Study On Estimation Of External Geometric Features Of Pummelo Based On 3D Reconstruction

China is the country with the largest planting area and the largest output of pummelo in the world.However,the grading of pummelo still depends on labor,which is time-consuming and laborious.It is urgent to improve the detection level and production efficiency of commercial processing equipment.Taking Guanxi pummelo as the research object,this paper studied the key technical problems such as 3D reconstruction and complement of Guanxi pummelo,and established the estimation method of volume,density,longitudinal diameter and transverse diameter of Guanxi pummelo based on 3D reconstruction.The main contents and conclusions are as follows:(1)A machine vision system for multi view 3D reconstruction of pummelo was constructed.The system was mainly composed of illumination system,rotating platform,image acquisition device and acquisition software.4 illumination schemes were compared,including transverse circumferential array strip light source.The average brightness of the pummelo surface were143.1,184.6,191.6 and 196.1,and the standard deviations were 16.4,25.7,23.3 and 5.8respectively.Therefore,the cylindrical uniform distribution surface light source scheme was determined at last,which could overcome the problem of image bright spots caused by strong reflection on the pummelo surface.MV-CA013-20GC industrial cameras and 8 mm industrial lens were selected.The image acquisition device and acquisition software could realize the synchronous trigger acquisition of three cameras of pummelo.(2)A real-time reconstruction method of Guanxi pummelo based on double view point cloud was proposed.Using the developed 3D image processing software of pummelo,this paper compared and analyzed the average residual error fluctuation and absolute distance error of four depth cameras at a depth of 0.55～0.95 m,including Azure Kinect DK,Real Sense D455,Real Sense L515 and HLT003S-001.It was found that the average absolute distance error was less than 1 mm,7 mm,3 mm and 2 mm respectively.It was determined to use Azure Kinect DK depth camera for reconstruction,so as to reduce the missing area of point cloud surface and realize real-time acquisition and reconstruction of Guanxi pummelo.(3)A high-precision reconstruction and completion model of Guanxi pummelo 3D point cloud was established.Based on the constructed uniform light environment machine vision system,a method to solve the scale uncertainty and reconstruction accuracy evaluation of pummelo point cloud by using two marker plates was proposed,and a multi view 3D point cloud model reconstruction scheme was established.and the high-precision reconstruction of Guanxi pummelo3D point cloud was realized.The accuracy error of reconstructed model point cloud was-0.66～0.72 mm.MSN point cloud completion network model was used for training,and the training model was used to complete the hole point cloud.The results showed that the average completion generation speed was 10.46 per second.(4)A estimation methods of volume,density,longitudinal diameter and transverse diameter of pummelo based on 3D model were established.A three-dimensional point cloud method based on minimum directed bounding box and triangulated convex hull was proposed to analyze the influence of the reconstructed model on the estimation of volume,density,longitudinal diameter and transverse diameter of pummelo.The results showed that the fine filtered,down sampling and complement point cloud estimated values were positively correlated with the measured values of pummelo,including volume,density,vertical and horizontal longitude.The average relative errors of volume were 5.04%,3.91%and 3.17%,while the average absolute errors of density were 0.031g/cm~3,0.025 g/cm~3 and 0.020 g/cm~3 respectively.At the same time,the average absolute errors of longitudinal diameter were 2.80 mm,1.69 mm and 3.82 mm,and the average absolute errors of transverse diameter were 2.51 mm,1.84 mm and 2.47 mm respectively.Therefore,it was determined that the completed point cloud could be used for volume and density estimation,and the down sampling point cloud could be used for longitudinal and transverse diameter estimation.