What is video analytics?

The plain-language definition.

Video analytics, also called video content analysis, is the automatic analysis of video to detect events in space and time. That is close to the textbook wording on Wikipedia’s video content analysis entry, and it holds up well for security work. The camera records light. Analytics decides what that light means: a person, a car, a crowd forming, something left behind, something that has not moved in hours.

For a guard or operator, analytics is the difference between a screen that waits to be watched and a screen that tells you where to look. It is also only one layer of the full workflow. Detection and rules produce events; AI video monitoring adds reasoning about the scene and routes the result to a person who decides what to do.

What video analytics actually does.

Motion detection

The oldest form. Compare each frame against the last, or against a learned background, and flag areas where enough pixels changed. Cheap and fast, and noisy: headlights, rain, shadows, swaying trees and camera exposure shifts all count as motion. It is the reason many sites turned their alerts off. The difference is laid out in AI video analytics vs motion detection.

Object detection

A trained neural network looks at a frame and says "person here, vehicle there," with a box and a confidence score for each. This is where modern analytics earns its keep, because a guard cares about people and vehicles, not pixels. We cover the mechanics in how object detection works on security cameras.

Tracking

Detection answers "what is in this frame." Tracking answers "is that the same person as a second ago." It links detections over time so a single walk across a parking lot becomes one event with a path, not forty separate alerts.

Rules and events

Rules turn tracked objects into events a site cares about: a person in a zone after hours, a vehicle stopped at a gate, someone lingering near a door. Rules are where site knowledge lives, and where most of the tuning effort goes. Specific patterns like loitering detection and after-hours trespassing are rule problems as much as detection problems.

Specialist analytics

License plate recognition, face recognition, crowd counting and specialist hazard models all sit in the analytics family too. Each carries its own accuracy, privacy and policy questions. Semaphor deliberately does not do facial recognition; it detects and tracks people without identifying them.

Where analytics runs.

Analytics can run in three places. On the camera, using a chip inside the device; convenient, but limited to that camera model and its firmware. On a server or appliance at the site, usually tied to the VMS or recorder. In the cloud, where video is streamed out for processing. Each has trade-offs in cost, bandwidth, privacy and what it takes to install; edge vs cloud video analytics goes through them.

There is also a fourth option that most buyers have not seen: analytics that reads the monitoring screen itself. Instead of connecting to each camera or VMS, the software captures what the guard already sees and analyzes each camera pane. That is how Semaphor works, and it is why it can run across mixed camera brands and different client VMS platforms without an integration project. The comparison is in screen-based vs VMS-integrated analytics.

Why input quality decides everything.

Analytics is only as good as the video it sees. Dirty domes, low bitrate substreams, cameras aimed into the sun, IR washing out faces at night, a lens fogged by a cold morning: all of it shows up as missed detections or strange ones. The Wikipedia entry makes the same point, noting that the quality of commercial analytics is hard to pin down because it depends on the use case, implementation, configuration and computing platform.

That has two practical consequences. First, any vendor figure measured on someone else’s footage tells you little about your sites. Second, the fastest way to improve analytics is often boring: clean the lens, fix the angle, check what the camera does at dusk. Our guide on night-time camera monitoring goes into the lighting side.

Analytics produces events. It does not produce judgment. Every event still needs someone who knows the site to decide whether it matters.

How to evaluate video analytics.

• Run it on your own cameras. Include your worst views: the backlit entrance, the far corner of the lot, the stairwell with one bulb.
• Count the noise. How many alerts per camera-night are things nobody needed to see? How are repeat alerts about the same parked car handled?
• Count the misses. Walk the site at night and check what did and did not alert.
• Look at the evidence. Does every event keep the original frame, or only a thumbnail and a label?
• Price the install. New cameras, new servers, per-camera licensing, VMS plug-ins and integration hours add up fast.

Where Semaphor fits.

Semaphor packages video analytics into a workflow built for guards. It reads the monitoring screen, separates each camera pane and runs person and vehicle detection locally with tracking that holds through short gaps. Duplicate suppression and cooldowns stop the same parked car from alerting all night. A preliminary alert goes out immediately, and an AI scene report follows on the same card with the original evidence beside it. The guard decides what happens next.

It all runs on top of the cameras and software you already own. Read what Semaphor is, then book a live demo on your own camera screens or start a pilot with one monitoring station. Unfamiliar terms are in the glossary.

Frequently asked questions.