3D Construction Of Fruit Tree Branches In Natural Growth Based On Kinect Camera

In recent years,China is undergoing reforms and adjustments to the agricultural industry.Under the influence of national policies,the area under the influence of national policies has continuously expanded,which has become a key industry in the economic development of my country’s forest and fruit producing areas and a new channel for rural farmers to increase income.Currently,there is a large gap in demand for fruit and fruit.However,fruit picking is generally manual picking,which has the disadvantages of long picking time and high cost.The labor shortage problem in an aging society is particularly prominent,and it has become an important reason for restricting the development of the fruit industry.one.As one of the important signs of modern agricultural automation,fruit and vegetable picking robots provide a good solution to this problem.However,some of the fruits of the forest are small and dense,so it is necessary to identify the distribution position of the fruit and the direction of the branches,so as to be fast and accurate.Recognizing and positioning fruit trees on the ground and planning the picking route of the picking robot manipulator in real time is still a hot and difficult point of current research at home and abroad,so fruit recognition and fruit tree construction based on computer vision are particularly important.The main research work of this article is:By combining the growth state of fruit trees under natural conditions,a depth camera is selected for camera calibration,the distance suitable for collecting data is determined,and the color point cloud database is established to prepare for subsequent further algorithm experiments,analysis and verification.Improve the point cloud matching algorithm.The improved point cloud matching algorithm registers point clouds from different angles to overcome the problem of tree occlusion under a single angle of view,so as to obtain complete tree point cloud data.This paper proposes an improved method based on the classic ICP(Iterative Closest point)algorithm.In the first stage of feature point matching,the source point cloud pose is adjusted according to the descriptor of the feature point to complete the initial matching,and then the ICP registration of the two point clouds is performed;In the second stage,the point cloud matching result of the previous stage and the third piece of point cloud are once again subjected to initial matching and fine matching,and the final point cloud matching result is obtained.The point cloud data matching experiment was performed on the point cloud data through two algorithms,which verified the accuracy and efficiency of the improved algorithm.The root mean square value of the improved ICP matching algorithm is generally lower than that of the classic ICP algorithm.Although some classic algorithms have a lower root mean square value,it is difficult to avoid the local optimal solution.The improved algorithm can achieve more accurate point cloud matching and has better robustness.The improved algorithm uses SIFT feature points for initial registration,which reduces a lot of running time compared to traditional algorithms.It can be seen that the accuracy and efficiency of the ICP matching algorithm are improved.Three-dimensional point cloud branch reconstruction of fruit trees.The three-dimensional tree reconstruction method based on cylindrical fitting is a three-dimensional image voxel reconstruction method for fruit tree reconstruction and trunk and branch recognition.The main steps of the algorithm are skeleton extraction and skeleton analysis.Skeletonization is based on a skeleton extraction algorithm,which provides the skeleton and skeleton features of trees.In the process of skeleton analysis,skeleton features are used to distinguish tree trunks from first-level branches,second-level branches,and so on.This article lists the total number of actual branches of each group of data and the number of branches determined by this algorithm.268 branches were evaluated on 20 trees.Among them,46 branches were misidentified,resulting in a false negative rate of 17.67%.The average branch recognition accuracy of this algorithm is 82.92%.Although there is a big difference in branch recognition accuracy(caused by false positives),the algorithm also achieves a recognition accuracy of 80% or higher for more than 65% of the trees evaluated in this study.