Research On Stereo Matching Algorithm Of Binocular Vision Based On Deep Learning

Since deep learning is applied to the field of stereo matching,the accuracy of stereo matching algorithm has been greatly improved,but there are still some problems,such as slow running speed,large video memory consumption,deviation from the actual application scene,and low matching accuracy of fast algorithm.To solve the above problems,this thesis has conducted in-depth research from three aspects: capturing the significant features of feature map,enhancing the information interaction between left and right image,and reducing the calculation cost.The main contents are as follows:(1)Aiming at the problems of slow running speed,high video memory consumption and deviation from the actual application scenario caused by preset disparity search range to reduce the calculation amount,a Multiple Attention and Iterate Residual Net(MAIRNet)algorithm is proposed.The algorithm consists of a cross-attention module,which can aggregate the global feature information between the left and right image more effectively and generate disparity maps without disparity range restriction.The iterative residual optimization module,which generates dense disparity maps at the smallest scale,and only generates disparity residual maps when iteratively recovering the disparity resolution step by step,reducing computational cost and video memory consumption.The contextual attention module,which can capture dynamic and static contextual feature information,reducing the number of floating point operations and the number of parameters.(2)Aiming at the problems of coarse-to-fine iterative algorithm matching error accumulation and lack of sufficient feature information for left and right image matching,the Adaptive Iterate Residual optimization Fast Net(AIRFNet)algorithm is proposed.This algorithm consists of a dual attention-guided feature aggregation module that can perform adaptive aggregation of multi-scale feature mappings.The iterate residual optimization module,which generates dense disparity maps at the smallest scale,and only generates disparity residual maps when gradually recovering disparity resolution by iterative means,reducing computational cost and memory consumption.The contextual attention module,which can capture dynamic and static contextual feature information and reduce the number of floating point operations and the number of parameters.In this thesis,the proposed algorithms are evaluated on three datasets,Scene Flow,KITTI2012 and KITTI2015,respectively.The experimental results show that the Compared with AANet algorithm,LEAStereo algorithm and STTR algorithm,the accuracy of MAIRNet algorithm is increased by 0.02% on average,and the running speed is increased by 42% on average,and generates feature maps without disparity range limitation;Compared with AAFS algorithm,RLStereo algorithm and BGNet algorithm,the accuracy of AIRFNet algorithm is improved by 0.2%,and the running speed is improved by 49%,and AIRFNet achieves fast on edge devices operation.