UAV Image-based Power Equipment Recognition by integrating Point Cloud Information

Aiming at the problems of significant structural differences between Unmanned Aerial Vehicles (UAV) images and Light Detection and Ranging (LiDAR) point cloud data, and significant differences in data recording information resulted overly independent processing of multi-source data in the fusion process, and insufficient utilization of complementary advantages of data, a multi-scale feature fusion deep learning network is constructed based on the idea of feature-level fusion. Its input is the characteristics of images and point clouds, and the output is the image features with fused point cloud information, which can be applied to pixel segmentation of power lines and power towers in UAV images. The test results show that the proposed feature-level fusion network structure can effectively extract deep features from point cloud data and integrate them into the image features, thereby improving the accuracy of power facility segmentation in UAV images. The True Positive Rate (TPR) of power line and tower pixels can reach 0.87 and 0.88, respectively. Meanwhile, the False Positive Rate (FPR) of power line and tower pixels has decreased to 0.16 and 0.14.