Study On 3D Reconstruction Method Of Trees Based On Sequential Images And Accuracy Assessment

Three-dimensional reconstruction of the real world is not only a rewarding job, but also a hot topic of the current research. As an important part of the real world, trees’ 3D reconstruction is very important. Three-dimensional modeling methods are mainly divided into three categories: geometry-based modeling, range-based modeling and image-based modeling. Among them, the image-based modeling method is rapidly becoming an important method for three-dimensional reconstruction because it is fast, convenient and low-cost. How to generate three-dimensional model from twodimensional images is the key problem of image-based reconstruction. Meanwhile, the precision of the 3D point cloud reconstructed from the images remains to be studied.This thesis presents an image-based 3D reconstruction solution of trees and evaluates 3D point cloud generated from it. First, this thesis introduces the basic concepts of homogeneous coordinate, projective geometry and camera imaging model, laying the groundwork for subsequent research methods. Then, the SIFT descriptor is used to extract feature points from the images and match them using the nearest neighbor method. After the matching process, these matched points can be used to calibrate the cameras automatically and solve the fundamental matrix and essential matrix, and then the camera projection matrix can be solved. In this process, we use RANSAC method to enhance the robustness of the calculation and the Bundle Adjustment method to optimize the parameters of the solution. When we are actually reconstructing a tree, sparse reconstruction can barely meet our demands because the sparse point cloud cannot reflect the whole picture of the object, hence we need to densify the sparse point cloud. We use the PMVS method to do this and its first step is to select the seed patches and reconstruct their three-dimensional information. Then the seed patches will be expanded to their neighboring cells and the patches in them can be reconstructed. This process can be repeated until all the patches meeting the requirement can be completely reconstructed. Finally, the geometric consistency constraints and the illumination consistency constraints are used to filter the reconstructed patches which do not meet the requirements. By designing a 3D reconstruction system,the thesis realizes the 3D reconstruction methods.The reconstruction results herein were evaluated. The Leica ScanStation C10 laser scanner was used to acquire 3D point clouds of trees of three different sizes. These point clouds were used as references of the point clouds generated by the image-based reconstruction method. The comparison results are: For the Hibiscus of the minimum height, the average distance between two models is 4.2mm and the standard deviation is 10.3mm; for the Narra Tree of medium height, the average distance is 2.2mm and the standard deviation is 5.6mm; for the highest ginkgo tree, the average distance is 0.7mm and the standard deviation is 22.6mm. Taking the 5% and 10% of DBH as the standard to filter outliers, the corresponding ratio of remaining points of the three trees are as follows: 66.8% and 87.6%; 69.2% and 96.2%; 76.3% and 94.2%. The remaining points of the trees are relatively intact, indicating that the reconstruction accuracy is good.