Research On 3D Reconstruction Technology Based On Computer Vision

3D reconstruction is a very important research direction in the field of computer vision,being compared with 2D image information,there is a stronger sense of reality and can present more information for 3D model.With the development of virtual reality and augmented reality technology,it can be predicted that 3D data will be the main form of data presentation in the future.The cost of 3D reconstruction of disordered images is low and the scope of application is wide,what's more,the characteristic of 3D is that only computing resources are needed.However,the speed of reconstruction is slow,so improving the efficiency of the algorithm and the accuracy of the reconstruction results is the focus of current research.At present,there are more mature algorithms in the research of 3D reconstruction technology based on disordered images,but many problems are presenentand,for example,long running time,low reconstruction accuracy and so on.In this paper,the related algorithms of 3D reconstruction are studied,the running speed and model quality can be improved.The specific research contents and improved methods are as follows:1)A matching method of growth features is proposed that in order to improve the stability of the whole 3D reconstruction process.The matching results of SIFT and SURF are combined,what's more,the matching pairs after fusion are applied to the SFM technology,which can be predicted that the accuracy of camera calibration is effectively improved,and the number of sparse point clouds is increased.In a word,the quality of 3D reconstruction results is improved.2)Solving the problem that pairwise matching of disordered images is needed in the process of SFM,an improved image pair matching strategy is proposed.By constructing the global hash sign on the data set by hash technology,the invalid image matching can be effectively screened out.By reducing the number of matching,the computation is reduced,the running time of SFM technology is greatly reduced and the running speed of the whole algorithm is improved.3)There may be insufficient image matching information and a large amount of noise which may lead to errors in the depth information of the image in the process of dense point cloud reconstruction.Therefore,an improved depth map fusion algorithm is proposed.By setting the confidence,the 3D point with high precision can be retained,the redundancy can be deleted and the accuracy of dense point cloud can be effectively improved.In this paper,three sets of common data sets and three groups of self-collected data sets are selected,and the effectiveness of the improved algorithm is proved by a large number of experiments.