Short answer: Type a description, like “a person in a white shirt” or “a white van at the loading dock”, and AI video search returns the moments across your security cameras that match. You review a short list of results instead of scrubbing hours of timeline.
How do people search footage today?
- Scrubbing the timeline at 8× or 16×, camera by camera. Slow, and easy to miss something.
- Motion search: draw a box and find when pixels changed inside it. Good for “did anyone open that door”, noisy outdoors where trees, rain and headlights all count as motion.
- Analytics filters: person or vehicle detection, line crossing. These have to be set up on each camera in advance, and often only work in that camera maker’s software.
What is AI video search?
Search by meaning instead of by motion. Most AI video search uses a model that understands both pictures and text, and compares your words with what is in the recorded frames. That lets you ask for things nobody set up a rule for in advance, on any camera.
How do I write a good search?
- Describe what the camera would see. Clothing colours, vehicle type and colour, things being carried. “Man in a dark hoodie carrying a box” works better than “suspicious person”.
- Describe what happened, not why. “Person climbing the fence” rather than “intruder”.
- One subject per search. Search for the van and the driver separately.
- Start broad, then add detail. If “red car” returns too much, try “red hatchback by the gate”.
- Narrow the time and cameras first. Only search the hours and cameras that could have seen it.
- Mind night footage. Cameras on infrared record in black and white, so colours won’t match after dark. Describe shapes and objects instead.
Some examples:
| Looking for | Try searching |
|---|---|
| A missing delivery | courier with a parcel at the front door |
| A vehicle | white van at the loading dock |
| A person | person in a red jacket with a backpack |
| Something left behind | black bag on the floor |
Once you’ve found it
A match on one camera is usually the start. Check the cameras next to it at the same moment, ideally played back together on one clock, to follow the person or vehicle. Then save what you found as an incident with clips from each camera, so it can be handed over.
Can it watch for something live?
Some systems let a search keep running against live cameras and alert you when something matches. It turns a search you’d otherwise repeat every morning into a notification.
What does AI search need?
- Recordings it can reach. It searches what was recorded, so cameras need to be recording.
- Compute. A GPU speeds it up considerably.
- A clear picture. Light, focus and resolution matter: AI can’t find what the camera couldn’t see.
How AI search works in Supervision
Type what you’d tell a colleague, and Supervision searches recorded footage on every camera at once, whatever the brand, including every channel of your NVRs. Narrow it by time range and camera so you only review what could be it.
Leave a search running and the desktop app tells you when a live camera matches. Open an incident straight from a match, attach clips from each camera that saw it, and export it as an MP4.
Common questions
Is my footage sent to the cloud for AI search?
Not with Supervision. AI search runs on your own Supervision server, which also does the recording, and footage never reaches us.
Does it work with any camera?
In Supervision, yes. It searches whatever Supervision records, from any ONVIF or RTSP camera and from every channel of an NVR.
Is AI search the same as motion detection?
No. Motion detection tells you that something changed in an area. AI search finds what you describe, wherever and whenever it appears.
Supervision is self-hosted video management software for any camera brand. If this is the kind of problem you're solving, tell us about your cameras and we'll talk it through, or email hello@bipolarfactory.com.