Watch a room for changes with the Anomaly Assistant

Motion detection reports every flicker. The Anomaly Assistant works the other way round: it learns what a place normally looks like and reports only meaningful differences that last. It suits places that should stay as they are: an empty house, a vault, a barn stall, a bed, a storeroom shelf.

It reports:

Dust, compression noise, small lighting changes, shadows, insects and the sun creeping over the floor are filtered out. The detection is computer vision and works without any AI model. A vision language model can add a plain-language description, and can double-check events before they are raised.

Step 1. Draw the areas to watch

Skip this step to watch the whole picture.

1. Open Cameras, choose the camera and open its Scene Builder.
2. Draw a zone around each area that has its own rules, for example “Bed” or “North stall”. Untick its measures if you do not need entry and exit counts for it.
3. Draw ignore regions over parts that change all the time: a TV, a monitor, a window with trees. The assistant never watches them.
4. Save the scene.

Step 2. Turn the assistant on

1. Open the experiment in Experiments (or create one for the camera).
2. Under Objects to track, tick the classes whose presence matters, for example Person, Dog and Horse. Pixel changes are found without them; presence needs them.
3. In the Anomaly Assistant panel, tick Watch for meaningful change.
4. Under Watch, tick The whole picture and/or the zones.
5. Open each area and write its Normal state, in plain words, for example The stall is empty. or The bed is made; nobody is here at night. The sentence is shown with every event and helps a vision model.
6. Click Save.

Step 3. Set how strict each area is

Each area has its own settings:

Settings for the whole camera, under Settings for the whole camera: how long the normal picture is learned at the start (8 s), how many analyses per second (4), how fast it follows daylight, and whether a covered camera or switched lights are reported.

Step 4. Start and let it learn

1. Keep the scene as it should be and click Start.
2. On Live, the run shows Learning the normal picture, then Watching for anomalies. When something is found the pill turns red (Anomaly now) and counts the events.
3. Re-learn normal takes what the camera sees now as normal, for example after you rearranged the room on purpose.

Step 5. Read an anomaly

Open Anomalies (or click the pill on Live). Each row is one event; click it to see:

• the normal picture from just before, the event with the watched zone in amber and the change tinted red, and close-ups before and at the event (and at the end, once it ended);
• the computer-vision summary, for example A dog entered North stall; lasted 18 s, back to normal.;
• when it started, how long it took to confirm, how it ended, the confidence and the changed area;
• the tracked objects. With the licensed recognition modules they show as Person A with “recognized person” or “unknown person”. Viewers with a recognition token see the name, everyone else does not.

Review the run jumps to the moment on the run’s review page, where the video clip plays if the experiment records clips. Tick anomaly among the clip events in Video recording to record one for every anomaly.

Mark each event Correct or False alarm. For a false alarm on a running camera, tick take the current picture of this area as normal so it stops reporting the same thing.

Step 6. Add a vision language model (optional)

A model turns the evidence into a sentence such as An animal entered the normally empty north barn zone and remained for 18 seconds. It interprets what the computer vision found. It never detects on its own.

1. Open Anomaly Assistant in the menu.
2. Choose a Provider:
   • Local model: runs on this computer, and pictures never leave it. Set it up in Local vision models on the same page, without leaving it: Install Ollama (Windows: it downloads the official installer, checks that it is signed by Ollama and installs it for your user; on Linux and macOS the page shows the command), then Download a model from the list and click Use. The list shows the download size, the GPU memory each model needs and whether it fits your card. A 12 GB card runs qwen2.5vl:7b next to the detector (about 5.5 GB for the model, about 1 s per event once loaded, 3 s with four pictures).
   • OpenAI, Anthropic Claude, Google Gemini, OpenRouter: paste an API key. The pictures of each event are sent to that provider.
   • DeepSeek reads text only, so it gets the written observations without pictures.
   • Other compatible API for any OpenAI-compatible endpoint.
3. Click Save, then Test. It sends a tiny made-up picture pair and should answer confirmed.
4. Back on the experiment, set each area’s Vision model:

If the model cannot be asked (no model, an error, the hourly budget or the queue is full), When the model cannot answer decides: raise the event anyway (default, safer), or hold it for review. Raise now on the Anomalies page raises a held event.

Step 7. Send alerts

Each raised anomaly is an ordinary event of type Anomaly: it appears on Live and in Data, is counted and exported, and starts video clips. To push it elsewhere, list webhook addresses under the area’s Webhooks. They receive the event, then anomaly.described with the model’s description and anomaly.ended with the duration. See guide 20 for receiving webhooks.

Try it without a camera

The command line runs the assistant on any video, outside the platform:

cvscope anomaly analyse samples/videos/people-detection.mp4 --out anomaly-test

It prints every event and writes the pictures to anomaly-test. Watch part of the picture with --zone "Door=0.0,0.2 0.3,0.2 0.3,1.0 0.0,1.0" (points from 0 to 1), and try --sensitivity high or --persistence 5.