Three-dimensional information extraction method of wire end for live-line working robots in strong illumination environments

Abstract

Live-line working robots require the 3D information of wire ends to implement grasping and connection tasks, whereas overexposed images and incomplete point clouds under strong illumination degrade extraction accuracy severely. In this paper, field-collected images and point clouds are processed via systematic data engineering to build a multimodal dataset with the help of adaptive Gamma correction. DCGAN is adopted for strong-light image restoration, and Grounded-SAM is used to segment wire-end regions. A point cloud completion algorithm integrating global feature perception and U-Net refinement is proposed to repair defective point clouds. With calibrated 2D-to-3D mapping, the three-dimensional coordinates of wire ends are finally extracted. Practical tests demonstrate that the proposed approach accurately obtains wire-end 3D data under intense illumination and satisfies the on-site operation requirements of robots.