59 / 2025-04-14 20:16:44

Body Parts Segmentation and Phenotypic traits extraction of pig Using an improved Point Cloud Segmentation Network with Multi-LiDAR

Computer vision,3D reconstruction,Point-cloud semantic segmentation,Deep learning

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

Precise and efficient acquisition of pig phenotypic traits is becoming increasingly essential in modern pig breeding and management. With the advancement of three-dimensional(3D) reconstruction technology, the automated collection of phenotypic parameters, is becoming increasingly feasible. Owing to the scene adaptability limitations of depth cameras and the influence of pig movement, mainstream point-cloud-based methods face limitations with incomplete and irregular point-cloud data, which affect the accuracy of body dimension measurements. Comparatively, LiDAR can accurately acquire point-cloud data with a field of view. Furthermore, Point-cloud segmentation technology enhances the accuracy and robustness of phenotypic parameters, calculations by dividing the point-cloud data into different regions. This study proposes an efficient pig phenotypic parameters extraction method within commercial pig farms based on a point-cloud segmentation network with Multi-LiDAR. First, we built a data collection platform using two LiDAR sensors with superior synchronization and reliability to produce high-quality point-cloud data for pigs. Second, we proposed an improved point-cloud segmentation network named PointVector++ to segment the pig point-cloud data to extract parts of the pig body. The improved model integrates a learnable downsampling module and the VPSA++ module, enhancing the rotation invariance, density balance, and feature adaptability to handle complex point sets effectively. Finally, a phenotypic parameters calculation algorithm for 10 key metrics was developed based on different parts after segmentation. Experiments were conducted using point-cloud data from 150 live pigs to evaluate the segmentation and phenotypic parameters assessment performance. The PointVector++ model achieved an mIoU of 90.5% on the pig dataset. In 10 metrics, the body length, chest width, hip width, abdominal width, and hip height exhibited high accuracy, with an MAE and MAPE below 3.61 cm and 6.88%, respectively, and R² values exceeding 0.64. This method can effectively partition pigs, extract body dimension information, and provide a novel approach for the automated acquisition of pig phenotypic traits.