Research On Weight Prediction Of Goose Based On Depth Camera

China is the largest goose breeding country in the world.The body size and weight of geese are important parameters in geese production.The traditional manual measurement of body size and body weight of meat geese is time-consuming and laborious,which causes great stress response to meat geese and is not conducive to welfare breeding.Therefore,based on image processing technology,an automatic body weight and image acquisition system for meat geese is studied in this paper.Two depth cameras are used to obtain two-dimensional and depth images of meat geese to extract body size parameters.The body size parameters can reflect the growth and development characteristics of meat geese directly.According to the body size parameters,the body weight prediction models of meat geese are established respectively,which provide a feasible method for rapid measurement of body size and body weight prediction of meat geese,so as to realize welfare breeding of meat geese.The main research results of this paper include the following points:(1)Due to the large stress response of meat geese during manual weighing,Intel Realsense D435 depth camera,resistance strain type pressure sensor,HX711 ADC and Raspberry Pi were used to design and build an automatic weight and image acquisition system for meat geese,so as to realize automatic image information acquisition and weight data acquisition for meat geese in the relatively open area.Save time cost,reduce stress response of meat goose,meet welfare breeding requirements.The maximum relative error of Raspberry Pi electronic scale is 0.17%,and the minimum is 0.04%.The error accuracy meets the requirements,which can be used in automatic weight detection of meat geese to provide data support for subsequent image processing.(2)The volume size parameter extraction method of two-dimensional image is studied.After OTSU threshold segmentation method was used in image morphology processing,the chest,abdomen and tail information of meat goose were missing.Based on this,an adaptive threshold segmentation method was proposed to extract the complete contour information of meat goose.Based on the maximum inner tangent circle and the body scale contour,the positions and calculation methods of body slanting length,chest depth,chest width,keel length and hip width in the image were determined.Correlation and stability analysis were carried out between the automatic extraction value and the manual measured volume size parameter value showed that the maximum R~2 of keel long-like was 0.86197 and the minimum R~2 of oblique length was 0.7798,indicating a high correlation.The stability of the thoracoid depth was the highest,σ=0.591,and the stability of the slanting length of the thoracoid body was the lowest,σ=2.427.(3)The depth of the image processing method uses the pass-through filter to the original point cloud image by cutting from the X,Y and Z axis to remove the background and fence effect,and uses Statistical Outlier Removal filter to remove outliers,so as to extract the goose surface model.Voxel Grid filter is used for subsampling to minimize the quantity of point clouds and improve the efficiency of subsequent 3D reconstruction on the basis of maintaining the meat goose surface model.ICP algorithm is used to register two point cloud images,and axis alignment bounding box method is used to register point cloud images.In the process of 3D model reconstruction,the triangulation algorithm is used to get the three-dimensional triangulation model of the goose,and then through Helen’s formula,the surface area and volume of the goose are calculated,which were used as the basis of the body weight prediction model based on the 3D body size characteristics.(4)Based on the body size parameters extracted from 2D images and the volume and surface area parameters extracted from 3D models,five weight prediction models of meat geese are established,which separately are BP neural network,RBF neural network,SVM support vector machine,stepwise regression analysis(SMLR)and partial least square(PLS)regression analysis.The minimum relative root mean square error of the SMLR-based weight prediction model is 12.2%and R~2is 0.92154.Among the meat geese weight prediction models based on volume and surface area,the R~2 of the weight prediction model based on BP neural network is the largest,which is 0.91278;and the relative root mean square error is the smallest,which is 3.8%.It provides a feasible method to predict the body weight of meat goose.