Detection & tracking

Detection#

The detector is tuned for the aerial case, which is meaningfully different from street-level vision. Targets are small (a car can be 25 pixels long), viewed top-down, and densely packed. Inference runs at high resolution, and confidence thresholds sit lower than you'd use on ground footage because a weak detection that tracking later confirms is worth keeping.

On top of the full-frame pass, the runtime runs localized re-detection: it crops the regions around active zones and locked targets, upscales them, and runs a second detection pass on the crop. Small and dim objects that the full-frame pass misses show up reliably in the upscaled crop. This is the main reason person tracking works at night.

Tracking#

Detections get associated frame to frame using motion prediction plus box overlap. Each confirmed association extends a track: one ID, one position history, one class. Association deliberately recovers low-confidence detections when they sit exactly where an existing track predicted they should be. That single trick roughly halves ID churn on aerial footage.

Stabilization#

Tracker output still isn't fit to show a human or feed to a rule. Boxes jitter, objects blink out for a few frames under a tree and come back, and the tracker occasionally hands out a fresh ID mid-journey. The stabilizer sits on top and enforces three things:

• An object must persist several frames before it exists at all.
• Positions are smoothed, so boxes don't vibrate.
• A brief dropout doesn't kill the object; it coasts on predicted motion and re-attaches.

Deduplication#

Duplicate boxes on one physical object are the most common source of overcounting, so the pipeline deduplicates in three places: after merging localized detections into the global set, before rule evaluation (boxes overlapping above 0.52 IOU collapse to one), and again before count aggregation. Redundant on paper. In practice each layer catches cases the others miss.