High Resolution 3d Scene Reconstruction Based On Dense Optical Flow

In recent years,naked eye 3D display technology has developed rapidly,which constantly gives the audience a shocking visual impact,and carry people’s expectations for the future display technology.However,due to the particularity of naked eye 3D display equipment,its display content needs to be specially made.In order to allow viewers to see images of objects at different angles on the display device,it is necessary to model the objects with high accuracy and collect images from different viewpoints.We usually use CAD,blender,unity and other graphics software to model and render objects,but the work takes a long time and it is difficult to build realistic scenes.The traditional 3D reconstruction method based on multi view can generate point clouds from a series of scene images to calculate the 3D structure of the scene.However,the method is limited by the number of features,resulting in the lack of details of the model,which can not be applied to 3D display.In this thesis,a high-resolution 3D scene reconstruction method based on dense optical flow is proposes and studies.The work mainly consists of two parts:First,a global attitude optimization method based on feature and optical flow matching is proposed to achieve camera calibration with high accuracy and robustness,which provides a basis for high-precision dense optical flow construction;Secondly,a high-resolution 3D scene construction method based on dense optical flow is proposed.Dense optical flow is used to match and generate high-quality 3D models for 3D display.The research contents are as follows:1.The camera pose solution is studied,and a pose optimization algorithm based on image pyramid features and optical flow is proposed,so that the pose can maintain robustness in multi-scale image and improve accuracy in small displacement areas.2.The dense point cloud reconstruction method is studied,which uses neural network to generate dense optical flow instead of traditional feature matching,makes full use of all pixels in the image,effectively expands the scale of point cloud,and provides more details for model construction.3.Through the study of 3D model construction,stereo vision principle,3D display and image coding technology,a high-resolution 3D model generation system is designed to provide content for 3D display.