基于改进PointNet++的大规模猪体点云部位分割

Segmentation of Large-scale Pig Body Point Cloud Parts Based on Improved PointNet++

PointNet++点云分割网络能直接处理点云并有良好的分类和分割效果,然而对于自由运动状态下获取的大型牲畜点云,存在较大姿态差异,PointNet++无法难以直接捕捉其局部特征,对牲畜体点云各部位分割效果不佳。针对此问题论文采用Octree结构改进了PointNet++集合抽象层中的分组与采样层,使得网络能够更好地捕捉目标点云的非刚性变化,充分提取不同层次下的局部信息,实现对大规模牲畜点云数据的自动分割。采用Octree改进的PointNet++模型分割活体猪点云数据的头部、耳朵、尾部、躯体和四腿。实验表明,500组猪体点云数据,采用320组不同姿态下的猪体点云进行网络训练,180组分割测试结果的平均mIoU达到了88.06%,OA达到了96.57%。

ion layer of PointNet++,so that the network can better capture the non-rigid changes of the target point cloud,fully extract the local information at different levels to realize automatic segmentation of the large-scale livestock point.The head,ears,tail,body and four legs of pig point cloud data are segmented by improving PointNet++model.Experiments show that 500 groups of pig body point cloud data are used for network training with 320 groups of pig body point clouds in different poses.The average mIoU of 180 groups of segmentation test results has reached 88.06%,and the OA has reached 96.57%.