Industries / Finance and business services / Security guard services
Watching cameras and alarms
This is the work of watching alarm signals and camera pictures, on site or from a monitoring centre far away, deciding which are real, and sending a guard or calling the police.
5
Milestones achieved, from 1858 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1858
The electric burglar alarm
On 21 February 1858 Edwin Holmes of Boston installed what CBS News calls "the very first electric burglar alarm", using a patent he had bought the year before from the inventor Augustus Pope. Holmes went on to use telephone lines to connect alarms to "ever-alert central offices", the forerunners of today's monitoring centres, and set up his own uniformed security service.
Credit
Edwin Holmes; Augustus Pope
1969
Cameras at the front door
Marie Van Brittan Brown and Albert L. Brown of Jamaica, New York, filed a patent on 1 August 1966, granted on 2 December 1969, for a "Home security system utilizing television surveillance": a camera at the door, a screen inside, and "a radio controlled alarm... for alerting a guard, policeman or watchman at a security station". The Lemelson-MIT Program calls it "the first closed-circuit television security system".
Credit
Marie Van Brittan Brown and Albert L. Brown
1996
A camera on the network
On 17 September 1996 in Atlanta, Axis Communications, a Swedish maker of print servers, launched the AXIS 200, in what it calls "the first-ever launch of a network camera": a camera whose pictures could be viewed in a web browser over a computer network instead of on a closed cable.
Credit
Axis Communications, by its own account
2014
The shared test set closes
The UK Home Office's i-LIDS datasets were sets of video for testing software that raises alarms from camera pictures against standard scenes, such as a "sterile zone" that nobody should enter. On 30 November 2014 the Home Office's Centre for Applied Science and Technology (CAST) stopped distributing them, though the building research body BRE kept offering evaluations against the sterile zone scenarios.
Credit
UK Home Office
2023
A score for every alarm
In January 2023 The Monitoring Association (TMA), the trade body for alarm monitoring centres, announced that its standard TMA-AVS-01 had "received official ANSI accreditation". It sets "standardized methods for calculating an alarm score" from Level 0, "No Call for Service", through Level 2, "confirmed or 'highly probable' human presence with unknown intent", to Level 4, "confirmed threat to life". The chair of the committee that wrote it expects AI to "assist automation platforms in determining an initial alarm level", with the operator as "the final arbitrator".
Credit
The Monitoring Association; American National Standards Institute
Watch for
AI that drops empty clips
Noonlight, a monitoring company, says its AI Person Filtering will "automatically dismiss any event where no person is detected", "delivering 99% recall and 97% precision". Recall here is the share of clips with a person that the filter keeps, and precision the share of kept clips that really show one. It gives no test design or data set, and it reports "early data" of a 45% fall in false alarms from a separate feature, Advanced Verification with "talkdown", in which an operator speaks to the site through a loudspeaker.
Achieved whenSomeone outside Noonlight tests the filter on clips from sites Noonlight did not choose and publishes recall and precision.
Claimed by
Noonlight, March 2026
Open
01
Nuisance alarms, labelled and public
A free, public set of short camera clips that set off motion alarms with no person in them (foxes, cats, rain, snow, spiders on the lens, swaying branches, car headlights), with staged intrusions acted out by mannequins or by figures drawn in a game engine, each clip labelled, and a scoring script for alarm filters.
The briefCredit · unclaimed02
Check AI alarm scores
A public set of invented alarm events (which sensors tripped and in what order, what the video software reported, what the keypad and the call list showed), each given an AVS-01 level by three experienced operators, and a measure of how often AI software gives the same level, with extra weight on Levels 3 and 4.
The briefCredit · unclaimed
Open 01 · the brief
Nuisance alarms, labelled and public
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free, public set of short camera clips that set off motion alarms with no person in them (foxes, cats, rain, snow, spiders on the lens, swaying branches, car headlights), with staged intrusions acted out by mannequins or by figures drawn in a game engine, each clip labelled, and a scoring script for alarm filters.
Why it is still open
The UK government's i-LIDS video has not been handed out since 30 November 2014. Vendors' figures come from their own tests: Noonlight gives 99% recall and 97% precision with no test described. At Y-12, the Inspector General linked "issues with false alarms" to complacency among the guards. No public, freely licensed set of nuisance alarm clips was found.
Built with what exists
Outdoor cameras on property the team controls; Blender or Unreal Engine for drawn figures; Meta's SAM 2 to speed up labelling; open detectors such as Ultralytics YOLO11 as a first filter to score.
First step
Mount three cameras on property the team controls, record a month of motion-triggered clips at night, and label them.
Done looks like
5,000 clips from at least 10 sites, at least 3,000 of them nuisance clips and 500 staged intrusions, released under an open licence; recall and precision published for at least three filters; any clip that shows a real person deleted before release.
Open 02 · the brief
Check AI alarm scores
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of invented alarm events (which sensors tripped and in what order, what the video software reported, what the keypad and the call list showed), each given an AVS-01 level by three experienced operators, and a measure of how often AI software gives the same level, with extra weight on Levels 3 and 4.
Why it is still open
AVS-01 was accredited in January 2023, and the chair of its committee expects AI to propose "an initial alarm level". No public set of scored alarm events, and no test of AI scoring against operators, was found.
Built with what exists
The AVS-01 standard, which TMA publishes; Python to generate the event logs; an AI model that returns probabilities for fixed answers, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), or a general model such as Claude Opus 5.5.
First step
Download AVS-01, write 50 events that cover every level, and have two operators score them separately.
Done looks like
1,000 events, each scored by three operators, with their agreement reported; at least three AI systems scored; every case where an AI gives Level 0 or 1 to an event the operators put at Level 3 or 4 listed.