【摘要】 针对现有深度学习网络结构单独提取点特征方式使网络缺乏度量空间中点之间邻域关系，以及几何信息的学习导致模型处理复杂场景能力较弱的问题，提出一种基于注意力机制结合邻域几何特征的点云语义分割算法，算法通过在邻域特征学习过程中引入注意力机制将中心点特征聚集为其邻域点特征的加权和，利用点之间相关性获得深层次细粒度局部特征；采用多尺度局部特征提取策略消除算法采样过程中信息丢失的影响。经实验验证，该文所提算法在室内场景数据集S3DIS上的平均交并比为53.12%，相较于PointNet++算法提升了5.96%。更多还原

【Abstract】 Aiming at the problem that the existing deep learning network structure separately extracts point features,the network lacks the learning of the neighborhood relationship between points in the metric space and the geometric information,which leads to the problem that the model has a weak ability to deal with complex scenes. Domain geometric feature point cloud semantic segmentation algorithm. The algorithm introduces the attention mechanism in the neighborhood feature learning process to gather the central point features into the weighted sum of the neighborhood point features,and uses the correlation between points to obtain deep and fine-grained local features;Multi-scale local feature extraction strategy is used to eliminate the influence of information loss in the sampling process of the algorithm;Experimentally verified,the average intersection ratio of the proposed algorithm on the indoor scene data set S3DIS is 53.12%,which is an increase of 5.96% compared to the PointNet++ algorithm.更多还原