UAV Inspection and Collision-Aware Visual Perception Using Topology-Preserving RGB–Infrared Power-Line Graph Reconstruction

Abstract

Accurate perception of overhead conductors is fundamental to autonomous unmanned aerial vehicle (UAV) inspection, trajectory planning, and collision avoidance in transmission corridors. Existing vision-based approaches predominantly formulate power-line detection as a pixel-wise segmentation task, despite the fact that operational safety depends strongly on the structural integrity of the detected conductors. This study proposes an RGB–infrared perception framework that reconstructs segmented power lines as topological graphs for collision-aware UAV operation. Modality-specific encoders are employed to extract complementary features from visible and infrared imagery, followed by an adaptive fusion mechanism for multimodal feature integration. A topology-aware learning objective is introduced to promote conductor continuity and mitigate fragmentation. The predicted masks are subsequently skeletonized and converted into graph representations comprising endpoints, junctions, edges, and connected components. Skeleton intrusion into a central image strip and normalized distance to the image centerline are evaluated as collision-aware visual indicators. Experiments are conducted on the public Ground Truth of Powerline Dataset (GTPLD) using a balanced subset of line-present and line-absent images. In addition to Dice and intersection over union, the evaluation incorporates skeleton F1, component error, edge-count-proxy error, image-level intrusion F1, and normalized centerline-distance mean absolute error to quantify structural fidelity and safety relevance. The graph-oriented model achieves a skeleton F1 of 0.6549, a component error of 1.500, a corridor F1 of 0.8966, and a safety-margin mean absolute error of 0.120. Although the pixel-overlap baseline attains a higher Dice score, it yields more fragmented and less stable graph structures. These results demonstrate that pixel-level overlap alone is inadequate for evaluating UAV-oriented power-line perception and highlight the importance of topology preservation and safety-corridor accuracy.