POINTNET++LR3D: AN IMPROVED POINTNET++ MODEL FOR INDIVIDUAL IDENTIFICATION OF PIG BACK POINT CLOUD
POINTNET++LR3D:一种改进的POINTNET++模型用于猪背点云个体识别
DOI : https://doi.org/10.35633/inmateh-75-98
Authors
Abstract
Individual pig identification is a key technology to realize fine farming management, which is of great value in the fields of animal behaviour tracking and health monitoring. Aiming at the limitations of traditional 2D vision methods in stereo feature extraction, this study uses pig back point cloud to effectively capture deep 3D features such as back contour and skin texture and proposes an improved PointNet++ model for pig individual identification, which explicitly captures the local geometric and feature differences through two-stream differential coding, refines the feature distribution by using low-rank bilinear decomposition and residual sharpening strategies, and then establish the two-way dependency between channel and space to generate the global perceptual map, and finally combine with Mish activation function to enhance the nonlinear feature extraction. The experiment takes the hybrid long white pig as the research object and uses the Intel D435i depth camera to collect data and construct the segmentation and identification model. The results show that the improved model PointNet++LR3D achieves an overall accuracy of 97.11% in the individual identification task, which is an improvement of 1.9% compared to the base PointNet++MSG model. In addition, extended tests on the ModelNet40 dataset show an improvement in classification accuracy to 93.1%, validating the generalization ability of the architectural improvements. This study provides an efficient solution for non-contact pig identification based on point cloud, demonstrating the potential for application in fine-tuned farming.
Abstract in Chinese