Automatic Measurement Method of Pig Body Size Based on Improved Point Transformer Segmentation Network

At present, most of the non-contact body size measurement methods based on the complete point cloud of pigs are complex and susceptible to interference from other parts of the point cloud when selecting key points. Therefore, this paper proposes an automatic measurement method of pig body size based on improved Point Transformer point cloud segmentation network. Firstly, a point cloud acquisition channel for pigs was built to collect point clouds on the left and right sides of 56 Yorkshire pigs aged 80-110 days. Secondly, the point cloud is preprocessed to remove the noise outside the pig point cloud, and the pig point cloud segmentation data set including head, chest, abdomen, buttocks and tail is constructed. Then, based on the Point Transformer, the adaptive convolution (AdaptConv) is introduced into the upsampling module, and the ACPT network is proposed for pig point cloud segmentation by combining the offset attention mechanism and the Focal loss function. Finally, the key points of the five-part segmentation part of the pig are found for body size measurement. The experimental results show that the accuracy of the improved segmentation network is 88.5 %, which is 5.5 % higher than that of the original network and better than other networks. The absolute error between the automatic measurement results and the measured values was within 4.8 cm, and the average relative error was within 5.1 %. The research shows that this method improves the measurement accuracy and has a better application prospect in non-contact body measurement.