Study On Stereo Matching Algorithm Based On Convolutional Neural Network

Binocular stereo vision technology refers to the use of a computer or related device to simulate a biological vision system,enabling it to perceive three-dimensional objects in a realistic scene from a two-dimensional planar image,which is an important form of machine vision.At present,binocular stereo vision technology is widely used in many fields such as robot navigation and control,vehicle automatic driving,and three-dimensional scene information reconstruction.This paper focuses on the calculation process of matching cost based on convolutional neural network,and designs a stereo matching network architecture based on asymmetric pyramid pooling.The main research work is as follows:(1)This paper proposes an asymmetric pyramid pooling algorithm.In the image processing process,the algorithm adaptively adjusts the pooling mode according to the image size to extract the image features comprehensively,and provides detailed feature information for the subsequent stereo matching steps.Combined with the Siamese Neural Network,the local area(image block)of the left and right views is used as the input of the network,and the initial matching cost of the image block is obtained through the asymmetric pyramid pooling network and the dot product operation.(2)A convolutional neural network structure with multi-scale extraction features is designed.After the improved pyramid pooling layer,two additional layers of convolutional layer are superimposed for multi-scale information fusion and parameter reduction,and the upsampling operation is used to recover.The original features of the image,the cascading operation integrates the feature information.In addition,two sets of contrast experiments were set to analyze the influence of additional convolutional layer parameters on network matching accuracy and convergence speed.Three sets of convolutional neural network depth contrast experiments were set to determine the depth of stereo matching network.Finally,all experiments are combined to select the optimal parameter setting network structure,and the network model is trained to calculate the matching cost of the image pair.(3)Finally,combined with the subsequent steps of stereo matching,the whole stereo matching algorithm based on convolutional neural network is completed.The initial generation value is processed according to the semi-global matching cost aggregation method;the parallax is calculated by the “winner-take-all” rule to obtain the initial disparity value;the disparity map is optimized by the disparity post-processing technique to make up for the defects of the traditional convolutional neural network,and the defect is obtained.A smoother parallax effect.This paper improves the pyramid pooling algorithm with reference to the classical convolutional neural network matching algorithm and advanced image processing technology.It is proved by experiments that the improved pooling algorithm can effectively reduce the initial matching error rate.Design an efficient stereo matching network to improve network matching accuracy and convergence speed by superimposing additional convolutional layers.The final experimental results show that compared with the benchmark algorithm,the convergence time of the proposed algorithm is shortened by about 50.1%,and the matching error rate is reduced from 6.65% to 4.78%.