Research Of Binocular Stereo Matching Method Based On CNN

Binocular stereo vision technology is a popular research direction in computer vision,and is widely used in robot navigation,3D reconstruction and other fields.The accuracy of the binocular stereo matching method has a direct impact on the effect of stereo vision,so improving the matching accuracy in binocular stereo matching is of great significance to the development of computer vision.This subject mainly studies the related methods of binocular stereo matching.Firstly,the basic principle of binocular stereo matching and research background at home and abroad are introduced.Secondly,a CNN-based binocular stereo matching method is designed and the network structure performance is experimentally analyzed.The main research contents of the paper can be summarized as follows:(1)A residual network structure based on 2D convolution is designed to calculate the matching cost.The structure of the twin network is used to share the weights in the network structure.At the same time,the 2D convolution-based residual network structure is used to extract the context feature information,which ensures that the extracted feature information has the correlation information of the binocular view.(2)A cost aggregation module based on 3D convolution and 3D deconvolution is designed.Using 3D convolution based on the residual network structure to perform cost aggregation on the initial matching cost set not only expands the receptive field of the convolution kernel,but also increases the amount of feature information that the network model can learn.Since 3D convolution increases the amount of calculation of the network,the method of downsampling first and then upsampling not only reduces the amount of calculation,but also restores the information in the small-size feature map after downsampling.(3)Multi-directional aggregation of semi-global stereo matching algorithm is added to the end-to-end network structure.Compared with the cost aggregation method of 3D convolution,the aggregation principle using semi-global stereo matching can obtain a larger aggregation window,and the end-to-end network structure can adaptively provide the required weights and parameters for the semi-global matching algorithm.,It is more flexible and efficient than the fixed weight aggregation method of 3D convolution.In order to prevent over-fitting problems during the training process,transfer learning is used to experiment with the network structure of the project design.The data in SceneFlow is pre-processed,and the KITTI training image set is used to fine-tune and analyze after obtaining the pre-processing model The influence of network structure parameters on network running time and mismatch rate.At the same time,different algorithms are used to compare experiments with the proposed algorithm to compare the error percentage and effect of the disparity map.The experimental results show that the method designed by the subject can effectively complete the task of stereo matching and obtain a more accurate disparity map.