"A few years ago, this type of artificial intelligence-based fire detection just simply was technically impossible in real time."

Pixel-by-Pixel Prevention

• By Ronnie Rittenberry
• Oct 01, 2005

Editor's note: Early warning is everything when it comes to fires, say the developers of SigniFire^TM, a software system that uses patented image-analysis technology designed for instant fire, smoke, and intrusion detection. Launched commercially in early 2005 by Baltimore-based axonX LLC (www.axonx.com), the system uses an advanced digital video recorder (DVR) platform to "see" threats much faster on average than smoke detectors and immediately sound alarms. With the software's instant visual verification, the prevalent problem of false alarms resulting from system malfunctions can all but be eliminated, axonX founder and CTO George Privalov and CEO Mac Mottley said in a July 29, 2005, conversation with Occupational Health & Safety's managing editor. The company recently developed a strategic partnering relationship with Milwaukee-based Johnson Controls' Fire & Security Division to provide distribution and integration capabilities for the SigniFire technology throughout the United States in non-residential facilities. Excerpts from the July conversation follow.

Video image detection is relatively new technology. Can you explain how SigniFire works, what it does?

Mac Mottley: It's basically artificial intelligence in software form. The software fits on a digital video recorder--a standard, Pentium-based digital video recorder--and the software analyzes standard CCTV [Closed Circuit TV] security camera feeds. The software looks frame by frame, pixel by pixel, and there's basically five algorithms that detect flaming fire, reflected firelight, two smoke-detection algorithms, and there's motion detection. So basically we can transform a standard security camera system into long-range, early warning fire and smoke detection.

Mottley: Yes. The first patent was issued in 2001. The processing speeds of today's computer technology basically got to the point two to three years ago where we could run it real-time. Before, images had to be stored and then run through the algorithms. Now, we can handle eight cameras at one time with a Pentium 4-based processor.

Mottley: Basically, the reflected firelight algorithm looks for some type of surface where the reflecting light is picked up at a certain frequency that's the same frequency as the fire that's being produced at the regular flaming site. . . .

George Privalov: Essentially, every frame has a flickering brightness, and what we are looking at is for the reflected surfaces that have the coherent reflection pattern. It's a mathematical procedure which we also patented and, essentially, once it exceeds some certain size and level of flickering and correlation, it triggers the alarm.

Privalov: Nobody was actually successfully doing any image-based fire detection before. There are a few companies right now in the world that have similar types of technologies, but we were rated the best among them.

Mottley: Right, there are two other technologies that we tested against at the Navy. One just did smoke detection; it didn't do any type of fire or flame recognition. The other company did one algorithm for flame but no reflected firelight or motion detection.

Privalov: The reason there's nobody there [offering the technology] is, as Mac said: A few years ago, it just simply was technically impossible in real time.

Mottley: We've been testing with them for the last two years. It's part of what they were calling their Advanced Damage-Control Volume Sensor Project. What they're looking at for the next generation of destroyers and carriers is for a sensing system that's volume-based. That means sort of like what you see in the whole compartment versus having point detectors, where the smoke or heat has to travel to that detector.

Privalov: The Department of Defense is looking at ways to reduce. The new generation of destroyers will have a very small crew on board, so they have many compartments but with relatively small staff to attend. So they need better situational awareness, obviously, and that's how it came around, the whole program. But what they are looking for is very close to what industry is looking for.

Mottley: It really doesn't have to be sophisticated at all. What we do is, we provide a digital video recorder, which is sort of the same thing that people are migrating to from analog-type taping--you know, VCRs--so we can use standard, analog cameras, the standard infrastructure that's already there. They can just feed the co-ax cable now into our box versus the box that they had previously. If they have a current security system, they can split the video feeds and keep their current system and then feed the cameras into our DVR, as well.

Privalov: It's an interesting point that we actually don't need very sophisticated cameras. We have better results with black and white, for example, than in color.

Mottley: Black and white can pick up fires at longer ranges. The software works with color; it's just that if we do special jobs--like an outdoor job--we'll specify black and white.

Privalov: Outdoors with bright sunlight is challenging for color cameras because, obviously, there's a lot of other light there so that the flame is not contrasting as well as with black and white, because they're sensitive to infrared; they're picking up flames much brighter than anything else. Any black-and-white camera is infrared-sensitive, actually. With a color camera, they suppress; they put a filter there to not allow infrared into the lens, because that will mess up the color balance.

Mottley: Right. They actually can get a security system as well as the fire detection capability.

Privalov: The detection is different. They're not related to each other. The fire is not detected based on motion. It's different things that the artificial intelligence looks for.

Mottley: The security part is basically just standard motion detection. You know, if you have a door that you want to monitor for entry, we can box that door with the software, so any time the door opens or somebody goes through it, it logs it as an event and will alert somebody that there's motion in that zone. The difference is, if you set up the fire algorithm--the fire algorithm covers the whole video--only fire sets it off. You could have hundreds of people walking through it and it won't set it off. Because a lot of times, obviously, people aren't going to set their motion detection during the day when everybody's walking around. They may have a schedule to set it up just at night. But you want to detect fire all the time.

Mottley: No, they don't. What we do is we give the immediate early warnings of the fire, so they can typically go out and put it out with other means besides sprinklers. In order for sprinklers to go off, the fire has to heat the element inside the sprinkler head to 200 degrees or more, so there's a raging fire in a building when sprinklers go off. Now, there are some applications for tying this technology into suppression systems automatically, but those are more specialized type hazards. Otherwise, the guard sees it on the video on his desk, and he can run put it out with a fire extinguisher before the whole place burns down.

Privalov: When the sprinklers go off, it's very substantial damage. The advantage of this technology is the guard can be sitting somewhere 1,000 miles away in a central monitoring station, and then he can call the local fire department to attend the event he will see firsthand on the video.

Mottley: Early warning is what this is all about. The whole ROI model is based on early warning--putting the fire out before it causes major damage.

Mottley: The range depends on the optics of the camera, the zoom angle of the lens, and the size of the fire, obviously. But to give you some examples: We can pick up an 18-inch fire at 125 feet. We designed for a pulp and paper mill that wants to protect their log yard, their raw material, and they're looking for a 10-foot-square fire. So we actually designed a system for a 30-degree-angle zoom lens, and we could go out 500 feet to pick up that fire. So it all depends on the size of the fire and the zoom angle.

Mottley: That's a good question. In a 20-by-30-foot room--to give you a sort of feel for the dimensions where most of the Navy tests were performed--we can detect a flaming fire in four seconds; the ion detectors detected in 10 to 15 seconds. However, the ion detectors took the longest to detect a smoldering type fire, where there's no flame and the heat and energy take much longer to build up. In those cases, half the time they would not detect at all, and half the time they would be from 10 to 15 minutes behind our technology. Now, the photoelectric [detectors] . . . do a better job at the smoldering fires, and they are, on average, about 5 minutes behind us on those.

Mottley: That's one of its biggest advantages. There's a major problem in the U.S. right now with false burglar and fire alarms. So what this video-based detection does is it can eliminate the majority of these false alarms, because when an alarm triggers at the central monitoring station, be it a motion-detection intrusion or a fire, the guard can look at the video and verify right away if there's an event or not. With other systems--a standard smoke detector system or something like these passive infrared detectors--you don?t know what's going on at the facility. . . . And you know, now, if you have a response to an alarm and it's a false alarm, you can get fined for it. There's definitely a large problem. I think your reader base will see this as a potential solution to this verified-response or false-alarm problem.

This column appears in the October 2005 issue of Occupational Health & Safety.

This article originally appeared in the October 2005 issue of Occupational Health & Safety.