Research On Segmentation And Registration Of Citrus Point Cloud Based On Kinect V2 Camera

In order to improve the low efficiency of Citrus picking,the development of machine automatic picking has become the current development trend.At present,the problems of fruit occlusion and clustering in automatic picking affect the accuracy of recognition and segmentation,and seriously restrict the rapid development of Citrus picking technology.This project sets citrus as the research object,and uses color characteristics and three-dimensional geometric features to identify and divide overlapping fruits through the Kinect v2 camera,which provides an important reference for the study of mechanized fruit picking.The research content of this topic is as follows:The point cloud segmentation method based on Euclidean clustering and supervoxel clustering is studied.A point cloud segmentation method combining Euclidean clustering,supers voxel clustering and LCCP algorithm is studied.Taking citrus as the research object,the citrus fruit and background are segmented by color threshold method,and the citrus fruit point cloud is obtained after outlier statistical filtering to filter out the noise points.The European clustering algorithm is used to divide the fruit point cloud into point cloud clusters independently,and the threshold is used to judge whether the segmentation is completed.The unfinished point cloud clusters are segmented again using supervoxel clustering algorithm and LCCP algorithm.Two groups of experiments are designed to verify the accuracy of the algorithm.In Experiment 1,869 citrus fruit point cloud samples were collected in a simulated environment to study the three situations of fruit non-contact,fruit clustering and fruit shielding.The results show that the average recognition accuracy of the improved algorithm is 97.12%,the recall rate is 98.13%,and the comprehensive score is 0.97.In Experiment 2,the shielding area was more than 50% and less than 50% respectively.The results showed that the citrus recognition effect was better when the occlusion area was less than 50%,but for the fruit with higher occlusion area,the fruit recognition accuracy was 89.84%,which was mainly due to the loss of point cloud information.Aiming at the problems of Euclidean clustering combined with supervoxel clustering,the point cloud registration method is used to improve the problem of lack of point cloud information.Point cloud registration is divided into coarse registration and fine registration.During the coarse registration of point cloud,the problem of large number of point cloud registration of fruit trees is solved by extracting feature points.Then the feature points are described by fpfh feature points,and then the sac-ia algorithm is used to complete the coarse registration.KD tree accelerated ICP algorithm is used to replace the traditional ICP algorithm,which improves the calculation efficiency and accuracy of fine registration,and the method of error point elimination is used to complete the fine registration.The experimental results show that the improved ICP algorithm proposed in this paper has improved in speed and efficiency.The comprehensive test results showed that the segmentation accuracy and comprehensive score of fruit occlusion area > 50% were 94.53% and 0.96 respectively,which were better than the corresponding results of 89.84% and 0.91 before no point cloud treatment.Therefore,the method of adding point cloud registration before fruit segmentation can achieve a better segmentation effect for large-area shaded fruits.Based on the above research,a citrus point cloud registration and segmentation algorithm is proposed,and a field experiment is carried out.A total of 245 citrus fruits were collected from 30 citrus pictures in the production site.The test results showed that the accuracy of covering area > 50% and covering area < 50% were 87.64% and 92.307% respectively,the recall rate was 92.85% and 95.36% respectively,and the comprehensive score was 0.90 and0.94.The comprehensive results meet the production technical requirements.The research content of this subject has certain application value for improving the fruit recognition efficiency of picking robot and further promoting the rapid development of Citrus picking technology.