Neural network can differentiate smoke from fog

A neural network-based fire notification system developed by Russia’s ITV | AxxonSoft has been installed in the Miloshevichi forestry station in the Gomel Region, Belarus. The detection distance reaches 50 km in good weather. The company also trains neural networks to recognize faces, cars, license plates, potentially dangerous behavior, etc. for video surveillance systems. ITV | AxxonSoft is engaged in video analytics based on artificial intelligence, as well as the development of software for security systems, and its business is thriving. In 2018, the company’s revenue exceeded RUR 1 bln ($16 mio), and a 30% growth is expected in 2019. Igor Falomkin, software development director at ITV | AxxonSoft, told Invest Foresight more about the project.

Neural network-based smoke detectors are operating on three reversible cameras combined with infrared cameras. They are installed on fire towers at the height of 35 meters. The cameras are rotating vertically in a 360° angle to detect forest fire smoke.

The neural network is trained to differentiate smoke from clouds, fog and dust. The video from the cameras and the interactive map with real time camera FOVs are displayed on the operator’s monitor. If smoke is detected, the system will send an alarm and the operators will receive e-mails with a screenshot of the map and the image on which smoke was seen. A fire can be localized by two cameras with an accuracy of several hundred meters.

The neural network was trained using video data provided by the customer: there were 10,000 images featuring smoke and 10,000 without it but with such similar events as fog, clouds and dust. The neural network analyzed the images and learned to differentiate smoke. Several tests were held, all of them successful except for one time when smoke was not visible to the camera. The customer was satisfied with the results.

ITV | AxxonSoft neural networks can also detect fire, which is useful both in forests and other open spaces and large rooms with an active air flow, such as open warehouses, large hangars and tunnels. Standard fire alarms are almost useless there, while the smart camera can minimize damage with early notification.

Along with training its own neural networks, ITV | AxxonSoft also uses solutions by other manufacturers, including the facial recognition system by the Moscow-based company Technologii Videonaliza (Tevian) and a vehicle plate numbers recognition module developed by the Moscow-based company компании Technologii Raspoznavaniya (Recognition Technologies).

“Recognizing faces, vehicle plate numbers or railroad car numbers is a narrow market involving companies that have vast experience in this field. There is no point in competing with them, and our company, as a manufacturer of integrated video monitoring systems, uses their developments for our product and offers the complete solution to customers. We develop detectors for specific tasks that have no commercial algorithms. One of our recent developments is the detector that differentiates pedestrians from cyclists or people using scooters and hoverboards, and allows preventing them from riding out into pedestrian paths. The project has been implemented in Singapore”, Igor Falomkin says.

So far this is a pilot project, with no large-scale introduction.

Despite neural network analytics being one of the mainstream areas of modern video surveillance’s technological development,  ITV | AxxonSoft has major customers that use the company products without its involvement. For instance, Sberbank’s Northeastern Bank has installed these video surveillance systems in over 400 officers and 4,500 ATMs, creating a unified monitoring center. The system integrator’s major clients include Gazprom with its headquarters at Lakhta Center and the gas filling station network, as well as Mexico City International Airport, Istanbul Metro, United Arab Emirates police, New York City schools, Arsenal Stadium in London, and many others.

The Russian company’s software has also found its application abroad. In late 2017, the South Korean city of Yangsan-si implemented a Safe City system based on ITV | AxxonSoft.  Analyzing footage from 3,200 cameras and using smart search, the system can quickly identify people based on their photographs, cars by their license plates and much more. Overall, more than 2.5 mln cameras are involved in the projects based on the company’s products all over the world.

“Urban video surveillance rarely requires real-time video analysis. First of all, in the chaos of city traffic it is very difficult to determine which incident should be considered ‘alarming’. Second of all, dozens of operators that could respond to such incidents is a very expensive resource, especially in countries with high wages. But fast search for video footage from thousands of cameras based on numerous criteria is a huge support for police investigating wrongdoings,” Igor Falomkin explains.

By Natalia Kuznetsova

Previous ArticleNext Article