基于改进PointNet++的输电杆塔点云语义分割模型

Semantic Segmentation Model for Transmission Tower Point Cloud Based on Improved PointNet++

针对现有输电线路点云提取精度不高、无法满足无人机自主精细化巡检需求的问题,提出一种改进的PointNet++的输电杆塔点云语义分割方法,以实现对导线、地线、引流线、绝缘子和杆塔塔身的点云分割。首先,对经典PointNet++模型参数进行调整,使模型在特征提取数量、感受野方面更适用于输电杆塔点云数据;然后,采用核心点卷积作为点云特征提取算法,进一步提升模型对点云特征的提取能力;最后,针对点云数据中存在的数据不平衡问题,采用focal loss作为损失函数,使占比较少的类别得到充分训练。为验证所提方法有效性,在2284基输电杆塔组成的点云数据集上进行了实验,实验结果表明:改进后的算法平均F1值达到97.26%,较经典PointNet++提高了3.95个百分点。

Aiming at the existing problem that the point cloud extraction accuracy of transmission lines is not high and cannot meet the needs of autonomous and refined inspection by unmanned aerial vehicles,an improved PointNet++semantic segmentation method for transmission tower point cloud is proposed,which realizes the segmentation of wires,ground wires,drainage lines,insulators and towers.Firstly,the parameters of the classic PointNet++model are adjusted to make the model more suitable for the point cloud data of transmission towers in terms of feature extraction quantity and receptive field;then,the kernel point convolution is used as the point cloud feature extraction algorithm to further improve the model's ability to detect point cloud features;finally,for the data imbalance problem in the point cloud data,the focal loss is used as the loss function,so that the categories with a small proportion can be fully trained.In order to verify the effectiveness of the proposed method,experiments are carried out on the point cloud dataset composed of 2284 transmission towers.The experimental results show that the average F 1 value of the improved algorithm reaches 97.26%,which is 3.95 percentage points higher than that of the classic PointNet++.