Research On Binocular Vision Ranging System Based On Deep Learning

Binocular vision ranging technology is one of the hot research topics in the field of computer vision with broad applications.This technology uses binocular cameras to capture the same target and obtains disparity information through various algorithms to perform distance measurement based on the principle of triangle similarity.However,traditional methods are prone to error in estimating disparity in low texture areas,leading to reduced ranging accuracy.Therefore,this study is based on the theory of deep learning and focuses on improving the stereo matching algorithm to enhance the accuracy of matching and obtain high-quality disparity,thus further improving the accuracy of binocular vision ranging system.The main research achievements are as follows:(1)To address the issues of low feature point detection in areas with low texture and the inability to effectively recover disparity of small objects,this paper proposes a disparity iterative update stereo matching algorithm based on Transformer feature optimization(TUNet).After feature extraction,the attention mechanism and position encoding in the Transformer algorithm are introduced to transform the extracted local features into contextdependent and position-dependent features,effectively improving the matching quality in low texture areas.Additionally,the correlation of features is calculated at different resolutions,and the gated recurrent unit(GRU)is used to perform correlation search and disparity iterative updates at multiple resolutions,resulting in a more accurate disparity.(2)To address the problem of decreased cross-domain feature consistency,this paper proposes a generalized stereo matching network based on whitening loss(ATUNet).The whitening loss function is introduced during feature extraction by calculating the feature covariance of the left and right images and selecting pixels with significant changes in the image.As the loss function decreases,the stereo matching network relies less on changesensitive features to form feature representations,thus improving the model’s feature generalization ability across multiple datasets.(3)According to the basic principles of binocular vision ranging,this study used a binocular camera to capture images,completed camera calibration using the Zhang calibration method in the MATLAB platform,and then studied the three stages of stereo rectification,stereo matching,and depth estimation in the binocular vision ranging system using the Python in the Py Charm development platform.The experimental results showed that the binocular vision ranging system constructed in this study can obtain distance information of target objects and achieve high precision ranging within a certain range.Therefore,this study has good application prospects.