AI PTZ Auto-Tracking Cameras in Abu Dhabi

An AI PTZ auto-tracker runs two coupled loops. The first is an object detector — typically a person and vehicle model running on the camera's on-board NPU, often in the 1 to 4 TOPS range, sometimes higher when the manufacturer fits a more capable inference chip. The detector outputs a bounding box and a class label every frame. The second loop is the PTZ controller: it takes the bounding-box centroid, computes a pan, tilt, and zoom command to keep the centroid in the middle of the frame at a target pixel height, and sends it to the motor controller. The two loops have to be tightly coordinated so the camera is neither slow to react nor jittering. Sensor choice matters: a 1/1.8" BSI sensor with an F1.4 to F2.8 zoom lens keeps detection working at the long end of the zoom and at low light, whereas a smaller sensor on a long zoom delivers a noisy image the detector mis-classifies. Abu Dhabi specifics push the configuration: heat shimmer from afternoon tarmac creates apparent motion that fools a naive tracker; palm fronds and shade canopies in compounds cause repeated false-locks unless the trigger geometry excludes them; sand and dust on the dome glass reduces effective resolution. We address those with classification-gated triggers — the tracker only locks on a confirmed person or vehicle, not just motion — and with carefully placed trigger zones tied to ONVIF Profile S/T events for VMS integration. There is also the question of bitrate: H.265+ smart codec saves storage but its quality floor must be set so the bounding-box region keeps the edge detail the on-board detector relies on. A sound acceptance test measures three numbers — lock latency from target entry to centred zoom, classification confidence at maximum standoff, and oscillation rate when two targets are present. Those three numbers, recorded against a representative run, are what should be handed over.