Research On Matching Algorithm Based On Convolutional Neural Network

Binocular stereo vision has the characteristics of easy expansion,low cost and strong adaptability in obtaining scene depth information,so it is widely used in many frontier fields such as unmanned driving and augmented reality.In recent years,methods based on deep learning have been extended to geometric problems such as stereo matching.Compared with traditional methods,the convolutional neural network method can significantly improve the accuracy and efficiency of stereo matching.The algorithm in this paper solves the existing problems of binocular stereo matching from the following two aspects.(1)Aiming at the problem of low matching accuracy of stereo matching algorithm in weak texture region,this paper designs a stereo matching algorithm based on multi-scale feature extraction,FPD-Net,by combining feature pyramid with cavity convolution.The fusion of feature pyramid and void convolution can not only realize multi-scale feature extraction,but also increase the receptive field of high-level features,obtain more global features,and effectively reduce the loss of details of small size objects.Secondly,in order to speed up the operation speed of stereo matching,the idea of Patch Match is introduced into the algorithm.By pruning the feature map,unlikely matching areas are removed,and sparse cost is obtained,which greatly reduces the computation and memory consumption of the algorithm.Compared with PSMNet,which is also a multi-scale feature extraction method,the proposed algorithm not only improves the accuracy of stereo matching,but also speeds up the algorithm.(2)Aiming at the mismatching problem of current end-to-end stereo matching algorithm in complex scenes,this paper proposes a stereo matching algorithm AFPD-Net based on multi-scale attention feature fusion.By the attention of both spatial information and channel information mechanism and combining the characteristics of the pyramid,improved algorithm in the feature extraction stage strengthen interested area extracting intensity,and suppress unwanted noise,capture abundant global context information and long range dependent,reduce matching error rate in the pathological area algorithm and makes the algorithm can meet the challenges of complex scenes.Secondly,in order to reduce the complexity of the model and reduce the computation of the algorithm,the nonlinear activation function in the network is replaced with a nonlinear function that can be activated adaptively,which can learn whether to activate adaptively,so as to improve the generalization ability of the network.In this paper,the effectiveness of the proposed algorithm is verified on three public binocular vision datasets,Scene Flow,KITTI2015 and KITTI2012.Because the PSM-Net algorithm is very similar to the PFD-Net algorithm proposed in this paper,the two algorithms are compared.Experiments show that the accuracy of the proposed FPD-Net algorithm is higher than that of PSM-Net algorithm,and the calculation speed is also improved.Since the AFPD-Net algorithm is improved on the FPD-Net algorithm,the two algorithms are also compared.Experimental results show that the accuracy of the improved algorithm is significantly higher than that of FPD-Net algorithm.