Target lock
Click an object once. The runtime owns it from there: through camera motion, through brief occlusion, through the frames where the detector sees nothing at all.
Zones express “tell me about anything that happens here.” Target lock expresses the other primitive operators actually use: “this one. Follow this one.” Between them they cover most of aerial overwatch.
What a lock is#
A lock binds three things to the clicked object: its current track, an appearance template sampled at lock time, and a motion model. The lock manager runs them as a voting system. While the detector and tracker keep producing the object, the lock simply rides the track. When they don't, the template and motion model carry it.
Localized re-detection#
The lock also changes how detection itself runs. Each frame, the runtime crops a region around the lock's predicted position, upscales it, and runs a dedicated detection pass on the crop. An object too small or too dim for the full-frame pass is usually obvious in the upscaled crop. This is what makes clicking a person at night work: the full frame shows a smear, the crop shows a person.
Losing honestly#
Every lock carries a confidence that decays when the object goes unobserved and recovers when it's re-confirmed. Below threshold, the lock declares itself lost and freezes at its last confident position instead of wandering onto whatever nearby object resembles the template. Reacquisition then works outward from that position. As with zones, the rule is that the system tells you it lost the target; it never quietly substitutes a different one.
Multiple locks#
Locks are cheap enough to run several at once. Each gets its own crop, template, and confidence. The practical limit is operator attention, not compute.