| Animal husbandry is the basic industry and characteristic advantageous industry in Xinjiang,among which mutton sheep breeding industry has always occupied an important position.Body weight is a key index to measure the growth and development of sheep in the process of mutton sheep breeding.In the traditional animal husbandry management mode,the measurement of body size and body weight of sheep mainly depends on manual work,which has high intensity and low efficiency.In addition,it will also frighten sheep and produce stress response.Therefore,it is necessary to seek a non-contact sheep body weight estimation method.In this thesis,a point cloud data acquisition platform for sheep was built based on Kinect V2 equipment,and the point cloud data of sheep side view and top view were collected.The point cloud filter was used to remove noise points and repair the missing part of the side view point cloud.The sampling consistency algorithm is used to register the side view point cloud and the top view point cloud.On this basis,three fine registration methods are used for fine registration,and the best registration point cloud is selected for mirroring and surface reconstruction.The body size parameters are calculated according to the key points,and the body size parameters are corrected based on the wool thickness.Finally,a variety of methods were used to es Tablelish a weight prediction model based on the modified body size parameters and sheep body surface feature information,and the best sheep body weight prediction model was selected.The main contents and results are as follows :(1)Build a sheep point cloud data acquisition platform,collect sheep point cloud data and perform preprocessing.In this thesis,Kinect V2 was used as the sheep body point cloud acquisition device.According to the experimental environment,the sheep body point cloud data acquisition platform was built to collect the side view point cloud and the top view point cloud.The useless information such as limit device,outliers and ground noise in the original point cloud is removed by straight-through filtering,statistical filtering and random sampling consistency segmentation algorithm respectively,and the processed point cloud data containing only sheep ’s side view and top view is obtained.For the occlusion part of the side-view point cloud railing,the moving least squares point cloud repair method is further used to repair the missing part of the side-view point cloud,and the complete sheep side-view point cloud data is obtained.(2)A target rod guided sheep point cloud registration and 3D model surface reconstruction method is proposed.The limit device rod with obvious edges and corners is used as the target rod for point cloud data registration,and the side-view and top-view point clouds are guided for iterative registration.The closest point iterative algorithm(ICP),the trimmed closest point iterative algorithm(Trimmed ICP)and the generalized closest point iterative algorithm(GICP)are used for fine registration.The best fine registration result is selected for mirroring processing to obtain a complete three-dimensional point cloud of the sheep body,and the Alpha shape algorithm is used to reconstruct the sheep body surface.The results show that the registration time of the initial registration is 5.086 s,and the registration accuracy is 3.259 cm.The registration time of GICP algorithm is 0.104 s,and the registration error is 1.162 cm.When alpha is 0.05,the sheep body surface model reconstructed by Alpha-shape algorithm has the best effect and can be used for the establishment of subsequent weight estimation model.(3)A coefficient weighted body size correction algorithm based on the difference of wool thickness between the neck and the leading edge of the scapula is proposed.The average Euclidean distance of the point cloud clusters at the neck and the leading edge of the scapula was calculated to obtain the difference in wool thickness.The linear coefficient weighting method was used to correct the body size of the reconstructed sheep torso.Five body size parameters of abdominal width,body height,body oblique length,body depth and hip height were selected,and the key point recognition algorithm was used to extract the sheep body size parameters.The results show that the measurement accuracy of body size parameters based on wool thickness correction is higher,and the average relative error is 2.45 %,which can be used for the establishment of subsequent weight estimation model.(4)Based on the sheep body surface feature information and the corrected body size parameters,the sheep body weight prediction model was established.The linear weight prediction model was established based on the envelope surface area and volume parameters of the intact sheep,and the envelope surface area and volume parameters of the sheep without limbs and head.At the same time,the linear and nonlinear weight prediction models were constructed based on the modified body size parameters combined with machine learning.The comparison results show that the support vector regression model based on body size parameters has the best prediction effect,and the absolute root mean square error is 1.031 kg.(5)Development and verification of sheep body three-dimensional reconstruction and weight estimation software.Using C + + programming language and Qt Creator framework combined with PCL point cloud library to develop sheep 3D reconstruction and weight estimation software,the graphical interface development of sheep point cloud data access,preprocessing,surface reconstruction,body size measurement,weight estimation and other functions is completed,and the core functions of the software are verified.The software realizes the integration of sheep 3D reconstruction,body size measurement and weight estimation. |
