| Potato is an important crop in our country.With the popularization of agricultural automation,potato nondestructive testing and quality grading based on machine vision have become the current research hotspots in the upgrading of the potato industry.The digital three-dimensional information of the potato plays an important role in the automatic classification of the potato.The main research content of this paper is to design a set of trinocular stereo vision system to complete the 3D reconstruction of potato.The thesis mainly focuses on trinocular camera calibration,feature matching algorithm improvement,camera pose estimation,sparse point cloud and dense point cloud reconstruction.The main contents of this paper are as follows:(1)CompIete the calibration of the trinocular camera.Camera calibration is an important part of 3D reconstruction.This article draws on the calibration methods of monocular and binocular cameras,uses three cameras to collect pictures of the calibration board at the same time,and combines the three cameras into three sets of binocular stereo vision systems.The two-calibration method completes the calibration of the trinocular camera,which ensures the accuracy of the camera calibration and lays the foundation for the subsequent 3D reconstruction work.(2)Aiming at the problem of long generation time of feature descriptors of feature matching algorithm,improve the SIFT feature matching algorithm.Since the collected potato icons have a single background and a small-scale transformation,the 128-dimensional feature descriptor generated by the original SIFT algorithm is reduced to 64 dimensions.By generating 64-dimensional feature descriptors,the SIFT feature matching time is greatly reduced,and then the RANSCA algorithm is used to eliminate Mismatched points:By comparing the running speed of the original SIFT algorithm after eliminating mismatches and the improved SIFT algorithm,the results show that the improved SIFT algorithm increases the speed of feature matching by nearly 36%.(3)Complete camera pose estimation,obtain potato sparse point cloud and dense point cloud to complete 3D reconstruction.Firstly,the essential matrix of the camera is solved according to the eight-point algorithm to estimate the camera pose relationship,and the 3D point coordinates and camera pose are nonlinearly optimized by beam adjustment to improve the accuracy of 3D reconstruction.Then use the optimized data to reconstruct the 3D point cloud of the potato to get the sparse 3D point cloud of the potato.In addition,the MVS algorithm based on the depth map is used to reconstruct the dense point cloud with the obvious surface features of the object.Finally,through comparative experiments,it is concluded that under the same conditions,trinocular stereo vision is better than binocular stereo vision to reconstruct the 3D dense point cloud of potato.In the process of potato 3D reconstruction,this paper proposes feasible solutions for trinocular camera calibration,feature matching,camera pose estimation and dense point cloud generation.It has certain application value in the field of potato non-destructive quality classification.The three-dimensional reconstruction of crops in machine vision is also useful for reference. |
