These compilations illustrate performance for both alarm and non-alarm scenarios. For comparison, identical, high, sensitivity settings were used in all these clips. (The only difference is that algorithm adaption functions are activated for the long-time clips which is not possible for short clips.) This is only a small sample of the total material used for developing SAF3. The clips do not correspond to the end-user GUI experience as they are generated directly from the algorithm development software. The bar to the right is the "dynamic-smoke-o-meter". When the bar reaches to the red level a smoke alarm is generated.

Smoke detection

These compilations demonstrate detection of dynamic smoke from real fires in various environments.
Source device: Web camera

Basic detection

This compilation demonstrates basic detection performance. It shows detection of smoke from early stage fires in a few different environments and ranges.

Source device: Axis P3384-VE

Realistic environment

These clips stems from live tests in realistic environments. The tests were performed in cooperation with the swedish emergency service.

Source device: Unknown

Random internet smoke

This compilation contains random clips from the Internet including a burning tire, a small smoke grenade and the Volgograd station bombing. It was created to show that no special setup or extreme requirements are needed for SAF3 to work.

Non-smoke clips

Detecting smoke is one thing. Not detecting non-smoke is another, but equally important: We don't want false alarms. These compilations contains samples from real, live, environments with no smoke. The clips are played at 10 to 1000 times the normal speed to shorten the run-time. (All processing is of course performed at nominal speed.) Note that these clips were generated without the adaptive sensitivity algorithm.
Source device: Unknown/Axis M1011

City timelapses

These clips come from Los Angeles (day/night), New York Times Square and an Oxford street. They span from 8 minute to 9 hour clips replayed at 10 to 1000 times normal speed. The point here is to demonstrate that no false alarms are generated regardless of light conditions and image activity.