Research On End-To-End Binocular Stereo Matching Algorithm Based On Deep Learning

Binocular stereo vision has the characteristics of easy expansion,low cost and strong scene adaptability in acquiring scene depth information,so it is widely used in many frontier fields such as unmanned driving and augmented reality.Compared with traditional methods,the method based on convolutional neural network can significantly improve the accuracy and operating efficiency of stereo matching.The algorithm in this paper solves the current problems of binocular stereo matching from the following aspects.The current stereo matching work ignores the problem that the aggregation network needs more multi-scale context information for feature similarity learning.This paper proposes a 3D convolutional aggregation network which integrates the unique idea of hole space pyramid pooling,and optimizes the processing effect of stereo matching in the inherent ill conditioned region.There is an imbalance between the number of pixels in the non-occlusion area(simple samples)and the number of pixels in the occlusion area(difficult samples)of the stereo image data set.The loss function Balanced L1 Loss is introduced to replace the original Smooth L1 Loss of the benchmark model to perform end-to-end supervised training of the model,prompting the model to strengthen the learning of pixels in the unoccluded area that occupy most of the stereo image data set.Aiming at the defect that most of the current cost construction methods lack sufficient similarity measurement features for subsequent aggregation networks,a group-by-group AD^2 cost construction method is proposed.The group-by-group AD^2 cost can clearly determine the similarity between the corresponding element values of the left and right unary features,and provide sufficient feature similarity guidance for the subsequent aggregation network.In addition,a 3D convolutional aggregation network based on feature reuse is constructed to suppress the phenomenon of image feature loss,which solves the problem of image feature loss caused by the feature map in the down-sampling stage of different hourglass modules,which makes it difficult for the model to perform more accurate disparity estimation.