| In recent years,with the rapid development of computer vision technology,it has been widely used in the fields of non-destructive measurement of crops,quality classification and so on.The main research content of this paper is to reconstruct the three-dimensional model of potato using the sequence image obtained by the monocular camera.The three-dimensional model will provide data basis for potato shape classification and growth state estimation.This paper designs a complete potato image reconstruction system from four aspects:camera calibration,feature matching,camera pose estimation,sparse point cloud and dense point cloud reconstruction.In view of the technical difficulties in the process of potato reconstruction,the research is mainly carried out from the following aspects:(1)An improved SIFT feature matching algorithm is proposed to solve the problem of long time-consuming in feature matching stage.The dimension of SIFT descriptor is reduced by changing the statistical strategy of image gradient direction.When extracting descriptor,only the gradient values of each window in four directions in 16 neighborhood windows of feature points are counted,and the rough matching results are eliminated by RANSAC algorithm.The experimental results show that compared with the original SIFT+RANSAC algorithm,the proposed algorithm improves the speed of feature matching by nearly 25%while guaranteeing the accuracy of feature matching.(2)A pose estimation method for multi-view cameras based on SFM algorithm is designed.The improved SIFT algorithm is used for image feature matching.After eliminating mismatching by RANSAC algorithn,the relative pose of cameras in two views is solved by classical 8-point method,and the perspective N-point method(PnP)is introduced to realize the pose estimation of multi-view cameras on the basis of two views.In order to ensure the accuracy of pose estimation,global optimization is carried out for dimensional points and camera pose.(3)The number of images required for the three-dimensional reconstruction of the potato was determined.Using the camera to collect 36,60,90,120,180,360 images of 10 different shapes of potato for 3D reconstruction,by analyzing the accuracy of the three-dimensional model of potatoes,and considering the speed of generation of three-dimensional point clouds,the number of images in three-dimensional reconstruction of potatoes was 90.(4)Constructed a software system for 3D reconstruction of potato,the PMVS algorithm,Poisson reconstruction algorithm and texture mapping algorithm required in the process of generating dense point cloud are analyzed in detail.Combining these existing algorithms with the improved algorithm in this paper,and the QT platform is used to implement each module.Finally,the accurate three-dimensional model was generated from the potato sequence image.In this paper,the accuracy of feature matching,pose estimation and density of three-dimensional model in the process of potato reconstruction were analyzed in detail,and effective strategies were put forward.The final reconstruction results prove the feasibility of the algorithm. |
