Detection & tracking
Mission Studio separates detections on sampled frames from the track histories used by area rules, subject following, and retained evidence.
Hosted detection#
The public worker samples the normalized source at a configured cadence and runs YOLO11 for the selected subject family: people, vehicles, people and vehicles, bicycles, boats, or animals. Each retained detection keeps its source frame, timestamp, label, confidence, and normalized box.
Hosted identity linking#
Detections are associated in time using predicted center motion, spatial distance, and box overlap. A history needs at least three observations before Mission Studio treats it as a stable track for rules and result rendering. The retained manifest keeps every observation with the stable track ID used by later stages.
Short gaps#
The tracker keeps recent motion state through a short sampling gap and uses that prediction when the subject appears again. It does not claim biometric identity or re-identification across long gaps. If the association cannot be supported, the later observation can receive a new track ID.
Broader runtime design#
The local Vision runtime design adds localized re-detection around active zones and selected targets. It can upscale a crop, run a second detection pass, and combine that evidence with a target template and motion model. Those target-lock behaviors are not part of the hosted Mission Studio worker.
Separate Challenge logic#
The Challenge replays checked-in detections, tracks, zone geometry, and scored entries. Its overlap and track-switch corrections belong to that prepared record and are not the hosted Mission Studio implementation.