Study On 3D Reconstruction Technology Of Low-Texture Target Based On Polarization And Multi-View Stereo

Passive 3D reconstruction technology based on multi-view stereo has emerged as a prominent research area within the field of computer vision,proving to be widely applicable in various practical domains,including terrain detection,digitized cultural relics,unmanned vehicle automation and virtual reality.However,the strong dependence on image features poses a major challenge when dealing with low-texture targets,leading to suboptimal reconstruction outcomes that may limit the applicability of multi-view 3D reconstruction technology in certain application scenarios.Fortunately,studies have shown that polarization information has been found to be capable of reflecting the geometry information of an object surface to some extent,while being relatively insensitive to textureless information.Hence,polarization can provide useful constraints to facilitate more accurate 3D information estimation in featureless areas,thereby enabling more comprehensive 3D reconstruction of the low-texture targets.Building on this foundation,the present thesis thus undertakes a deep exploration into 3D reconstruction technology specifically targeting such low-texture objects..Firstly,this thesis delved into multi-view 3D reconstruction techniques based on the camera imaging model and the theory of multi-view 3D imaging,which included methods for estimating camera pose information via motion recovery structure,as well as for performing dense 3D point cloud reconstruction based on depth map estimation and fusion.The initial depth map of the target image was generated through the multi-view3D reconstruction technique.In addition,this work analyzed the ambiguity problem of polarization information through the polarization reflection model,and outlined the development of a novel neural network model that leveraged polarization information for normal map estimation.The proposed model advanced solutions to the aforementioned ambiguity and perspective projection problems,incorporating multiple self-attention modules and perspective coding and so on.Ultimately,the normal map of the target image was estimated leveraging this neural network model.Building on the studies mentioned above,the present thesis established a framework for 3D reconstruction of low-texture targets based on polarization multi-view technology.This framework utilized the normal map constraint to propagate and interpolate depth information from feature-rich to featureless regions,while optimizing the depth map results through optimization theory.Ultimately,the 3D point cloud model of the target was reconstructed based on a fusion of the resulting depth maps.Finally,an experimental framework of 3D reconstruction based on polarization and multi-view was established,and experiments were conducted on low-texture targets.By comparing the results of the reconstruction method proposed in the thesis and the traditional multi-view 3D reconstruction method,the study was able to demonstrate that the 3D reconstruction effect of the method proposed in the thesis on low-texture objects is improved compared with the traditional multi-view 3D reconstruction method.At the error threshold of 0.005 m,the F₁ score of the reconstruction method proposed in the thesis was 16.45%and 1.91%higher than that of colmap and open MVS respectively in the reconstruction of paper bag.As for the kettle with lower texture information,the F₁score of the reconstruction method proposed in the thesis was 32.71%and 20.59%higher than those of colmap and open MVS,respectively.