【摘要】 图卷积神经网络虽然可以对初始点云数据进行直接处理，但其只在局部尺度上独立提取点特征，未将局部点互相关联起来则会影响点云数据分类精度。因此，本文在考虑点云数据整体几何关系和拓扑信息的基础上，提出了一种基于滑动图卷积神经网络的输电线路点云分类模型。首先，通过最远点采样方法从原始点云中迭代子采样点集，有效地降低模型复杂度和时间消耗；其次，利用多尺度K近邻对子采样点集建立局部有向图；再次，采用边缘卷积滑动地提取局部图特征，计算点云上的每一个点与其相邻点之间的边缘特征；最后，利用全局最大池化层进行点云分类。所提模型首先在公用数据集上进行预训练，之后再用标注过的由激光雷达实地采集的输电线路点云数据进行验证。实验结果证明，本文所提模型在公用数据集和实际数据集上均取得较好的分类效果，分类准确率比通用的ECC、PointNet、PointNet++等方法高出至少1.5%。更多还原

【Abstract】 Although the graph convolutional neural network can directly process the initial point cloud data, it only extracts point features independently on a local scale. Failure to associate local points with each other will affect the classification accuracy of point cloud data. Therefore, this paper proposes a transmission line point cloud classification model based on sliding graph convolutional neural network based on the overall geometric relationship and topological information of the point cloud data. First, the sub-sampling point set from the original point cloud is iterated through the farthest point sampling method, effectively reducing the complexity and time consumption of the model. Secondly, the multi-scale K-nearest neighbor pair sub-sampling point set is used to build a local directed graph. Again, edge convolution is used to extract local map features in sliding mode, and to calculate the edge features between each point on the point cloud and its neighboring points. Finally, the global maximum pooling layer is used to classify the point cloud. The proposed model is first pre-trained on a public data set, and then validated with the labeled point cloud data of the transmission line collected by the lidar on the spot. Experimental results prove that the proposed model has achieved good classification results on both public data sets and actual data sets, and the classification accuracy is at least 1.5% higher than general ECC, PointNet, PointNet++ and other methods.更多还原