Research On 3d Dynamic Scene Reconstruction Method Based On Multi-view Video

The three-dimensional reconstruction technology is a technology for obtaining 3D information of an object from a real physical world through cameras.In recent years,with the development of technology and increasing demand for 3D data,reconstruction technology has attracted attention and made great progress.However,3D reconstruction still faces some problems.For example,in the reconstruction of static scenes,there are some mismatches in feature matching,and the point cloud holes caused by the unclear features of weak texture or repeated texture regions.In addition,for dynamic scenes,there are some problems of inaccurate target detection and tracking,and immature dynamic reconstruction algorithms.In this paper,for the matching problem of multi-view images reconstruction,structural points optimization is added to feature matching process.The semantic information provided by the structural points and the geometric information of the objects are used to fill the holes of the reconstructed point cloud,reducing the influence of outliers.For the reconstruction of multi-view video dynamic scenes,an optical flow algorithm is added to optimize the process of 3D scene point cloud reconstruction.The optical flow algorithm is used to track and extract moving objects,and the multi-frame scene background point cloud is merged,which effectively improves the scene information.Considering the limitations of traditional optical flow algorithms,the optical flow based on deep learning is used to improve the effect of scene reconstruction.The main work of this article is as follows:(1)The object structure points are used to optimize the feature matching process.Aiming at the problem that the features in the weak texture region are difficult to match,the object structure points are extracted,and the matching points are added to alleviate the problem of insufficient number of feature matching in the weak texture region.(2)A point cloud reconstruction method based on structural point optimization is proposed.Aiming at the problem of outliers and holes in the reconstructed point cloud,a method based on structural points optimization is proposed.The target structure points extracted by the convolutional network are used to remove outliers in the reconstructed point cloud according to the associated information in the three-dimensional space,and to fill the holes existing in the area where the target object is located.(3)A three-dimensional dynamic reconstruction method based on optical flow is proposed.Use optical flow to track scene targets in multi-viewpoint videos,extract three-dimensional motion points,and merge scene background point clouds.In the original three-dimensional static scene,the point cloud information at multiple moments is merged,which increases the effective information amount of the scene.