Implementation Of 3D Reconstruction Algorithm Based On Multi-view Image

With the development of computer vision and the improvement of computing power,3D reconstruction technology based on multi-view images has been rapidly developed.This technology is low in cost and easy to obtain texture information of scenes.It is suitable for large-scale scenes,and 3D reconstruction has a wide range of applications in driverless car,digital architecture,digital cities,games and movies.Reconstructing a scene 3D model from multi-view image includes three stages of image feature extraction and matching,sparse reconstruction,and dense reconstruction.In large scenes,image feature extraction and matching take more time.In the subsequent sparse reconstruction stage,the traditional incremental SfM algorithm calculates the camera parameters and scene structure in turn.The algorithm has high precision,but the calculation takes a long time.The global SfM algorithm calculates all camera parameters and scene structure according to the camera constraint relationship.The algorithm is fast,but has poor robustness and low precision.Aiming at the problems existing in traditional SfM,a hybrid SfM algorithm is proposed.The main work includes:1.Aiming at the above problems,in feature extraction stage,the CUDA implementation of SIFT feature extraction is studied to accelerate the feature extraction.In the matching stage,the data structure of the vocabulary tree is studied to speed up the feature matching.2.In sparse reconstruction stage,a hybrid SfM is proposed to segment a large-scale image set into small subsets.First,the camera parameters of each subset are calculated by the incremental SfM,then all camera parameters are calculated by the global SfM.The algorithm avoids the problem of incremental SfM time-consuming and error accumulation,also avoids the problems of low accuracy and poor robustness of global SfM,balances the accuracy and efficiency of reconstruction.3.In dense reconstruction stage,the depth map fusion MVS algorithm is used to densify the sparse model obtained by the hybrid SfM to obtain a dense model of the scene,which makes the effect of the three-dimensional model more realistic.Tested on the Tanks and Temple benchmark dataset,the results show that SIFT's CUDA implementation saves more than 70% time compared to the CPU.The vocabulary tree matching strategy saves more than 50% time compared to the exhaustive strategy,and the matching results are almost same.The hybrid SfM saves more than 40% time compared to the incremental SfM with a comparable number of point clouds.Compared with the global SfM,the hybrid SfM has an error reduced by about 2% when the number of point clouds is about twice.Finally,the model obtained by the hybrid SfM is densified by the MVS algorithm to obtain a dense model,and the real scene is more accurately restored.In addition,this thesis collected the actual scene for testing,and the results show that the contour of the reconstructed model is clear and meets the research work requirements.