Enabling monocular collision avoidance in a smart hangar

Abstract

Abstract Hangar-based maintenance, repair and overhaul (MRO) operations expose aircraft to close-proximity interactions with ground support equipment (GSE), where operator errors can cause structural damage and disruption. This study presents a monocular vision framework for collision-risk monitoring in a smart hangar, coupling instance segmentation, domain-adapted monocular depth inference and deterministic clearance reasoning. A paired dataset comprises synthetic top-down hangar renderings with pixel-aligned red-green-blue (RGB)-depth supervision and a matched miniature testbed containing scaled aircraft/GSE models and 3D-printed calibration objects, imaged by custom-mounted RGB cameras. The dual-head pipeline comprises YOLOv8-seg for six-class instance masks and DepthAnything-V2 for metric depth under the trained model convention. On the scene-disjoint synthetic test split, the custom-trained depth model achieves 0.135 m MAE, compared with 8.385m for pretrained weights, without materially increasing depth-inference time. WIthout dense RBG-camera depth ground truth, testbed-image transfer is reported as an ordinal/structural calibration-object proxy: relative step ordering is often preserved, whereas inter-step metric spacing is compressed and deemed inadequate for direct clearance-threshold decisions. Custom YOLOv8-seg variants achieve mean Intersection over Union (IoU) of 0.88–0.92 across synthetic and testbed splits, while the off-the-shelf baseline yields near-zero overlap in transfer tests. Collision inference is evaluated on deterministic 30-image samples per synthetic/testbed scenario, combining depth-derived height envelopes and boundary-only ground-plane distances to produce clear/warning/alert states and overlays. Median no-render compute latency is 0.736–1.037 s; profiling identifies post-processing, particularly GSE clustering and geometry construction, as the dominant runtime component. The study provides an auditable perception-to-geometry pipeline and highlights the need for uncertainty-aware threshold calibration before deployment.