Research Of SFM 3D Reconstruction Technology Based On Deep Learning

The 3D reconstruction based on vision is an important research topic in the field of computer vision.It captures the image of the surrounding scene through the camera and uses the 3D reconstruction algorithm to reproduce the 3D model of the scene.Vision-based 3D reconstruction technology has been widely used in robot navigation,autonomous driving,reverse engineering,heritage protection and other fields.Therefore,the study of 3D reconstruction technology based on vision has important theoretical and practical value.In this paper,SFM 3D reconstruction method based on monocular vision is adopted to reconstruct the surrounding scene through images taken from different angles.This paper focuses on the key steps of SFM 3D reconstruction based on structure from motion,it mainly includes camera model and camera calibration method,stereo matching algorithm,depth prediction algorithm and Camera pose estimation.The details are as follows:(1)The hole camera model and camera distortion model are discussed in detail.On this basis,the calibration principle and process of several camera calibration algorithms are elaborated and their advantages and disadvantages are compared.Camera calibration experiments are carried out according to the theory.(2)This paper explores the application of deep learning method in 3D reconstruction,and proposes a convolution network model "single-peak matching network" for stereo matching.The input of "single-peak matching network" is two image blocks to be matched,the network structure is designed in the form of full convolution,and the model training adopts the end-to-end training method.In this paper,the pixel resolution of the proposed network model is tested experimentally and compared with other feature matching methods.Experimental results show that the model has a good matching effect when matching image block pairs with a distance of more than 8 pixels,and the proposed model has a higher matching accuracy compared with similar matching methods.(3)This paper studies the prediction of image depth by deep learning method,proposes a convolutional network model for image depth prediction,and conducts experimental verification on the model.The experiment shows that the depth trend of the predicted depth map is basically consistent with that of the real depth map,and there is a certain gap between the predicted depth value and the real depth value,with the root-mean-square error of 1.183.In addition,other potential applications of deep learning in three dimensional reconstruction are discussed.(4)On the basis of the above research,the paper has written an 3D reconstruction program based on structure from motion.With the help of SIFT for feature point detection,feature matching convolution network model for feature matching,triangulation algorithm for point cloud coordinate calculation,PnP algorithm for camera pose calculation,incremental 3D reconstruction is finally achieved.