Research On Stereo Matching Of Binocular Stereo Vision

Binocular stereo matching is an important branch of stereo vision.The main factors that affect the accuracy of stereo matching are divided into two categories: one is external factor,such as lens distortion;the other is internal factor,such as the unreasonable construction of the matching window.Firstly,this paper proposes a camera calibration algorithm based on the improved particle swarm optimization.By introducing a first-order tangential distortion model and an improved particle swarm optimization algorithm,image correction and parameter calibration are implemented,which lays the foundation for stereo matching.Subsequently,a Simplified ICA Based Local Similarity Stereo Matching Algorithm(SILSSM)is proposed to improve Disp Net C.The Simplified Independent Component Correlation Algorithm(SICA)cost aggregation was introduced,which combined the matching cost volume pyramid and SICA loss function.Next,the regionwise loss function combined with the pixel-wise loss function,is defined as a local similarity loss function Finally,the SICA loss function combined with the local similarity loss function,which is defined to estimate the disparity map.The main contents of this paper are as follows:(1)For the problem that Zhang' calibration algorithm is easy to fall into the local optimum,this paper introduces the first-order tangential distortion based on the original second-order radial distortion to form a mixed distortion model.At the same time,based on the the particle swarm optimization algorithm,the convergence degree judgment and roulette wheel selection stratage is introduced to avoid premature convergence of the algorithm.(2)For the problem of insufficient utilization of context information in disparity prediction,the deconvolution results of different layers are stacked during the decoding stage of the Disp Net C to form a matching cost volume pyramid to obtain multi-scale information.(3)For the problem of unsatisfactory prediction of disparity maps in weakly textured regions,a loss function inspired by ICA is introduced,combined with a pre-built matching cost volume pyramid to obtain more accurate weights,and form a complete SICA matching cost aggregation process.(4)For the problem that the edge of the disparity map is not clear,the region-wise loss function is introduced.In combination with the pixel-wise loss function,a local similarity loss function is defined to improve the spatial structure of the disparity map.