Fire prevention video analytics systems.
What is “fire prevention video analytics”? Computer vision algorithms, from which the term “video analysis” itself originates, have been used for a long time: recognition of license plates, faces, detection of passage through a border, selection and tracking of moving objects – all this, if not the past, then at least the present. But progress does not stand still, and in the last few years, algorithms have been created to identify signs of smoke and fire from an image. This is what will be discussed in this article.
“Classic” fire safety systems have existed for a long time, and in Russia their use is enshrined in federal law. But let's turn our attention to how we can improve their efficiency using video surveillance and, in particular, fire safety video analytics.
First, let's look at the facts and statistics.
PROBLEM
According to the U.S. Fire Administration (FEMA), although the number of fires in commercial premises has recently decreased to 3% per year (but the mortality rate is growing to 8.5% per year), this number is estimated at hundreds of thousands, and the losses are hundreds of millions of US dollars. The most interesting thing is that, despite the relatively high percentage of buildings equipped with automatic fire extinguishing systems (up to 60% in educational institutions), they do not work in 8% of cases. Let's think: what was there where it was not at all? The percentage of equipment of warehouses and office buildings with fire extinguishing systems does not exceed 20%, and for some classes of premises (such as garages) is close to zero. In 2000, more than 70% of such buildings were not equipped with fire safety systems at all.And what about Russia? In 2010, there were 179,098 fires, which is 4.5% less than last year, in which over 12,983 people died (a decrease of 6.9% compared to last year). 13,067 people were injured in fires. The State Fire Service units saved 84,548 people and material assets worth over 44.6 billion rubles.
What can be seen? According to statistics, even in the US the penetration of “classic” fire safety systems is far from 100%, let alone video surveillance (on the basis of which fire safety video analytics could work).
Here, as everywhere, the rule works: everyone pays for their own level of security. Some do nothing, some start and stop at the fire system, some install a security alarm and ACS in addition, and some install video surveillance. However, wherever video surveillance is already installed, as a rule, there is also a classic sensor-based fire protection system, and this is where video analysis can help us. Let's assume that a person already has fire sensors and a video surveillance system installed.
It is worth noting separately the cases when it is impossible to install a sensor-based fire protection system in principle – for example, an open parking lot, an inner courtyard, an open warehouse and other open spaces. It is in such cases that video analysis shows its unique advantages. Another example is detecting fires in forests.
Another significant advantage of a system with fire video analytics is a potentially faster response time compared to a sensor. If we do not take into account expensive multi-criteria fire sensors, then, for example, in the case of a draft in the room or simply a large volume of it, the “classic” fire protection system may either not work at all, or work only when the smoke concentration is such that even a draft is not an obstacle for it – too late! But the camera can see a whitish fog in the room much earlier.
HOW IT WORKS, OR NOT ALL DETECTORS ARE EQUALLY USEFUL
How does the smoke and/or fire detection algorithm work? The system tries to see the characteristic signs of a fire in the room. There are many different approaches: for example, primitive smoke/fire detectors simply record movement in the frame where there should be none (for example, during the day there are usually people in a warehouse, they will detect a fire anyway if it occurs, but what kind of movement can there be at night?). Of course, such an approach is unacceptable. You will get nothing but a bunch of false alarms and the inability to distinguish a shadow from dangerous signs of fire. A good detector differs from a bad one precisely by the number of false alarms. It is absolutely clear that a detector that constantly gives “false alarms” will not be taken seriously and will simply be switched off after a couple of days.
Due to the fact that there is usually no smoke without fire, developers combine smoke and fire detectors into one. But there is a big difference in their detection principles.
First of all, let's pay attention to the fact that smoke, like fire, can be different.
There is so-called fast smoke, typical for open spaces, where it dissipates and moves relatively quickly, and slow smoke, typical for closed spaces. It is important to understand that these two phenomena are perceived completely differently by the computer algorithm that is trying to see them. A system that sees smoke well indoors will most likely work worse in open spaces, and vice versa. The general criterion for smoke is usually a decrease in contrast in some local area of space, which changes its shape.
Note Fig. 1 – there is not enough smoke to trigger the sensors, but the system already sees clear signs of smoke (approximately the same way a person would see them) and gives an alarm.
Fire, for example, can be detected as a flickering area with changes in brightness intensity. Of course, one can imagine examples of scenes where a typical detector will give false positives when factors coincide. For example, a flickering CRT monitor, curtains fluttering in the wind, etc. But all these factors can be reduced to zero, for example, by masking these areas of the image.
How can a detector with image-based fire detection provide an additional level of security for an object?
PROSPECTS
There is an interesting law: even a weak programmer can write “some detector” in a couple of days, but sometimes only a large team with good brains and an impressive amount of time can make a “good detector”. And the only difference between them will be the number of false positives.
Unfortunately, it is difficult to say when the computer eye will be able to replace the good old operator. I would not bet on a computer in complex video analysis algorithms, but, on the other hand, it is obvious that where a system of 16 cameras begins, human attention “ends”.
Verdict: the future is in human-machine symbiosis. Only in this way will complex systems be able to bring benefits and not require dozens of operators. The future (or rather, almost the present) is that a computer, using video analysis tools, will provide a person (operator) with alarming incidents that require action, and the person will not be distracted by a multi-picture of 16 cameras, but will act on specific alarms that the computer has selected for him. For example, the system gives an alarm to the operator: “smoke was detected in camera 13” or “fire was detected in camera 37”, while the fire alarm can be set off specifically for a classic sensor, and the operator can respond to an early alarm or reject it if he considers it false. Verdict: the future is in human-machine symbiosis. Only in this way will complex systems be able to bring benefits and not require dozens of operators. The future (or rather, almost the present) is that a computer, using video analysis tools, will provide a person (an operator) with alarming incidents that require action, and the person will not be distracted by a multi-picture of 16 cameras, but will act on specific alarms that the computer has selected for him. For example, the system gives an alarm to the operator: “smoke was detected in camera 13” or “fire was detected in camera 37”, while the fire alarm can be set off specifically for a classic sensor, and the operator can respond to an early triggering or reject it if he considers it false. In the case when you are confident in your sensors, this system can even work in a delayed mode without an operator at all, i.e. send SMS/MMS with a request to “see what’s going on there”. And it will still be useful, because you know that “classic” sensors will trigger in the event of a fire, but if you have time and desire, you can get ahead of them and assess the situation based on the video image.
On the one hand, in closed spaces, systems with fire-fighting video analytics are clearly not capable (and should not) replace “classic” fire-fighting systems based on sensors. Here, their function is to duplicate the system and potentially trigger it earlier. This will not be superfluous, given that even according to US statistics, in one case out of 12 (these are the same 8% of cases), the “classic” system does not trigger at all. In Russia, the situation is most likely even worse.
But, on the other hand, when working in open spaces, computer vision algorithms can greatly help in saving property and the lives of our fellow citizens. And here they have few competitors.
