| Airborne laser radar measurement system focuses on laser ranging technology,computer science technology,high dynamic attitude determination technology and high-precision GNSS(Global Navigation Satellite System)dynamic positioning technology,which can obtain multi-source data fusion of 3D point cloud data acquisition tool.Because it has some advantages of high efficiency,high precision and low cost,it has been widely used in the current inspection of power line corridor.Data processing is an important link of power line inspection after the collection of point cloud data,and the classification of point cloud in power line corridor is the core of data processing.Point cloud classification in power line corridor is the process of extracting key elements such as wires,ground wires,towers,vegetation and buildings from the collected point cloud data,which provides data basis for the subsequent dangerous point detection,3D reconstruction and intelligent management in power line corridor.However,the complex terrain,unbalanced density and irregularity of point cloud in power line corridor cause low automation,accuracy and efficiency of conventional point cloud classification methods.Deep learning technology has great potential in the field of point cloud classification due to its high computational efficiency and ability to extract features independently.Aiming at the existing problems in point cloud classification,this paper realizes the point cloud classification of power line corridor on the basis of deep learning technology.However,the lack of suitable deep learning data sets is an important factor that restricts the popularization and application of this technology in practical applications.According to this,this paper builds a set of POWERLINE-ALS data set by using power line corridor point cloud collected in Jiangsu Province,and then studies the classification of power line corridor point cloud based on this data set,so as to applys deep learning technology to the classification of power line corridor point cloud,and improves the accuracy and automation of point cloud classification,and at the same time,provides an important technical means for intelligent power.The principal research contents and results of this paper are as follows:(1)Based on the airborne laser radar measurement system,five flight platforms(manned helicopter,self-rotor aircraft,unmanned helicopter,unmanned fixed-wing,unmanned multi-rotor)and four positioning methods(Qianxun CORS,provincial CORS,dynamic post-processing differential base station positioning,precise point positioning)are adopted to collect the data of power line corridor in a place in Jiangsu Province.Through experiments on acquisition strategies of different flight platforms and different positioning modes,the results show that the acquisition strategy based on the unmanned helicopter flight platform and the base station differential positioning mode has the best effect.And,the cloth simulation filtering algorithm is used to filter out the ground points,and 32.46 million non ground points are manually marked,including 54 training sets and11 test sets.The categories are divided into six categories: conductor,ground wire,tower,vegetation,building and low power line.So that the power line corridor point cloud data set POWERLINE-ALS is constructed.(2)The point cloud data set of power line corridor is classified according to the original Point Net++ algorithm.Because of the worse classification effect,the Point Net++algorithm is improved to transmit the point cloud data of single file into the network,use ball query in the bottom network,and use k-nearest neighbor to obtain neighborhood points in the high-level network.The overall accuracy OA of the original Point Net + +network is 89.59%,the overall accuracy OA of the improved Point Net++ network is92.26%,the m Io U of the original Point Net++ network is 53.04%,and the m Io U of improved Point Net++ can reach 73.82%.The accuracy of improved network m Io U is 39%higher than that of original Point Net++.In terms of efficiency,the test data set of the original pointnet++ algorithm takes 4691.37 ms,while the time of the improved pointnet++algorithm increases by 650 ms,reaching 4052.83 ms,and the efficiency increases by about16%.On the one hand,the improved network classification accuracy is greatly improved,on the other hand,it also apply to the point cloud classification of power line corridor constructed in this paper.(3)The data set of power line corridor is classified based on Rand LA-Net algorithm,and the classification effect of Rand LA-Net on POWERLINE-ALS data sets is tested for different network layers.Among them,the four-layer Rand LA-Net network structure m Io U is 59.5 %,and the five-layer Rand LA-Net network structure m Io U is 82.25 %.The accuracy of the five-layer network structure m Io U is 38 % higher than that of the four-layer network.In terms of efficiency,the five-layer Rand LA network structure takes784.30 ms,and Rand LA-Net is five times higher than the improved pointnet++. |
