摘要
针对传统三维激光点云场景分割算法容易忽略目标边界模糊的问题,采用边界对比学习算法设计了三维激光点云场景分割网络,旨在通过对比学习提升模型在边界处的预测性能。首先采用PointNet++作为主干网络,通过多尺度的降采样特征编码和上采样特征解码来学习点云中不同类别的语义特征,并逐点预测目标类别,实现场景整体分割;然后引入对比学习算法,采用迭代的方式捕获子场景点云的边界,并挖掘出模糊的边界点;最后在网络训练阶段利用对比学习损失函数实现边界类别增强,大幅提升了对三维激光点云场景分割的精度。在公开的三维激光点云场景分割数据集上进行了大量实验,结果表明:所提分割算法在19个语义类别的点云中有15个的分割性能是最佳的,整体的指标性能均优于对比算法,消融实验和可视化结果也验证了所提算法可以有效改善三维激光点云场景分割任务中边界的预测性能,充分说明了所提算法的优越性。
To solve the problem that the traditional 3D laser point cloud scene segmentation algorithm tends to ignore the blurring of target boundaries,a 3D laser point cloud scene segmentation network is designed using a boundary contrastive learning algorithm,so as to improve the model's prediction performance at the boundaries through contrastive learning.Firstly,the PointNet++is taken as the backbone network,multi-scale downsampling feature encoding and upsampling feature decoding are used to learn the semantic features of dfferent target categories in the point cloud,and the prediction of target categories is conducted point by point,so as to achieve overall scene segmentation,Then,a contrastive learning algorithm is introduced to capture the boundaries of sub-scene point clouds through iterations and mine fuzzy boundary points.Finally,the contrastive learning loss function is used to enhance the differentiation of boundary points belonging to different categories at the network training stage,which significantly improves the accuracy of 3D laser point cloud scene segmentation.A large number of experiments are conducted on the publicly available 3D laser point cloud scene segmentation dataset,and the results show that the proposed algorithm has the bests egmentation performance in 15 out of 19 semantic categories,with overall performance indicators superior to the comparison algorithms.The ablation experiments and visualization results also verify that the proposed algorithm can effectively improve the category prediction performance of boundary points in 3D laser point cloud scene segmentation tasks,which fully demonstrates the superiority of the proposed algorithm.
作者
张迪
刘婷婷
宋家友
ZHANG Di;LIU Tingting;SONG Jiayou(School of Computer and Software Engineering,Sias University,Zhengzhou 451000,China;School of Information Engineering,Zhengzhou University,Zhengzhou 450000,China)
出处
《电光与控制》
CSCD
北大核心
2024年第5期54-59,共6页
Electronics Optics & Control
基金
河南省科技攻关项目(212102210150)。
关键词
三维激光点云
场景分割
深度学习
对比学习
点云边界
损失函数
3D laser point cloud
scene segmentation
deep learning
contrastive learning
point cloud boundary
loss function