A lightweight real-time accurate deformation detection approach for pantograph horns

Abstract

The pantograph is vital for urban rail transport, with the horn deformation directly impacting safety. In this paper, a novel method for detecting horn deformation is proposed. First, to determine the horn pose, keypoints on horn surfaces are extracted using a lightweight, real-time, and accurate network. A novel module is introduced to improve the performance of post-processing. Second, view transformation of horn keypoints is proposed. It calculates the three-dimensional angles of the horns in monocular images based on the vanishing points. Finally, the experimental results demonstrate a sound performance of the proposed horn keypoint detection network, achieving 97.73% accuracy and 2.05% false positive rate with an inference time of 3.8ms and a parameter count of 3.41M. Under different views, the converted horn angle can account for 95.75% under the error of 1°and 100% under the error of 1.5°. Therefore, the proposed method reaches the level of practical industrial application.