Point Cloud Classification Method for Transmission Towers based on CAA-PointNet

Abstract

In the filed of smart grid, the accurate classification of transmission towers is one of the hot research topics. However, to extract the different features of different towers in the process of classification is still a difficult task, in this paper a point cloud classification method for towers based on CAA-PointNet is proposed. Using PointNet as the basic framework, multi-scale local neighborhood is generated by sampling and grouping, and combined with the channel-wise affinity attention to enhance the the differential feature weight between the categories, so as to achieve accurate classification of towers. This method has good classification results for five different categories of tower point cloud data sets, with the overall accuracy of 95.0% and the average accuracy of 94.2%.