A. Rouhi and D. K. Han
Aethon Perception, Pennsylvania, United States
Keywords: counter-UAS, detect-and-avoid, test and evaluation, drone detection, synthetic data validation
Long-range aerial perception fails on small, distant, low-contrast objects, intermittent detections, and false tracks from birds and clutter. We present a detection-and-validation architecture in which a temporal EO/IR detector/tracker is developed within a frozen, difficulty-resolved evidence harness separate from training. Our preregistered three-seed study found synthetic world-model data increased held-out medium-object recall from 0.122 to 0.206 (+69%), yet all arms scored zero below ~24 pixels across 1,237 real instances. Replication prevented two false capability conclusions. The result defines the gap: a frame detection is not a trustworthy operational track. We are extending this work into a perception layer that accumulates weak motion/contrast evidence across frames, maintains and reacquires tracks after missed detections, and reports staged identity, uncertainty, and track quality through an API. Each supported configuration would be assessed for warning time, continuity, false tracks per sensor-hour, and edge latency on held-out, rights-cleared data. This is vendor acceptance evidence, not third-party certification. Current evidence establishes the evaluation method and single-frame boundary; tracking and live metrics remain unproven. We seek integrator and government design partners offering lawful test access, a baseline, and a path to paid evaluation. Authorized response integration is a future partner option; interceptor development is outside scope.