How object detection works on security cameras.

What object detection is.

Object detection is a computer vision task: given an image, find every object of the classes you care about and say where each one is. On security cameras those classes are almost always people and vehicles, sometimes bikes, bags or animals. A detector is a neural network trained on large numbers of labeled images until it learns what a person looks like from above, from the side, walking, sitting, half in shadow or wearing a hood.

Detection is the foundation of modern video analytics. It is what lets an alert say “person at the loading dock” instead of “motion on camera 14.”

Bounding boxes.

For each object it finds, a detector outputs a rectangle, called a bounding box, drawn around it, plus a class label. The box is the model’s best estimate of where the object’s edges are. It is not a cut-out of the person and it is not exact; boxes drift a little from frame to frame, clip a foot, or swell when someone carries a large bag.

For monitoring, the box does three jobs. It tells the guard where to look. It lets the system crop the object for a closer view and for evidence. And it gives tracking something to follow from one frame to the next. You can see real boxes on real footage in the recorded demo on our home page.

Confidence scores are not probabilities.

Every box comes with a confidence score between zero and one. It is tempting to read a 0.8 as “eighty-in-a-hundred chance this is a person.” It is not. A confidence score is the model’s internal ranking of how strongly the image matches what it learned. It is useful for comparing detections and for setting a threshold, but it is not calibrated to real-world odds, and the same score can mean different things on a sharp daytime view and a grainy night one.

In practice, systems pick a threshold: detections above it become candidates for an event, detections below it are ignored. Set it too low and a traffic cone becomes a person at 3 a.m. Set it too high and a person in the far corner of the lot never shows up. The right threshold depends on the camera, and it is one of the things a pilot on your own cameras should settle.

A confidence score says how sure the model feels, not how often it is right. Treat it as a ranking, and judge the system by its alerts on your own footage.

Tracking: from boxes to movement.

A detector looks at each frame on its own. It has no idea the person in this frame is the same one as in the last. Tracking fixes that by linking detections over time, usually by matching position, size and motion between frames, and gives each object a track ID. That is what turns forty boxes into one person walking across a lot, and it is what makes events like loitering possible to reason about at all.

Good tracking also handles short gaps. Detectors blink: a person turns, passes a pole, sits down, and for a frame or two nothing is found. Semaphor keeps numbered, clip-local track IDs and recovers through short detection gaps using pixel continuity (optical flow and color), with regional rechecks that keep seated or still people supported. Track IDs are per camera and do not identify anyone. There is no facial recognition.

Occlusion.

Occlusion is when something blocks part of the object: a car door, a shelf, a column, another person. Detectors are trained on partly hidden people, so they often still find them, but with lower confidence and less stable boxes. In a crowded lobby, two people standing close can merge into one box or swap track IDs. In a parking lot, a person crouched between two cars may be almost invisible to a high camera.

The answer is partly camera placement and partly software. Cameras aimed down a row find more than cameras aimed across it. Tracking that holds through gaps keeps a person’s event going while they pass behind a pillar.

Night, weather and distance.

Night is where detection gets hardest. IR illuminators turn the image grayscale and wash out anything close to the lens; headlights blow out whole regions; low light makes cameras drop shutter speed so moving people smear. Rain, snow, fog and insects on the dome add clutter. Distance matters too: a person who is only a handful of pixels tall at the far fence line gives the detector very little to work with.

None of that is a reason to skip AI at night; it is where guards need help most. It is a reason to check your night views honestly and to fix what you can. Our guide to night-time camera monitoring covers lighting, IR and placement.

From detection to an alert.

Detection on its own would still flood a guard. A parked car is a vehicle all night. A receptionist is a person all day. Useful systems add rules on top: is this object new, has it moved, is this camera’s view even the same scene as a minute ago? Semaphor combines detection with pixel-change checks and event routing, including duplicate alert suppression and cooldowns, so the same parked car does not alert every few seconds.

Where Semaphor fits.

Semaphor runs person and vehicle detection locally on each camera pane it finds on your monitoring screen, using a YOLOX detector, with tracking and gap recovery built in. The alert publishes immediately, and a vision-language model adds a scene report to the same card with the original frame beside it. The guard reviews and decides. It all runs on top of the cameras and VMS you already have.

Read what Semaphor is, then book a live demo on your own camera screens or start a pilot. Terms like bounding box and track ID are in the glossary.

Frequently asked questions.