Artificial intelligence for video surveillance, on your existing cameras
- Object classes, not identities
- Confidence shown as returned
- No face recognition, ever
Interior, after hours
Fixed camera, low light, two subjects
Analysing the clip
Clip
Decoded in your browser. Your video is never uploaded.
Sets the luma delta threshold and the smallest region that counts.
Kills the one frame flicker that no operator wants paged about.
Exclude zone
Put it over the road, the tree, the neighbour's window. Events inside it stop alerting.
The frame an event was found on shows up here.
Nothing selected yet
Hover a row in the event table to preview its frame, click it to hold the frame and jump the clip there.
- Length
- 0.0s
- Confidence
- 0%
- Area
- 0.0%
0
Motion events found
0
Dropped by zone
0
Dropped, too short
0
Alerts you would get
| # | Start | Length | Class | Conf. | Area |
|---|
Reading the clip. Events appear here as they are found.
Get these alerts from your own cameras.
AI video surveillance software is a contract, and this is ours in full.
Given to the model
- Two or three small cropped keyframes from one detected event, not the video.
- No location, no camera name, no site name, no customer identity.
- No history of previous events, so nothing is inferred from a pattern of behaviour.
Returned by the model
- One class from a fixed list: person, vehicle, animal, package, other motion.
- A one line description in plain language.
- A confidence value, displayed exactly as it came back.
Separating the measured half from the model half is the whole design. It is why an event survives a failed classification call with its data intact, and why nothing on screen is ever a plausible guess dressed as a measurement.
Four things this will not do, and they are not on a roadmap.
No face recognition
No face templates are computed, stored or compared. In Illinois, Texas and Washington this is not a preference, it is the difference between a purchase your counsel signs off and one they do not.
No re identification
The system does not link a person seen at 01:47 on camera two to a person seen at 02:31 on camera five. Each event is classified on its own.
No behaviour scoring
Nothing here rates a person as suspicious, loitering or intent on anything. A model that infers intent from posture will be wrong about a real person on a real night.
No guessed labels
If the classification call fails or is rate limited, the event is labelled unclassified and keeps its real motion data. We never substitute a plausible class.
The version written for a security questionnaire, with retention, encryption and sub processors, is on the enterprise security camera systems page.
Five conditions where accuracy drops, and what to do about each.
Any vendor who tells you their detection is always right has not run it through a wet November.
Very small objects at distance
A person at the far end of a long yard may occupy too few pixels to classify confidently. Fix: a second camera covering the far end, or accept a lower confidence threshold there and read the description.
Heavy occlusion
Someone half behind a vehicle or a pallet stack is a harder problem for any model. Fix: camera angles that cross the space rather than looking down it.
Extreme weather at night
Driving rain under infrared reduces contrast for everything. Duration filtering handles most of it, and confidence values will honestly be lower on those nights.
Unusual camera angles
A camera looking almost straight down produces a shape a model sees less often than a normal oblique view. Fix: mount at an angle where a person looks like a person.
Ambiguous classes
A parcel held by a person, a bicycle, a person in a wheelchair. The description line usually carries the nuance the class cannot, which is why every row has one.
Dirty or fogged optics
No model is better than its input. A clean dome is the cheapest accuracy improvement available and it is free.
The arithmetic that decides whether artificial intelligence is worth paying for here.
The model does one job: it removes the alerts that thresholds cannot remove. Duration and zones are free arithmetic and they take out most of the volume. What is left over is the set of cases where a pixel counter genuinely cannot tell the difference: a moth against a person in dark clothing, headlights against a torch, rain against somebody walking through rain.
That last slice is small in count and it is the entire reason the channel gets muted, which is why it is worth about ten dollars per camera per month and not a hundred.
Fifteen cameras, one site
The arithmetic, with your own numbers in place of ours
The enterprise figure is that market's published per camera positioning, not a quote we have seen. Get your own and put the two next to each other.
The questions that come before a trial.
01
Which model is doing the classification?
A general purpose vision model, called through a thin server side proxy so no key is ever in your browser. The important part is the contract described above: it receives cropped keyframes and nothing identifying, and it returns one class from a fixed list plus a confidence value.
02
Does our footage train anything?
No. Frames are sent for a single classification call and are not retained by us for training. In the demo the video itself never leaves your browser at all.
03
Is this compliant with BIPA and similar state laws?
We are not your lawyers, so the honest answer is that you should ask them. What we can tell you is the fact those statutes turn on: this product computes no biometric identifiers of any kind, because it never builds a face template.
04
How accurate is it?
We are not going to publish an accuracy percentage, because a single number across every camera, angle and weather condition would be meaningless and we have no honest way to measure yours. Run your own clips through the demo and judge it on your own footage, which is the only number that matters to you.
Related pages on the model
Video management software for IP cameras
Six VMS platforms on licence model and price, and which half of the job you already own.
Artificial intelligence video surveillance
Connect, detect, classify, alert, stage by stage.
Person detection camera software
The class that carries most alerts, without face recognition.
AI security camera software
The camera side: compatibility and setup.
Enterprise security camera systems
Retention, encryption, SSO and the DPA.
AI video analytics software features
The full capability list, grouped by job.
Best video analytics software
How this compares against seven other products.
Judge the model on your own footage.
The demo runs the whole loop on a clip you choose, in your browser. Read the confidence values yourself before you believe anything on this page.