Research And Implementation Of Stereo Matching Algorithm Based On Deep Learning

Stereo matching,as a key technique in binocular stereo vision,is used to find matching point pairs in a pair of corrected images and has important applications in 3D reconstruction,automatic driving and vision measurement.With the development of deep learning,the performance of learning-based stereo matching has been continuously improved,but there are still deficiencies.In the feature extraction stage,it is difficult for the current algorithm to combine less computational consumption,rich context information of features,and greater similarity between features of the same-named points.In the cost aggregation stage,when the cost volume is a joint cost volume constructed by connecting different cost volumes,if it is directly regularized,the advantages of different cost volumes cannot be fully utilized,because the data distribution of different cost volumes is very different.In the disparity refinement stage,the popular algorithm provides the refinement network with the implicit confidence of the initial disparity estimation through the Warp operation,but if the refinement network can be provided with the explicit confidence of the initial disparity estimation,it will be able to better guide the refinement network makes correct adjustments to the initial disparity to improve prediction accuracy.In view of the above shortcomings,the work of this paper is as follows:(1)Aiming at the shortcomings in the feature extraction stage,this paper proposes stereo matching network based on attention fusion.Attention fusion consists of multiple attention fusion modules,which consist of full attention and convolution.Full attention calculates the attention weights once on the epipolar lines of the left and right feature maps,and then uses the weights to update the left and right feature maps simultaneously.Since the full attention is only calculated within the epipolar lines,a convolution is added after it in order to capture the contextual information between the epipolar lines.Attention fusion uses less computation,enabling greater similarity of feature pairs with the same name while capturing rich contextual information.Applying the proposed attention fusion to GwcNet,EPE and D1 are reduced by 23.1%and 24.6%.(2)Aiming at the shortcomings in the cost aggregation stage,this paper proposes stereo matching network based on two-stage cost aggregation.The two-stage cost aggregation first regularizes different cost volumes separately,so that their data distribution tends to be consistent.Then use the regularized cost volume to construct the joint cost volume and regularize it to make full use of the data of different cost volumes.Experiments show that the EPE is reduced by 18.2%with a 21.7%reduction in the number of parameters.Better performance is achieved with fewer parameters.(3)Aiming at the shortcomings in the disparity refinement stage,this paper proposes stereo matching network based on information entropy uncertainty estimation.In this paper,we first propose an uncertainty estimation based on information entropy,which is used to evaluate the uncertainty of the disparity probability distribution and serve as an explicit confidence measure for the initial disparity estimation.Then,the information entropy is added to the refinement network during disparity refinement to guide the refinement network to make more accurate adjustments to the initial disparity.The information entropy refinement network proposed based on information entropy can be seamlessly embedded to the model based on cost volume,and the performance of the model can be greatly improved with only a few parameters and calculations.The model combined with attention fusion,two-stage cost aggregation,and information entropy refinement network achieves state-of-the-art performance with EPE of 0.46 on Scene Flow,and it achieves competitive performance on KITTI benchmark.(4)In this paper,we design and implement a deep learning-based stereo matching system that combines attention fusion,two-stage cost aggregation,and information entropy refinement network.The system provides users with a variety of functions.Users can not only directly use the system to do high-precision reasoning,but also reconfigure and train the network model to meet actual needs.