Research On 3D Scene Rendering And Reconstruction Technologies Oriented To Spatial Multi-view Images

Three-dimensional scene rendering and reconstruction is a structure model by using graphic or image information to render the scene image in a particular observation viewpoint and reconstruct three-dimensional scene.It is an important research problem in computer vision and has a very important significance for pattern recognition,virtual reality,adventure and rescue,military reconnaissance.According to different types of basic processing units,the classical three-dimensional scene rendering and reconstruction methods are divided into two categories:one of models employs with pixels as the basic processing units to render and reconstruct.However,the performance of this mehtod is relatively slow.The other regards grid as the basic processing units to render and reconstruct images,which is calculated fastly and can basically meet the requirements of real-time rendering,But when the grids contain the target boundary it will lead to distortion of the rendered image while rebuilding the model.Furthermore,Many existing rendering and reconstruction algorithms require that the coordinates of different observational viewpoints which may not be able to be accurately gotten in real situations.For instance,it's difficult to determine the coordinates of the underwater robot due to water shocks.This paper takes granules of the image(ie connected regions of irregular shape)as the basic processing units and three-dimensional scene rendering and reconstruction method for spacial multi-viewpoint images is researched.The main contents are as follows:(1)Image granularity based on the theory of morphological connectivityConsidering the features of hue?brightness and texture,a granulation method is proposed by using morphological connected operator of color image.The pixels in the same region have homogeneous or same texture structure,while pixels in different granules have different characteristics.Then,the method of determining the granule boundary is discussed based on Hidden Markov model which can achieve the consistent segmentations for color images in different viewpoints.(2)Image registration based on spatial topological structureThe spatial topological structure between multi-views images of granules(connected region)is established based on directed graph.The features of directed trees and loops(ie generalized vertex)are analyzed by introducing reachable matrix.By fuzzying granules of images,the global optimum matching problem between different viewpoint images is discussed based on fuzzy reasoning.On the basis of granule matching between images,the matching relationship of feature points among granules is established.This method eliminates the false matching problem of self-similar structures in different viewpoints images.(3)Three-dimensional scene rendering and reconstruction with unknown relationship between multi-view camera positions.Aiming at the impact of radial distortion of camera,a calculating method of the relative position between cameras in different viewpoints is proposed based on weighted Sampson approximation.A method of scene depth estimation is put forward based on the view tension drift,and a technique of three-dimensional scene rendering is developed combining image granule based spatial transform with scene depth map.Moreover,three-dimensional reconstruction technique is studied which combines the model surface fitting and point cloud model correction.The proposed method reduces the impact of cumulative errors under the premise of unknown relative positional relationship of multi-views,truly achieving three-dimensional rendering and scene reconstruction.