the-art global stereo matching framework based on binocular vision. The fusion of brightness and gradient of input images harnessing three color channels is used to compute matching cost value which is taken as the data item of the global energy function, and the weighted difference of the adjacent pixels is taken as the smooth term of the energy function. A fast global smoothing filter based on weighted least squares is the solver to the objective function, and the disparity map need to be refined using the weighted median filter with guided image. We show that such solutions can be efficiently achieved by the specific framework, but also very effective.3) We harness the parallel capacity of GPU to optimize the newly proposed global stereo matching framework. Compared with the traditional algorithm, the specific parallelization method can improve the computing speed and ensure the real-time effect.