Abductive identity of small aerial targets for refining neural network classifier labels

Abstract

Relevance. Optical channels of counter-drone systems distinguish small unmanned aerial vehicles from birds with neural network classifiers. At long range a target of a few pixels appears in isolated frames, and in a group of objects labels belong to different targets. Purpose of the article. To refine classifier labels on a small aerial target by abductive identity, without retraining. Objectives. To formalise target identity as an abductive choice and determine its effect on track maintenance, labels and class conclusion. Methods. In each frame either the silhouette near the prediction is the target or the target is not visible. An integrity constraint refutes a silhouette deviating from the prediction by over two body lengths, the measured margin of the true target; the minimal explanation is chosen. Labels of one target and the flapping fact feed an abductive logic program. Scientific novelty. Identity rests on a physical motion constraint scaled by the measured body length, not on appearance, which tiny silhouettes lack. Practical significance. The same classifier held on one target gives more correct labels and explained conclusions. Results. For vehicles of nine to twelve pixels frames off the target fell by one quarter to one half without harming bird tracks. The attention window raised correct vehicle labels from one fifth – one third to one third – three fifths. The abductive program removed wrong vehicle conclusions at the cost of fewer decisions; six-pixel vehicles are beyond the method. Conclusions. Maintaining the track of one target is a precondition for refining classifier labels; the main limitation is frequent “unknown” answers at low confidence.