Research On Airborne LiDAR Point Cloud Classification Method For Transmission Line

Airborne Li DAR(Light Detection and Ranging)system can directly and quickly acquire high-density and high-precision three-dimensional spatial information of target surface by integrating the laser scanning system,global positioning system and inertial navigation system.It has become an effective means in electric power inspections.The classification of transmission line point cloud is the process of extracting key elements(e.g.,power lines,towers,vegetation,and buildings,etc.)from the original airborne point cloud.It aims to provide accurate basic data for further applications,such as threedimensional reconstruction of transmission lines,digital management,operational inspections,and early warning analysis.However,the terrain of transmission line is complex and its point cloud data has distinctive attributes of unbalanced class distribution,large difference of data sources and so on.The exiting point cloud classification algorithms cannot meet the demands for accurate classification of transmission line point clouds.Based on the above background,this study carried out the research on the classification of transmission line based on airborne Li DAR data.It aims to migrate the artificial intelligence algorithms to point cloud classification of transmission lines,which further improve the algorithm accuracy and generalization ability and provide important technique supports for intelligent power inspection.The main contents of this paper are as follows:(1)The study on point cloud classification solution of transmission line based on traditional machine learning.This paper systematically compares the results of four classifiers(random forest,K nearest neighbor,logistic regression,and gradient boosting decision tree),two category distributions(balanced and unbalanced),three feature sets(constructed by random forest feature selection,principal component transformation,and the complete feature set),and various neighborhood radius.The results demonstrate that the algorithm with random forest classifier,original unbalanced class distribution and complete feature sets is chosen as the optimal classification solution.(2)The study on the transmission line point cloud classification based on deep learning.Deep learning algorithm is the further development of traditional machine learning.In this paper,the Point Net++ algorithm using the weighted cross-entropy loss function is applied to the transmission line point cloud classification.Compared with the best traditional machine learning classification scheme in terms of accuracy and efficiency,the average F value is increased by 11.15%,and the classification time cost is decreased by about 61 times.The results indicate that the Point Net++ algorithm is more accurate and lower time-consuming,which opens new horizons for large-scale transmission line point cloud classification.(3)The study on the transmission line point cloud classification based on transfer learning.A general point cloud classification method based on multi-level domain adaptation is proposed.This method imgrates the idea of adversarial training and domain adaptation in image processing to point cloud classification,and design the domain adaptive modules of point level and point set level to realize the alignment of different domain features,which can be applied to a variety of point-based point cloud deep learning algorithms.The results of the cross-terrain adaptation experiments show that the mean intersection over unions(m Io U)increase by 10.45%,indicating that the generalization ability of the model has been significantly improved.