Object detection is the analytic that finds and locates objects in a video frame — drawing a box around each one and labelling it, for example "person" here and "vehicle" there. It is the foundation almost every other analytic builds on: tracking needs something to track, behavioural rules need to know a person crossed a line, search needs objects to index. Detection answers both "is it there?" and "where exactly?", which is what separates it from plain classification.
In surveillance, detection's biggest practical win is replacing crude motion detection. Old motion triggers fire on swaying trees, rain, and shadows; an object detector that only alerts on people or vehicles cuts those nuisance alarms dramatically — commonly by 80–95% — so operators chase real events, not weather. It is light enough to run on a camera's NPU (a few TOPS) and surfaces results as metadata over ONVIF Profile M.
Two caveats matter. First, accuracy is a range, not a guarantee: detection quality depends on resolution, lighting, angle, and occlusion, and is never perfect — plan around realistic figures for the actual scene. Second, detecting a "person" is not recognising who they are: plain person/vehicle detection is generally not biometric data, which keeps it on the right side of the privacy line that face recognition crosses. The detection model internals (the YOLO lineage and beyond) belong to the AI for Video Engineering section.
