3D Reconstruction Of Tree Structure Based On RGB-D

The RGB-D-based 3D reconstruction is the most extensive research direction in recent years.The main advantage of the RGB-D camera is that it is convenient to acquire depth data and can acquire color images at the same time.However,due to the accuracy problem of the camera,the deep noise of the collected data is serious,especially for the tree-like structures,the data loss occurs in the finer parts of the branches,which seriously affects the reconstruction effect in the later period.In the registration phase,the color of the tree structure is monotonous,the shape of each branch is similar,and the extraction of feature points is difficult,resulting in low registration accuracy.This paper proposes to use the complete information of the color image to optimize the missing depth data,and proposes to use the target boundary point as the feature point for coarse point cloud registration.The main research contents of this paper include data acquisition and processing,optimization of depth data based on color images,and feature matching based on boundary points.Because there are few researches on the three-dimensional reconstruction of the tree-like structure and the data set is not perfect,at the same time,the research needs to use the color image and the depth data at the same time,and it is necessary to perform the data collection of the tree-like structure object.Use the Kinect SDK of the Kinect depth camera to acquire data,and use point cloud development libraries to convert depth data into point cloud data files.For the noise and lack of data in the tree structure data,this paper proposes to use color images to optimize the depth data.By calculating the coordinate mapping of the color image and depth data,the pixel coordinates of the color image are mapped to the coordinates of the midpoint of the point cloud space.To achieve the purpose of optimizing missing depth data.For the registration of point clouds in tree-structured objects,this paper considers that boundary contours can also represent the shape features of objects,and proposes the use of boundary points as feature points for coarse registration of point clouds.