高低维双局部特征融合的点云分类分割网络

Point cloud classification and segmentation network based on double local features fusion of high-dimensional and low-dimensional

点云的分类和分割是三维点云网络的基础步骤,是三维重建、目标识别和自动驾驶等研究的关键环节。针对现有三维点云网络进行分类和分割时无法充分利用点云局部信息而导致精度不高的问题,提出了一种高低维双局部特征融合的点云分类分割网络。本文网络利用图卷积提取点云基于位置信息的低维局部特征,并通过注意力机制提取点云基于语义信息的高维局部特征,同时使用多层感知机提取点云的全局特征,通过多个特征融合的方式来提高点云分类和分割的精度。实验结果表明:本文网络在点云分类任务中,总体精度达到93.4%,平均精度达到90.3%;在点云分割任务中,平均交并比达到86.2%;相较于现有的主流网络具有较好的分类和分割精度。同时对于稀疏点云输入的问题,本文网络具有较强的鲁棒性。

The classification and segmentation of point cloud is the basic step of 3D point cloud network,which is a key link of 3D reconstruction,target recognition and automatic driving.A point cloud classification and segmentation network based on double local features fusion of high-dimensional and low-dimensional is proposed to address the problem that the existing 3D point cloud network cannot make full use of the local information of point cloud for classification and segmentation,resulting in low accuracy.In this paper,graph convolution is used to extract the low-dimensional local features of point cloud based on location information,and the attention mechanism is used to extract the high-dimensional local features of point cloud based on semantic information.At the same time,the MLP is used to extract the global features of point cloud,and the accuracy of point cloud classification and segmentation is improved by multiple feature fusion.The experimental results show that the overall accuracy of this network is 93.4%,and the average accuracy is 90.3%in the point cloud classification task;and the average intersection and merging ratio is 86.2%in the point cloud classification task,which has better classification and segmentation accuracy compared with the existing mainstream networks.At the same time,the network has strong robustness for the problem of sparse point cloud input.