Research Of Polar Scene 3D Reconstruction From Ice Type Constraint And Aerial Optical Images

Aerial photography is an important way to obtain the polar information.But usually the aerial optical images cannot embody the terrain features intuitively.In the field of computer vision,multi-view reconstruction based on optical images is the trend of current research.This technology enables the conversion of image’s 2D planar information into intuitive 3D stereo information.It has been widely used in telemedicine,virtual reality,etc.It also brings possibility of achieving 3D reconstruction of sea ice terrain by using aerial optical images.This thesis focuses on the procedures and related improving strategies of 3D reconstruction of sea ice scenes.Based on the analysis of the specific characteristics of sea ice images together with the theory of structure from motion and multi-view stereo,the goal of 3D reconstruction of sea ice scenes that meets the needs of practical applications has been achieved.This thesis of research mainly includes the following parts:1.With the characteristic analysis of sea ice images,this thesis studies the factors that affect the correctness during the image selection period and the deficiency of the traditional classification standards of sea ice during the reconstruction procedures.On that basis,a new classification standard aiming at 3D reconstruction is put forward and a pre-processing procedure based on the neural network classification model is designed to automatically screen the images needed.The results show that new classification standard contributes to a better screening outcome.It even displays a better performance in the thick ice and thin ice scene screening which is hard to distinguish in manual operation.2.This thesis introduces the principle and characteristics of the sparse reconstruction process as well as different methods for different sea ice scenes under the new classification standard.It also puts forward correspondent reconstruction strategies regarding different scene types.It is justified in this thesis that sea ice scenes with relatively complex terrains from our data is more suitable for the incremental structure from motion method.For the ones with relatively flat terrains,the adaptability of illumination model is demonstrated.Thus,it proposes an improved strategy that recovers shapes from shade with sparse point cloud data reconstructed from complex sea ice scenes as prior constraints.The results show that the new reconstruction method based on types works well and archives better results.3.To get a complete reconstruction procedure,based on the introduction of multi-view stereo theory,this thesis analyzes its influential factors.It studies the inter-regional relationship of sea ice images in different scenes and has proposed a dense reconstruction strategy of sea ice scenes based on region constraints.In the region segmentation phase,the sparse point cloud generated by the original view is used to project and segment the candidate reconstruction area on each corresponding image to obtain segmented reconstruction regions.In the reconstruction phase,the segmented images are used as the dense matching input,and the sparse point clouds are used as reliable initial parameters.Then we combine Cluster Multi View Stereo and Patch-Based Multi-View Stereo to achieve the complete reconstruction procedure of sea ice scenes.Relevant results based on the data set show that our strategy can effectively reduce redundant computation and improve operating speed of the dense reconstruction stage.It also reduces the outliers in the dense point cloud data.