Point cloud-based pig body size measurement featured by standard and non-standard postures

• This paper studies the influence of point cloud-based pig's standard and non-standard posture on the accuracy of body size automatic measurement. • This paper proposed a point cloud-based pig posture quantization algorithm, which based on the median skeleton extraction gets 22 key points of the pig's skeleton and calculates all joint point vector eigenvalue set. • Both linear and nonlinear regression can effectively adjust the non-standard body length to the standard posture results. Automatic body size measurement of livestock using 3D vision technology is one of the current research focuses. Point clouds obtained from freely walking livestock involve a variety of postures. However, measuring livestock body sizes requires their upright and straight standard postures, and non-standard measurement postures such as bending their body or lowering their head lead to inaccuracy in body size estimation. In this paper, we analyze 207 groups of point cloud data collected from 25 landrace pigs, and propose a standard posture classification algorithm. By using the algorithm based on the median skeleton extraction, 22 key points of the pig's skeleton were obtained and all joint point vector eigenvalue sets were calculated. Then Bagging-SVM Posture Classifier was constructed, and by three experts’ votes, comparisons were made between the automatic classification results and the manual classification results based on standard and non-standard postures of pigs. According to the automatic calculation of four parameters including body length, body height, body width and abdominal girth in standard and non-standard postures, the experimental data showed that the body size measurement results in standard postures presented high stability and consistency, while the results in non-standard postures fluctuated considerably, which significantly affected the accuracy and stability. The experiment showed that the ratio of standard postures to non-standard postures was 1:3. Finally, based on joint point vectors eigenvalues, the regression models were applied to adjust body length, body height, body width and abdominal girth in non-standard postures. The linear regression and nonlinear regression methods such as BP regression and SVR support vector regression were used for comparison. The experimental results indicated that various regression methods can significantly improve the accuracy of automatic measurement by correcting the pig's body size measurement results in non-standard postures. Therefore, it is essential to classify and adjust the livestock point cloud postures to improve the accuracy of three-dimensional measurement.