基于改进PointNet++的电力线场景语义分割

A Semantic Segmentation of Power Line Scenes Based on Improved PointNet++

从机载激光雷达点云数据中自动准确地提取电力线和杆塔是常规电力线检测的关键步骤,针对目前基于无人机的电力线点云场景中输电线提取精度不高,难以满足无人机自主精细化巡检需求的问题,本文提出了一种新的基于深度学习的电力线点云语义分割方法。本文提出一种改进PointNet++网络架构,完成了对导线、地线、地面和杆塔塔身的分割。首先在输入电力线点云数据的基础上,对经典PointNet++模型参数进行调整,使该模型在特征提取数量、感受野方面更适用大场景输电线现场。然后,对set abstraction(SA)模块进行改进,增强模型对点云数据的特征提取能力。针对该模块主要做了两点改进,一是在每个SA模块中增加残差连接,二是在SA模块中增加了倒置瓶颈设计。基于自建的电力巡检数据集进行测试,结果表明,本文提出的方法在电力线点云语义分割任务上取得了比较好的性能。对比原网络,改进的网络在平均准确率上提升了5.1%,在平均交并比上提升了7.2%。

Automatic and accurate extraction of power lines and towers from airborne LiDAR point cloud data is a key step in conventional power line detection.A new deep learning based semantic segmentation method for power line point cloud is proposed to solve the problem of low extraction accuracy of transmission line in current unmanned aerial vehicle based power line point cloud scenarios,which makes it difficult to meet the autonomous fine inspection needs of unmanned aerial vehicles.An improved PointNet++network architecture is proposed to complete the segmentation of wires,ground wires,ground,and tower bodies.Firstly,based on the input power line point cloud data,the parameters of the classic PointNet++model are adjusted to make the model more suitable for large-scale transmission line field in terms of feature extraction quantity and receptive field.Then,the set abstraction(SA)module is improved to enhance the model's feature extraction capability for point cloud data.Two main improvements have been made to this module:1)Adding residual connections in each SA module.2)Adding inverted bottleneck design in the SA module.Based on the self built power inspection dataset,the test results show that the proposed method has achieved good performance in the semantic segmentation task of power line point clouds.Compared with the original network,the improved network has increased average accuracy by 5.1%and average intersection to union ratio by 7.2%.