From Object Detection to Conflict Assessment: A Safety-Oriented Uncertainty Framework for Monocular UAS Detect-and-Avoid

Monocular cameras are attractive sensors for Detect-and-Avoid (DAA) on lightweight unmanned aircraft systems (UASs) because they are passive, low-cost, easy to integrate, and can support the detection and tracking of non-cooperative or electronically non-conspicuous airborne objects. Their safety-relevant performance, however, cannot be inferred from object-detection metrics alone because errors propagate through the perception chain: image acquisition and formation, target detection and classification, bounding-box localization, physical-size prior selection, monocular ranging, temporal tracking, and closing-speed estimation. This paper develops a safety-oriented uncertainty framework for monocular airborne conflict detection as a residual onboard layer in a layered UAS conflict-risk mitigation architecture. The framework distinguishes uncertainty arising from detection failures, class confusion, physical-size priors, target pose and aspect angle, bounding-box localization, temporal association, and camera and image-formation characteristics. These include sensor resolution and distance-dependent loss of projected target detail, intrinsic and extrinsic calibration, optical distortion and off-axis degradation, exposure time and frame rate, motion blur, sensor noise, compression, contrast loss, and background clutter. A large purpose-built labeled air-target image dataset, supplemented by suitable public datasets, is used to estimate component-level error distributions across target classes and operating conditions. These distributions are propagated through a pinhole-based ranging model and temporal closing-speed estimators to quantify total system error as a function of distance, relative geometry, target class, image quality, and camera configuration. The framework is evaluated using a curated flight-test dataset with RTK-derived ground-truth separation between observing and intruding aerial platforms. The analysis yields class- and distance-dependent uncertainty budgets, safety margins, and operating limits for UAVs, light and transport aircraft, helicopters, gliders, balloons and airships, birds, and kites. By identifying the dominant error sources across relevant regimes, the framework supports evidence-based assessment of monocular sensing for low-altitude UAS conflict detection.