ConvNeXtV2-EMA Model for Airport Runway Object Detection with Structure-Preserving Representation Fusion

Abstract

A Structure-Preserving ConvNeXtV2-EMA Model is Proposed for Airport Runway Object Detection Under Multiscene Visual Disturbance. Runway monitoring is very important for safety because in the same long-distance view, small foreign object debris, service vehicles, personnel, wildlife, lighting equipment, shadows, markings, wet reflections, and surface damage may all appear. The proposed framework normalises the runway images and creates a runway structural mask based on line direction, surface boundary, and marking continuity. Then, it combines this with the target-region candidates before they happen. ConvNeXtV2 backbone removes hierarchical features, and an effective multi-scale attention module refines object-like regions and eliminates runway textures false activations. A dataset of 13,600 images from daylight, night, rain-after, backlight, and long-distance monitoring scenes has been annotated with six target categories and safety-critical missed-detection labels. The proposed model has improved average precision to 91.3%, small-object recall to 86.9%, and safety-critical missed detections to 2.1% compared to Faster R-CNN, YOLOv5, Swin-based detection, ConvNeXt, and a non-EMA variant. Therefore, while runway geometry and focus-enhanced feature fusion may improve visual inspection reliability, real airport camera data, temporal verification, and operational records are still needed before deployment.