Study On Binocular Stereo Vision Matching Algorithms Based On Convolutional Neural Network

Binocular stereo vision is an important research topic in the field of computer vision.It can reconstruct 3D scene through multiple 2D images captured in a scene from different perspectives.It has a wide range of application prospects in the fields of driverless,virtual reality and robot.Stereo matching is the most important and difficult part of binocular stereo technology.The goal of stereo matching is to calculate the disparity of each pixel in the reference image and get the disparity image to calculate the depth by giving two images captured by the stereo camera.Traditional methods generally follow four steps: matching cost calculation,cost aggregation,disparity calculation and disparity optimization.However,this manual method of feature descriptors is often plagued by low efficiency.In recent years,due to the strong feature extraction ability of convolutional neural network,it once replaced the traditional manual method of feature descriptor.The method based on neural network can regard the disparity estimation as a learning task,use a large number of data to optimize the model parameters continuously for learning,and use the neural network model after learning to process the input image output disparity map.Compared with the traditional stereo matching method,the convolutional neural network method can capture the local context information better,and it is more robust to the performance of the ill posed areas(occlusion,weak texture and depth discontinuity,etc.),and has a good improvement in matching speed and accuracy.In this paper,we have a deep understanding and Research on the domestic and foreign research status and key and difficult problems of binocular stereo vision matching algorithm.Aiming at some problems caused by the traditional four-step methords,it is easy to cause error accumulation and complicated calculation process.In this paper,we design an end-to-end disparity prediction framework to output a good disparity map,which does not need any post-processing steps.Based on the previous work of convolution neural network stereo matching algorithm,this paper improves the network structure.In order to solve the problem that the feature extraction ability of the stereo matching algorithm based on convolution neural network is insufficient in the ill posed area,this paper makes full use of the feature information of different levels of the extracted image,designs a pyramid layer in the front part of two branches of twin network structure,extracts multi-scale low-level structure features,and merges the high-level semantic features of the following three layers in the back-end of the network.It enhances the ability of image feature representation and extraction of convolutional neural network structure.After extracting the left and right image features,we measure the similarity and output the disparity map.The effectiveness of the proposed algorithm is verified on the KITTI 2015 dataset,and compared with other similar algorithms,the accuracy of stereo matching is improved.