一种基于空间特征的三维点云语义分割模型

A Semantic Segmentation Model of 3D Point Clouds Based on Spatial Features

摘要: 真实的点云场景通常包含非常复杂的环境，加上三维点云数据所具有的离散性、无序且分布不均匀的属性，针对点云的详细信息分析往往非常具有挑战性。针对以上问题，提出一种基于编码器–解码器的分割网络。首先，通过关键点提取模块获取点云的形状特征并通过全连接层提取形状关键点；然后，通过空间细节提取模块获取基于关键点的空间细节特征；最后，结合形状特征和空间细节特征，获取点云丰富的空间信息及上下文联系，提高网络的泛化能力。实验结果表明，我们的网络在公共数据集Urban Semantic 3D (US3D)和ISPRS Vaihingen 3D semantic labelling benchmark (ISPRS)上的平均交并比分别为93.44%和81.45%。我们的网络分割性能良好，且具有较好的泛化能力。

Abstract: Real point cloud scenes usually contain very complex environments. Coupled with the discrete, dis-ordered and uneven distribution properties of 3D point cloud data, detailed analysis of point clouds is often very challenging. Aiming at the above problems, an encoder-decoder-based segmentation network is proposed. First, the shape features of the point cloud are obtained through the key point extraction module and the shape key points are extracted through the fully connected layer; then, the spatial detail features based on key points are obtained through the spatial detail extraction module; finally, the shape features and spatial detail features are combined to obtain the rich spatial information and contextual connections of point clouds improve the generalization ability of the network. Experimental results show that our network achieves an average intersection ratio of 93.44% and 81.45% on public datasets Urban Semantic 3D (US3D) and ISPRS Vaihingen 3D semantic labelling benchmark (ISPRS), respectively. Our network has good segmentation performance and good generalization ability.