Skip to content
Camzify
Now shipping · AI detection

Behavioral anomaly detection

Describe the behavior you want watched, in your own words, and Camzify monitors for it. Type “alert me if anyone starts fighting” or “tell me if someone is smoking in the loading bay”. Natural-language processing interprets what you have asked for, the system watches the people appearing in the cameras you selected, and a notification is raised when that activity is observed.

  • Describe the behavior in plain language
  • No rule syntax, no zones to draw
  • Watches the cameras you pick for it
People moving through a public space with an AI overlay tracing behavior patterns

Every other detection model on the platform is built for one job decided in advance. This one is defined at the point of use, which means you are not restricted to the behaviors someone anticipated when the catalog was written.

Describe it the way you would say it

There is no rule syntax and no zone geometry to draw. These are the kinds of descriptions operators actually write, and what each one puts the system on watch for.

Alert me if anyone starts fighting

Watches for
Physical altercation between two or more people

Tell me if someone is smoking in the loading bay

Watches for
Smoking in a zone where it is prohibited

Notify me about vandalism or damage to property

Watches for
Deliberate damage to fixtures, vehicles or signage

Watch for anyone climbing the fence

Watches for
Trespassing over a boundary rather than through a gate

Flag people hanging around the entrance after closing

Watches for
Sustained presence with no apparent purpose

Let me know if someone is tampering with the cameras

Watches for
Interference with equipment

How it works

Interpreting the description

Natural-language processing turns what you typed into what the system should watch for. Two differently-worded descriptions of the same behavior resolve to the same monitoring, so you do not have to learn a phrasing convention.

Watching the subjects

The people appearing in the selected views are tracked using multi-object tracking, so a subject keeps a persistent identity across frames while their behavior is assessed against the description rather than judged from a single frame.

Raising the notification

When the described activity is observed, an alert enters the notification queue naming the camera, the time and which description it matched, routed to the contact assigned to that camera.

Why this matters operationally

Every site has behaviors that matter locally and appear on no vendor’s feature list. Smoking beside a fuel store. People climbing on stacked pallets. Someone propping a fire door. These are obvious to whoever runs the site and invisible to a fixed model catalog.

Historically the only options were to accept the gap or commission a custom model. Describing the behavior in a sentence removes that trade-off, and it means the system can be adjusted by the person who understands the site rather than by the vendor.

In a patrol round

Behavioral alerts run continuously rather than only at patrol time, but they feed the same record. Anything flagged between rounds is logged against the relevant camera in the next patrol report, so the round reflects what happened while nobody was checking rather than only what was true at the moment of the check.

FAQ

Frequently asked questions

Behavioral anomaly detection lets an operator describe, in ordinary language, a behavior they want monitored — for example "alert me if anyone starts fighting" or "tell me if someone is smoking near the loading bay". Camzify interprets that description, watches the people appearing in the selected camera views, and raises a notification when the described activity is observed. It differs from a conventional detection model because the behavior is defined by the operator in words rather than chosen from a fixed list.

The other models are purpose-built for one thing each: a line-crossing model watches a tripwire, a PPE model watches for helmets and vests. Behavioral anomaly detection is defined at the point of use. You describe the behavior and the system interprets it, so you are not limited to the behaviors someone anticipated when the model catalog was built.

Behaviors that are visible in the camera view and describable in a sentence. Fighting, smoking, vandalism, climbing a fence, loitering near an entrance and interfering with equipment are all typical. Behaviors that depend on information the camera cannot see — intent, identity, or anything happening off-frame — are outside what any video system can determine.

No. The input is ordinary language. Natural-language processing interprets the description into what the system should watch for, so "alert me if people start fighting" and "notify me about physical altercations" resolve to the same monitoring behavior. Being specific about the location or time window narrows it usefully.

It runs continuously on the cameras you activate it for, and notifies in real time when the described behavior is observed. It also contributes to virtual patrolling: anything it flagged between rounds is logged against the relevant camera in that round's report, so the patrol record reflects what happened while nobody was checking.

Through the same queue as every other detection, with severity, site, camera and a timestamp, routed to the contact assigned to that camera. It carries an acknowledgment state and can be marked a false positive, which records the system's behavior on that camera and that description.

Yes. A camera can carry several descriptions at once — a loading bay might watch for both smoking and vandalism — and each is evaluated independently, so a notification names which description it matched.

Ready to patrol your site 24/7?

Book a 15-minute demo and see a live patrol run on your own cameras.

This site is being updated

We are rebuilding pages as you read this, so an image, a link or a section may look unfinished for a while. The product itself is unaffected. If something important is broken, tell us at the contact page and we will fix it.