AI Loitering Detection CCTV in Abu Dhabi

A loitering analytic chains three operations on the AI camera. The detector model returns bounding boxes with class labels and confidence per frame at four to ten FPS. A short-term tracker (IOU-based or a tracking transformer on a higher-spec camera) associates boxes across frames, giving each person in the scene a stable track ID for as long as they stay visible. The rule layer then watches whether a track ID’s footprint stays inside the polygon ROI for longer than the loitering dwell — usually 90 to 300 seconds in real operation, not the 30 seconds shown in vendor demos. Only then is an alarm raised, with the dwell counter as part of the metadata. Why this matters in Abu Dhabi: most after-hours loitering is genuinely innocent — drivers waiting on a fare, delivery riders checking a map, residents on the phone outside a building. A correctly tuned dwell respects that and only alarms on the durations that are actually anomalous for your venue. Sun glare, IR over-exposure on white compound walls, and reflective glass facades create plenty of nuisance trigger sources; sensor selection (starlight CMOS, true WDR) and IR balance fix more of those problems than the model ever will. Emirtech draws the polygon tight to the threat geometry, sets the dwell against the venue’s real pace, ties an arming schedule, tests the rule on site, and revisits after go-live with adjustments recorded.