Research On Multi-view 3D Reconstruction Technology Based On SfM

3D reconstruction technology has been widely used in driverless,construction industry,game movie modeling and other fields in recent years.With the improvement of computer computing ability and the continuous development of computer vision,the multi-view 3D reconstruction technology based on Sf M has achieved rapid development.Compared with active methods such as structured light,this method has the advantages of low cost and strong applicability.A number of images from different angles are used as input to reconstruct the 3D model of the target object.The process includes three stages: extraction and matching of image feature points,sparse reconstruction order and dense reconstruction.When the number of images in the dataset is large or there are many feature points,the operation time of feature point extraction and matching will be too long.The traditional incremental structure from motion(Sf M)adds images one by one and solves its camera parameters and point cloud model,which has high reconstruction accuracy but long reconstruction time.The traditional integrated Sf M solves all camera parameters and point cloud models at one time,The running time is short,but the accuracy and robustness are low.In order to solve the shortcomings of traditional Sf M,this paper designs a hybrid algorithm.Specific work contents: 1.In the feature extraction part,in order to shorten the time spent on extracting feature points,the CUDA(Compute Unified Device Architecture)based accelerated feature extraction is implemented.In the feature matching part,in order to reduce the time spent on feature matching,the feature matching algorithm based on vocabulary tree is studied;2.In the sparse reconstruction part,the hybrid Sf M algorithm is designed and implemented.The picture set is divided into multiple subsets,and the camera parameters of each subset are solved in the subset.The process uses the additive Sf M,and then solves the overall camera parameters of all cameras.The process uses the integral Sf M,and then carries out the sparse reconstruction,which solves the problems of the additive time-consuming and low overall accuracy and robustness;3.In the dense reconstruction part,input the sparse model and camera parameters,and perform dense reconstruction through the Multi-View Stereo(MVS)algorithm based on depth map fusion,finally obtain the dense model of the reconstructed object,making the 3D model realistic.This paper conducts experimental verification on seven sets of data sets.The analysis of the results shows that using GPU to complete the CUDA-based accelerated SIFT feature extraction method reduces more than 65% of the time compared with using CPU to extract SIFT features.The feature matching algorithm based on lexical tree overcomes the timeconsuming problem of exhaustive matching and reduces more than 50% of the running time.The number of point clouds obtained by the hybrid Sf M and the additive Sf M is basically the same,but the hybrid Sf M can reduce the reconstruction time by more than 40%compared with the additive model.The hybrid Sf M reduces the reconstruction error by about 2% compared with the integral Sf M,and its number of point clouds is about twice that of the integral model.The MVS algorithm based on depth map fusion is used to complete the dense reconstruction,and the final dense point cloud model is obtained.The3 D model has a clear structure and meets the research requirements.