Correlated Verifier Errors in Counter-UAS Perception

Clustering, Cross-Sensor Dependence, and Label-Free Auditing

Abstract

Counter-UAS perception systems re-query classifiers on the same track and fuse visible and infrared evidence as if each query and each sensor were independent. We test this assumption with random-effects and latent class models from diagnostic-test evaluation, on two public drone recording sources (Halmstad and Anti-UAV RGB-T) plus AVID, a paired visible–infrared subset of the Halmstad recordings that adds bird negatives, five model families and three detector architectures. Errors cluster strongly within tracks (track-level random-effect standard deviations from 0.7 to above 5 on the logit scale), and false alarms cluster as strongly as misses. With out-of-session predictions for all 315 Anti-UAV RGB-T sequences in 37 recording sessions, recording conditions account for about two thirds of infrared clustering and half of visible clustering. For trained detectors, infrared and visible misses co-occur on the same frames far more often than independence predicts (odds ratio 6.26 across 315 sequences), so a second sensor is not independent redundancy. As a consequence, a standard label-free (Dawid–Skene) audit overstated detector sensitivity by 8 to 18 points on real votes and misestimated accuracy by up to 57 points when bird negatives were available. We show that dividing summed frame evidence by Kish’s design effect is the exact Bayes rule under Gaussian, calibrated, equally correlated evidence; on real data it cut track-level overconfidence from 29 to 10 points without changing accuracy, and error dependence decays with time in sensor-specific ways. Counter-UAS systems sho