UPKD: Unsupervised pylon keypoint detection from 3D LiDAR data for autonomous UAV power inspection

Abstract

The automatic extraction of inspection points for pylons is essential for intelligent Unmanned Aerial Vehicle (UAV) power inspections, especially in generating inspection route. However, current researches primarily focus on power scene perception and power component fault detection, with little attention paid to the inspection point detection. The primary reason are the lack of public pylon datasets for keypoint detection and the neglect of distinguish the inspection points from the keypoints used in feature extraction. Regarding that, this paper proposes a two-stage unsupervised pylon keypoint detection (UPKD) method to improve the efficiency of power inspections. In the first stage, the UPKD Network processes the point cloud to generate candidate keypoints, which comprises two main components: a data normalization module and an unsupervised keypoint detection network (UKD-Net). The data normalization module compresses information based on the symmetric structure of pylons, thereby reducing instability in inspection point detection. The UKD-Net incorporates a Point Transformer layer that uses self-attention mechanisms to extract features from the point cloud. In the second stage, a convex optimization strategy is applied to filter and acquire inspection points. These inspection points are then interconnected using a shortest-path strategy to generate the UAV inspection route. Our dataset is obtained using the Riegl VUX-1 laser measurement system and comprises 3,296 pylons of 10 types. Each pylon’s point cloud contains up to 25,000 points, with a high point density of 100 pts/m2. Extensive experiments show that the UPKD Network achieved state-of-the-art performance on our dataset, with repeatability achieving 90.39%, effectiveness (the ratio of effective keypoints to annotation points) achieving 69.95%, and completeness (the ratio of detected annotation points to keypoints) achieving 79.55%.