Research On Three-dimensional Real-time Reconstruction Algorithms Of Binocular Stereovision Based On CNN

Machine vision is a hot issue in the field of biomedical engineering.Binocular stereovision technology,which simulates the principle of human binocular perception,has become an important research branch of machine vision,which has important application value in the fields of medical surgical robots and unmanned driving and so on.In binocular stereo vision,the traditional dense stereo matching algorithm generates massive data in the process of calculation,which is slow and difficult to be applied to real-time three-dimensional reconstruction.Sparse stereo matching algorithm has fast calculation speed,but it has the poor matching effect and is sensitive to noise.In order to overcome the above problems to a certain extent,the convolutional neural network technology is introduced into stereo matching algorithm to improve the accuracy and speed in the research,realizing low-cost three-dimensional real-time reconstruction.Firstly,the chessboard calibration method of Zhang Zhengyou is adopted to calibrate the binocular camera.Binocular camera with USB interface which can collect two images at the same time is the hardware foundation of this research,but the optical distortion caused by the manufacturing process will seriously affect the matching results.For this reason,after analyzing the mathematical model of camera imaging and common camera calibration algorithm,the binocular camera used in this study is calibrated by Zhang's calibration method based on MATLAB toolbox.The intrinsic and extrinsic parameters of binocular camera are obtained by feature point extraction and iteration calculation,which lays a good foundation for subsequent research.Then,aiming at the problem of low accuracy and time-consuming of traditional matching algorithm,a matching cost algorithm based on convolutional neural network is designed.A good software platform can accelerate the design and verification of algorithms.Therefore,the OpenCV3.3.0 machine vision development library under the development environment of vs2013 of win7 system is configured,and the convolutional neural network framework Caffe under Ubuntu 16.04 system is built.Using image pairs with real disparity in KITTI data set,a large number of small image pairs positive and negative sample data sets are constructed,the siamese convolutional neural network structure is designed and the parameters of each layer are set and the network model is trained to calculate the matching cost of image pairs.Finally,the semi-global matching algorithm is used for cost aggregation,Winner-Take-All strategy is used to find the optimal corresponding points,and left-right consistency checking algorithm is used to solve the problem of missing matching points caused by object occlusion.The disparity map is further optimized by smoothing filter to obtain the final disparity map.High-precision disparity map combined with triangulation principle obtains 3D point cloud data of human skull model and brain anatomical model and carries out three-dimensional reconstruction.The experimental results show that the stereo matching algorithm based on siamese convolution neural network model is effective and feasible to improve the matching accuracy and speed,and preliminarily realizes the real-time three-dimensional reconstruction of binocular stereo vision.