A Pose Estimation Framework for Multitype Transmission Electrical Towers Based on Elementalizing Substructure Strategy and Part Affinity Fields (PAFs)

The integration of Unmanned Aerial Vehicles (UAVs) and deep learning technologies has become increasingly prevalent in the inspection of transmission lines. However, existing methods often overlook the positional information of images within the overall structure of transmission towers. Manually annotating such positional information is time-consuming and labor-intensive, limiting the parallel inspection of multiple lines and regions, thereby reducing overall inspection efficiency. This paper proposes a pose estimation framework for multi-type transmission electrical towers based on elementalizing substructure strategy and part affinity fields (PAFs) to accurately estimate the poses of various transmission towers. Specifically, it decomposes the substructures of 5 types of transmission towers into 14 keypoints and 37 skeletons, offering two significant advantages. Firstly, it elementalizes hundreds of keypoints from five types of transmission towers into 14 types, providing the encoding method with greater flexibility and generalization capabilities, and supporting few-shot and even zero-shot learning. Secondly, it leverages PAFs to predict skeleton categories, positions, and directions, fully exploiting image features to enhance keypoint prediction accuracy and instantiate keypoints and skeletons. Comprehensive experimental results on a high-resolution image dataset collected from the State Grid demonstrate that the proposed framework achieves precise pose estimation for various transmission towers. Additionally, the framework exhibits excellent zero-shot and few-shot learning capabilities, validating its effectiveness and strong generalization performance. The framework facilitates UAVs in understanding tower layouts and autonomously annotating inspection images, enabling more efficient and automated transmission line inspections.