Research On 3D Reconstruction Based On Binocular Stereo Vision And Deep Learning

Binocular stereoscopic vision obtains the three-dimensional depth information of the object by simulating the way human eyes perceive the target object.Binocular stereoscopic vision technology has become the most concerned direction in the field of machine vision,and has been widely used in the fields of nondestructive flaw detection and unmanned aerial vehicle.Stereo matching is an important part of binocular vision technology.The complexity of the traditional dense stereo matching algorithm leads to slow calculation speed,which makes it difficult to obtain the 3d information of the observed target in real time.Although the sparse stereo matching algorithm is simple and fast in calculation,its calculation accuracy is poor,and it is easily affected by noise.In order to solve the above problems,this paper applies convolutional neural network models to stereo matching technology to achieve highprecision and rapid 3d reconstruction of the target object.Aiming at the issue of poor accuracy of binocular cameras calibration,this paper firstly makes a comprehensive comparison and analysis on the existing camera calibration technologies,and Zhang's method with high accuracy is chosen to extract the feature points of image sequences.Then by combining Matlab software and Opencv4.0 machine vision development library,the parameters such as the internal reference matrix,the distortion coefficient matrix and the translational rotation vector of binocular cameras are computed for the subsequent stereo matching and 3D reconstruction technology.In order to solve the problems of low accuracy and poor real-time performance of the traditional stereo matching computation,in the stage of the initial matching cost,this paper designs an efficient and lightweight convolutional neural network,which stacks multiple convolutional layers to extract feature maps of different scale for performing the more accurate estimation of initial matching costs.Standard image sequences from the KITTI database are adopted to iteratively update the constructed network model parameters,which effectively achieves the disparity estimation precision.Compared with the listed comparison methods,the estimation accuracy of the proposed algorithm is improved by 28% and the running time is less than 1 second.For the issue of low accuracy of initial disparity estimation,this paper uses the crossbased semi-global matching algorithm to perform cost aggregation for extracting the optimal matching points.Considering the lack of matching points caused by occlusions,this paper further adopts left-right consistency algorithm to refine the cost volume for better disparities.Additionally,for the phenomenon of mismatch generated by noises,this paper performs smooth filtering operation on the extracted disparity maps for improving the protection performance of the image edge.Compared with the listed comparison methods,the estimation accuracy of the proposed algorithm is improved by 35% and the running time reaches the second level.Finally,based on the principle of triangulation,this paper uses the high accuracy and strong robustness of the computed disparity maps to perform the 3D reconstruction of the target objects in real scenarios.This paper utilizes multiple standard databases such as KITTI and Middlebury to conduct a comprehensive experimental comparison between the proposed method and other classic algorithms.The experimental results illustrated that the convolutional neural network introduced into the stereo matching calculation technology may effectively improve the accuracy of the matching algorithm and satisfy the real-time requirements of the binocular stereo vision technology in real scenes.