Vision-Based Inspection of Airfield Ground Lights: Detection, Colour Classification, and Edge-Computing Performance

Abstract

Airfield ground lighting is safety-critical electrical infrastructure and requires reliable inspection. We benchmarked lamp detection, colour assignment and embedded execution using 59 GoPro HERO 8 recordings acquired through scientific cooperation with Poznań-awica Airport. The dataset contains 3612 frames and 18,141 instances in four colour classes; training-only augmentation produced 8668 images. Four YOLO variants and three classical pipelines were compared on a 362-image test subset using region-of-interest filtering and greedy one-to-one matching at an intersection-over-union threshold of 0.45. At 640 by 480 pixels, YOLO26 achieved the highest observed F1 score (0.8080) at 64.0 frames per second, followed by YOLOv12 (0.7178) at 53.8 FPS. Brightness thresholding reached 622.4 FPS with F1 score = 0.3057. Raspberry Pi 5 CPU execution reached 1.1–1.3 FPS; Hailo-8 reached 48.6 and 36.4 FPS for YOLOv8 and YOLO11, respectively, with a reduced F1 score after conversion. FPS includes working-image preparation, detector execution and post-processing, but excludes video decoding and rendering. These fixed-split measurements quantify the trade-off between detection quality, colour coverage and headless throughput for human-supervised inspection and edge-platform selection.