Research On 3D Point Cloud Reconstruction Of Single Tree Based On Multi-angle Point Cloud Data

With the continuous research and development of machine vision and computer graphics,3D reconstruction technology is widely used in many fields,and its importance is becoming more and more prominent.Therefore,this paper uses Kinect V2 sensor to obtain multi-angle single tree point cloud data,and carries out 3D point cloud registration,3D point cloud model reconstruction and branch and leaf separation research.The specific research contents are as follows:(1)The capture and pretreatment of point cloud is the first step in the reconstruction of three-dimensional point cloud model of single tree.In this paper,the depth data obtained by Kinect V2 sensor is firstly converted into three-dimensional point cloud data and the point cloud data is preprocessed.The preprocessing of point Cloud includes using Cloud Compare software to segment 3d point Cloud data.Then,the point cloud statistical filtering algorithm is used to filter the single tree,and the 3d point cloud data of statistical filtering is compared before and after,and the discrete points of the single tree after filtering are effectively reduced compared with the pre-filtering point cloud data.(2)Point cloud registration is the second step of 3d point cloud model reconstruction.This paper studies the influence of different registration methods and acquisition methods on the registration accuracy.Firstly,3D scale-invariant Feature Transform(3D scale-invariant Feature Transform,3DSIFT Feature point extraction algorithm,3DHarris Feature point extraction algorithm and Normal Aligned Radial Feature(NARF)Feature point extraction algorithm were used for rough registration.Experimental results show that NARF Feature point is the worst.Secondly,the root mean square error of the corresponding point pairs after using 3DSIFT feature points and using 3DHarris feature point algorithm is 1.44 cm and 1.70 cm respectively,which shows that 3DSIFT feature points have the best registration effect.Then,the Iterative Closest Point(ICP)algorithm was used to study the precise registration.In addition,comparative experiments with different acquisition distances,heights and perspective differences were also discussed,and the data acquisition method suitable for this experiment was obtained: the distance from the camera to the center of the tree was 2m,the height of the camera was 10 cm,and the interval of each scan was 45°.In this acquisition method,eight angles of the tree were taken.The experimental results show that the root mean square error and time of accurate registration using3 DSIFT feature points and 3DHarris feature points combined with ICP algorithm are 1.00 cm and1.01 cm,7.90 s and 4.17 s,respectively.(3)Aiming at the reconstruction of multi-angle point cloud model,this paper designs a multi-angle stitching method by using the adjacent angles mentioned above and the registration results of 3DSIFT feature points and 3DHarris feature point algorithm,and carries out 3D model reconstruction of multi-angle single tree point cloud data.In this paper,point cloud data from 8angles are used.After registration of adjacent point clouds,multi-angle point cloud stitching method is used for stitching,and finally a complete THREE-DIMENSIONAL point cloud model is obtained.Secondly,the Voxel Cloud Connectivity Segmentation(VCCS)algorithm and the Locally Convex Connected Patches algorithm are adopted.LCCP algorithm was used to separate branches and leaves from the 3d point cloud model,and good results were obtained.