Video analytics.
As for video analytics, the situation is not simple. Reality (for example, the same motion detectors), having barely appeared, immediately becomes overgrown with myths about their supposedly unlimited capabilities. In fairness, it should be noted that some myths, in turn, quickly become reality. What is the reason? Most likely, because the demand for the intelligence of video systems, the most, so to speak, accessible to mass understanding, is very high. Today, it many times exceeds the capabilities of the equipment. This, plus superficial coverage in the mass media and not always adequate representation in feature films (including good ones), partly gives rise to myths. Another important circumstance, sometimes forcing experts to use formulations like «more yes/no than no/yes» is that many developers actually have algorithms that allow us to talk about the «intelligence» of systems. But how do they work, and not on exhibition stands, but on real objects?
For example, not long ago one of the authors of the technical specifications told about a system created on the basis of neural algorithms. This is, if you like, a classic example of a self-learning system. It showed remarkable results in a prototype version, but to train it, it was supposed to be shown several thousand videos shot in various conditions, and, what is most unpleasant, half of them had to contain the target detectable situation — an intruder. Imagine the task: at a specific object, at least a thousand times to film a person imitating an intruder, penetrating in various ways, it is desirable to cover as widely as possible all possible manners and methods of penetration. And in order to generalize such a procedure, it is necessary to conduct the same on thousands of objects. This is millions of experiments. What is even worse: the experimental sample, capable of processing only a few dozen examples, worked relatively acceptably on a 3-GHz Pentium. And if you expand its capabilities to master millions of examples, you will need a cluster of at least a thousand computers — and all to process a single video signal.
But let's get down to business. The floor is given to the expert on duty for this column, the head of the technical department of the ISS Company, Alexander KOSOVSKY.
Intelligent video surveillance systems provide facial recognition and can help in finding a criminal.
Today, this is more a myth than a reality. At least, as far as objects with a high throughput of people (train stations, city streets, metro stations) are concerned. Face recognition is, in general, perhaps the most difficult of the tasks for which video analytics technologies are used. In order for a face recognition system to work, it is necessary to ensure a number of conditions: correctly select and install the equipment, ensure that the database contains high-quality photographs used for identification. The organizational side of the process is also very important: the flow of people must “correctly” fall into the field of view of video cameras so that the system can “capture” each individual face from the crowd.
Therefore, I repeat, in conditions of a large crowd of people, for example, in the flow of passengers at a metro station during rush hour or at a stadium during a major sporting event, it is almost impossible to do this. However, it is justified and logical to use a facial recognition system in the airport passport control area, where it is quite possible to create the necessary lighting conditions, correctly install video cameras, and ensure that each passenger's face is captured. Identification is carried out unnoticed, there is no physical contact between the user and the system. This also makes facial recognition systems indispensable for access control to specially protected facilities with a pass system. All this is true, but let's not forget about such a condition as the availability of HIGH-QUALITY photographs of the wanted persons. Such are far from always found even in the databases of law enforcement agencies. Let's say they are looking for a repeat offender who was last photographed ten years ago. Then he was fat and had a short haircut, and now he has lost weight, grown a head of hair, and even grew a moustache. Will the system identify him? Hardly, the necessary algorithms for this do not exist today. And based on the photofits of suspects, you can safely detain a third of the population of Russia — any artificial intelligence would go crazy based on them.
Video analytics are mainly motion detectors or abandoned object detectors.
Everything that we mean today by the generalized concept of «video analytics» began with detectors. And today, motion detectors are an essential component of any video surveillance system. What are modern motion detectors? These are devices with an unlimited number of detection zones and individual sensitivity settings for each of them, regardless of the number of simultaneously processed video channels, this is a setting for detecting objects of specified parameters. Such a motion detector is a truly effective modern technology for intelligent processing of video information. Today, the priority task of developers is to «force» the video surveillance system to function as much as possible in automatic mode, without human intervention. Therefore, various systems include not only motion detectors, but also rotation detectors, video camera backlighting/blocking. There are successful attempts to implement intelligent mechanisms for searching in the video archive, so that the operator does not need to view the archive of video recordings, but only specify a parameter, an event, for example, the fact of the appearance of a certain person in the video frame, to view the corresponding video fragment.
An abandoned object detector is a more complex component of a video surveillance system; it is designed to establish the presence of an object or objects of certain parameters in the field of view of video cameras. Such a detector can be used to ensure the safety of crowded places, technological premises, highways, where the unauthorized presence of people or any objects poses a potential threat. An abandoned object detector has every reason to be used for video surveillance of the subway, train stations, airports, shopping, entertainment, sports complexes, and venues for mass events. Another “untapped” area of application for this module is the prevention of emergency situations by detecting unauthorized movement of people, vehicles, and detecting foreign objects in technological and production areas, on runways, and highways.
The transition to IP video signal transmission technologies will give a powerful impetus to the development of intelligent systems
For now, this is more of a myth than a reality. The fact is that compression algorithms cause distortions and artifacts. Moreover, they naturally arise precisely at those moments when something interesting happens, important from the point of view of ensuring security, control of the object.
That is, there are still unresolved problems, but it is absolutely clear that the use of video analytics technologies in IP systems is an established trend. And this, in my opinion, is logical: the development of IP technologies gives impetus to the development of video surveillance in general. And, therefore, specialists are faced with the task of implementing all their achievements, developments now on the IP platform. IP projects are becoming more and more in demand every day, the share of orders for intelligent systems is increasing, so there is every reason for these two areas to actively develop «hand in hand».
The technical characteristics of IP cameras create certain difficulties for video analytical processing of the video images received from them. For example, at night, video recordings from IP cameras cannot be used to recognize license plates. But I am sure that these problems are temporary and will be solved by IP equipment manufacturers in the near future.
Intelligent systems allow you to recognize car license plates
This is a reality, first of all, at least because the systems with the function of recognizing car numbers are the most popular and in demand. They can solve a wide range of user tasks. The demand from a variety of customers is really very high, naturally, the market has formed and is constantly improving technical proposals. Current car number recognition systems, both domestic and imported, are strikingly different from their predecessors, say, five years ago.
Developers have managed to solve many complex problems, and current systems are no longer afraid of glare from car headlights, darkness, dirt on license plates. The systems can operate in a wide range of external conditions, integrate with radars, security and technological equipment, actuators and external databases. They can keep track of cars in a parking lot, control the transport fleet of an enterprise and register cargo, monitor the traffic flow of an entire city.
The probability of unconditional recognition (and this is the main indicator of work efficiency) in some modern systems exceeds 90%.