Skip to content
AI analytics

AI video analytics on the cameras you already own

Habile iEYE counts people and vehicles, notices tampering and after-hours presence, and asks a vision model you host about PPE, falls, fire and smoke and blocked areas. Every alert says which rule fired and keeps the image that proved it.

Last updated

Available in licensed installations

Available in licensed installations (self-hosted appliance, Windows Server, multi-server). Each AI use case is a licence feature, with a limit on how many cameras may run one.

Not yet on AWS Marketplace

Not yet in the AWS Marketplace appliance: as of version 4.4.0 it does no detection of its own, and its alerts come from the alarms your cameras raise.

No accuracy figures yet

We do not publish accuracy figures yet. The use cases were checked on 28 October 2026 against open models on Amazon Bedrock, a replayed recording and a live camera stream, not yet on a customer's site footage. We tune and measure them on your footage during a pilot.

Detection and counting

What the server detects without a vision model

These run in the inference service beside the recorder, on the camera's own stream.

  • Detector

    Motion and intrusion

    Motion with a sensitivity and a smallest area, judged per region you draw, with lighting changes absorbed. Intrusion is motion inside a region you mark as off limits.

    Raises: A motion or intrusion alert with a snapshot

  • Detector

    Camera tampering

    Frozen, covered, very dark, overexposed or blurred pictures and a changed scene, each check switchable per camera.

    Raises: A tamper alert naming the check that fired

  • Detector

    People and vehicle counting

    People and vehicles are tracked across the picture and counted once as they cross a line you draw, in and out, with the number in view sampled over time. Hourly counts are kept for 400 days.

    Raises: Counts and reports, and a crowd alert above a threshold you set

  • Detector

    After-hours presence

    Anyone present outside the hours you set for a camera raises an alert, with a cooldown per camera.

    Raises: An after-hours alert

Vision-model skills

Questions a detector cannot answer

A skill sends a crop of a person, a named area or the whole frame to a vision model with a question, and gets an answer in a fixed structure that rules can act on.

  • Detector + vision model

    PPE compliance

    Checks each tracked person for the protective equipment your site requires, such as a helmet.

    Raises: An alert naming what is missing

  • Detector + vision model

    Dress code

    A company dress code, or a hospital staff code: ID card, uniform and footwear. Chosen per camera.

    Raises: An event or alert per item that fails

  • Vision model

    Falls

    Asks the model whether someone in the frame has fallen or collapsed.

    Raises: A fall alert with the frame as evidence

  • Vision model

    Fire and smoke

    Asks the model about the whole frame on a timer for visible fire or smoke.

    Raises: A fire or smoke alert

  • Detector + vision model

    Odd-hour movement

    When someone is present after hours, the model describes how they are moving, and the description goes into the alert.

    Raises: An after-hours alert that says what the person was doing

  • Vision model

    Occupancy timer

    Asks about each area you name on a camera, on a timer: goods left in a bay, or a passage that stays blocked. The timer survives a restart.

    Raises: An alert when an area stays occupied longer than you allow

How an alert is made

From a frame to an alert that explains itself

  1. The camera's stream is decoded once

    The inference service reads each camera once and shares the frames between every use case switched on for it.

  2. A detector finds and tracks

    People and vehicles are detected and tracked from frame to frame, so a person is one person, not thirty detections a second.

  3. A vision model answers a question

    For a vision skill, a crop of the person, a region or the frame goes to the model with a prompt, and the answer comes back in a fixed structure.

  4. Rules decide

    Rules on the answer decide between nothing, an event and an alert, with a cooldown so one situation is one alert.

  5. The alert says why

    Each alert keeps the rule, the use case and the evidence image, and goes to the notification rules like any other.

Where the model runs

Your cameras' images go to a model you choose

The vision skills call a model you choose. By default a provider outside your network is refused, so an air-gapped site cannot send camera images out by mistake.

A model server you host

Any OpenAI-compatible server, such as vLLM or Ollama, on your own GPU. Camera images stay on your network.

Amazon Bedrock

A model in your own AWS account and Region, switched on explicitly by a platform administrator.

Kept under control

  • Provider secrets are stored sealed and shown by name only
  • Each organisation can be given its own provider and model for each use case
  • The evidence image behind each answer is kept with the alert for 30 days
Make it yours

Write, test and measure your own checks

  • Custom skills

    Write your own: a prompt, the shape of the answer, and rules that turn answers into alerts or events. Every save keeps a version.

  • Test on a camera

    Try a use case or your own skill on a live camera before switching it on. Nothing from a test is stored.

  • Count accuracy checks

    Replay a stretch of a recording through a camera's counting lines and score the result against your own hand count.

  • Budgets and statistics

    A calls-per-minute budget for each organisation, and run statistics per model on the health page.

Not yet released

Built, and waiting for a release

These are in the code and described in the changelog as unreleased. They are listed so you know what is coming, not as something you can use today.

  • Face recognition

    Known and unknown faces against a face gallery per organisation, with in and out across a camera's counting lines. Needs a separate face service on a GPU server.

    Built, not yet in a release.

  • Editing the built-in skills

    Changing the prompt, minimum confidence, cooldown and rules of a built-in skill per organisation, running a custom skill on each drawn region, and choosing an alert's severity.

    Built, not yet in a release.

Questions

About AI in Habile iEYE

Does Habile iEYE include AI video analytics?
Licensed installations on your own servers can run AI use cases: motion, intrusion and tamper detection, people and vehicle counting, after-hours presence, and vision-model skills for PPE, dress code, falls, fire and smoke and occupancy. The AWS Marketplace appliance does no detection of its own yet (version 4.3.0); its alerts come from the alarms your cameras raise.
Where do camera images go when a vision model checks them?
To a model you choose: an OpenAI-compatible server you host yourself, such as vLLM or Ollama, or Amazon Bedrock in your own AWS account. A provider outside your network is refused by default and has to be switched on explicitly by a platform administrator.
Get started

Measure it on your own footage

Tell us which situations matter on your site. We set the use cases up on a few of your cameras, tune the rules with you, and count the right and wrong alerts together before you decide.

Or email support@habilelabs.io · Monday to Friday, 10:00–19:00 IST

  1. Subscribe on AWS Marketplace30 days with no software charge on one instance; AWS bills the instance and storage.
  2. Launch the stackPick a Region and a size. A working console in about six minutes.
  3. Add your camerasGive the stack a route to your site (a VPN, Direct Connect or a peered VPC) and add each camera by its address.