Research On Binocular Stereo Matching Calculation Technology Based On Densely Connected Convolutional Neural Networks

Binocular stereo vision is an important task in the fields of computer vision and image processing,and aims to recover the scene depth by calculating the disparity of the pixel points in the left and right viewpoint images.It contains rich geometric structure information in the image.Thus,the key of studying on binocular stereo vision is to extract accurate and dense disparity maps from rectified stereo image pairs,and then provides reliable prior information for more advanced vision tasks.Since the 21 st century,with the rapid development of computer hardware and software,binocular stereo matching calculation technology has been widely applied in many fields,such as augmented reality,autonomous driving,UAV navigation control system,3D reconstruction,and target recognition and tracking.In recent years,with the continuous development of deep learning theory and technology,the accuracy and robustness of stereo matching estimation based on deep convolutional neural networks have been significantly improved.However,when the binocular images include complex scenes,such as texture-less,illumination changes and occlusions,the accuracy and robustness of the existing stereo matching algorithms need to be further enhanced.To solve the above issues,this paper mainly focuses on binocular stereo matching calculation method based on densely connected convolutional neural networks,which aims to deal with the problems of matching ambiguity and disparity map detail loss of stereo matching calculation under complex scenes.The main research contents of this paper are as following:1.The research significance and background of binocular stereo matching calculation technology are firstly introduced.Then,the research status and some remained key problems of binocular stereo matching calculation technology are summarized.Finally,the main contributions and chapter arrangement of this paper are briefly described.2.The basic knowledge of binocular stereo matching calculation technology is elaborated.Specifically,the binocular stereo vision theory,the binocular stereo vision system,the used stereo constraints and the existing typical stereo matching models are discussed and analyzed in detail.3.To improve the accuracy and robustness of disparity estimation under complex scenes,such as texture-less,illumination changes and occlusions,a stereo matching algorithm based on densely connected convolutional neural networks is proposed in this paper.Firstly,the advantages of densely connected convolutional neural networks are discussed.Then,in the process of the initial matching cost calculation,the dense fully convolution module with feature reuse is constructed,and the captured feature maps from the shallow layers are cascaded to subsequent layers through the skip connection mechanism to compensate for the loss of local feature information in the deep convolutional layers.Finally,combined with the post-processing pipeline,the cost function of network prediction is further modified and optimized to obtain the optimal disparity,which can effectively improve the accuracy and robustness of disparity estimation under complex scenes.4.To optimize the parameter learning of the initial matching cost network model,a hybrid cross entropy loss strategy with a certain slack is proposed in this paper.The network parameter learning is divided into two stages,and the loss functions with different value of expectation are designed to iteratively update the network parameters for better distinguishing matching and nonmatching training samples and improving the performance of the network model.5.The standard test image sequences provided by Middlebury and KITTI databases were adopted to compare and analyze the proposed method and the existing representative stereo matching methods.The experimental results show that compared with other comparison methods,the proposed method has high accuracy of disparity estimation,especially in complex scenes,it performs better accuracy and robustness of disparity estimation,and can effectively recover disparity map details.