Research And Implementation Of Disparity Estimation Method Based On Binocular Images

With the development of artificial intelligence,people more and more demand for smart devices.The autonomous mobility of driverless vehicles and UAV,the accurate grasping ability of industrial robots all depend on non-contact measurement technology.Because of high precision and low cost,binocular distance measurement has become one of the most promising methods in the field of non-contact measurement.This thesis introduce convolution neural network method into the research field of binocular distance measurement,and also innovate and optimize the algorithm's accuracy,real time and robustness.The main content of this thesis is as follows:1.After fully analyzing the limitations of the traditional stereo matching method,according to the feature extraction ability and expressive ability of convolutional neural network,this thesis replaced three modules of matching cost calculation,matching cost construction and disparity prediction in traditional method by deep learning method.On this basis,a complete frame using convolution neural network to predict disparity is formed,and a tiny end to end disparity prediction network is designed.The experimental results show that it is feasible to predict the disparity with the convolution neural network.2.This thesis proposed a deep disparity prediction network based on the structure of coding and decoding,and the network is designed for high precision and high real time capability scene.The whole network uses coding and decoding structure as general framework.In the step of geometric feature extractio,a shift-concat binocular feature fusion method is proposed and the separable convolution is used to generate the matching cost.This method obtains a more accurate result of disparity prediction and has a certain generalization ability on different data sets.3.This thesis proposes a high robustness disparity prediction network based on low resolution images aiming at the low performance processors and binocular cameras equipped on portable mobile devices.First,in order to deal with low resolution input images,the dilated convolution is used in the image feature extraction module.Second,the super resolution function is added to optimize the output result in the disparity prediction module.Finally,in order to improve the robustness of the system to the dithering of the binocular camera,a two-dimensional geometric feature extraction method is used in the network.The experimental results show that the network can predict the exact disparity with low resolution and inaccurate input images,and has certain practical application value.All the networks in this thesis are trained by FlyingThings3 D dataset,and verified the validity of the trained network on the FlyingThings3 D data set and Middlebury data sets.