摘要
三维点云无序不规则的特性使得传统的卷积神经网络无法直接应用,且大多数点云深度学习模型往往忽略大量的空间信息。为便于捕获空间点邻域信息,获得更好的点云分析性能以用于点云语义分割,文中提出Graph⁃PointNet点云深度学习模型。Graph⁃PointNet在经典点云模型PointNet的基础上,结合二维图像中聚类思想,设计了图卷积特征提取模块取代多层感知器嵌入PointNet中。图卷积特征提取模块首先通过K近邻算法搜寻相邻特征点组成图结构,接着将多组图结构送入图卷积神经网络提取局部特征用于分割。同时文中设计一种新型点云采样方法多邻域采样,多邻域采样通过设置点云间夹角阈值,将点云区分为特征区域和非特征区域,特征区域用于提取特征,非特征区域用于消除噪声。对室内场景S3DIS、室外场景Semantic3D数据集进行实验,得到二者整体精度分别达到89.33%和89.78%,平均交并比达到64.62%,61.47%,均达到最佳效果。最后,进行消融实验,进一步证明了文中所提出的多邻域采样和图卷积特征提取模块对提高点云语义分割的有效性。
Since the traditional convolutional neural network cannot be applied directly due to the disorder and irregularity features of 3D point cloud,and most point cloud depth learning models often ignore a large amount of spatial information,a point cloud deep learning model⁃Graph⁃PointNet is proposed to capture the neighborhood information of spatial points and obtain better performance of point cloud analysis for the point cloud semantic segmentation.On the basis of the classical point cloud model PointNet,and in combination with the clustering idea in two⁃dimensional image,a graph convolution feature extraction module is designed in Graph⁃PointNet to replace the multi⁃layer perceptron embedded in PointNet.In the graph convolution feature extraction module,the adjacent feature points are searched by means of the K⁃nearest neighbor algorithm to form the graph structure,and then multiple groups of graph structures are sent into the graph convolution neural network(GCNN)to extract local features for the segmentation.In this paper,a new point cloud sampling method(multi⁃neighborhood sampling)is designed.The multi⁃neighborhood sampling is used to divide point clouds into feature regions and non feature regions by setting the threshold of an intersection angle between point clouds.The feature region is used to extract features,and the non feature region is used to eliminate noise.The results of extensive experiments on indoor scene S3DIS and outdoor scene Semantic3D data sets show that these overall accuracy reached 89.33%and 89.78%,respectively,and the average intersection and merging ratio reached 64.62%and 61.47%,respectirely,which have all achieved the best effect.The ablation experiment is carried out to further prove that the proposed multi neighborhood sampling and graph convolution feature extraction module is effective in improving the semantic segmentation of point cloud.
作者
陈苏婷
陈怀新
张闯
CHEN Suting;CHEN Huaixin;ZHANG Chuang(Nanjing University of Information Science&Technology,Nanjing 210044,China)
出处
《现代电子技术》
2022年第6期87-92,共6页
Modern Electronics Technique
基金
国家自然科学基金项目(61906097)。
