Neural networks to monitor work safety

VizorLabs startup helps companies to monitor safety at production facilities, including whether staff wear protective gear such as hard hats, gloves, uniform and other items, using neural networks.

The company’s clients include Rosenergoatom (Kola NPP), Gazprom Neft and others. During the very first year the startup broke even and plans to make RUR 35 mln ($555K) by the end of this year, VizorLabs CEO and co-owner Vasily Dolgov told Invest Foresight.

VizorLabs team consists of alumni and postgraduate students of the Lomonosov Moscow State University’s Department of Computational Mathematics and Cybernetics and the Department of Mechanics and Mathematics. They already have experience in creating video analytical systems in electric energy and oil and gas production using neural networks. The startup requires RUR 30 mln ($476K) in investment.  

Vasily Dolgov himself has been in IT for 15 years, including 12 years as IT Director at Polyus Gold where he thought of using neural networks for detecting protective gear at hazardous production facilities.

How does VizorLabs work – or its software, VizorLabs Health &Safety?

The company links up to a client’s video surveillance system and analyzes conduct of the client’s staff. It will then create and send to the client video reports on any detected violations. Neural networks analyze whether employees wear protection – a task for which the networks were specifically developed by VizorLabs. The key detector developed by the startup is a human skeleton detector. Markers outlined skeletons with squares on thousands of stills of video that were later used for teaching neural network to recognize body parts by their skeleton. The job was primarily done by the company employees. The next stage was to separately mark (outline) gloves on hands and hands without gloves, hard hats on heads and heads without hard hats. The same with shoes, clothing, etc. It took 50,000 images to teach all the necessary neural networks. Initially, the startup created its own system for data separation but then switched to Yandex.Toloka.

As a result, several trained neural networks in the system work together to identify heads, hands and legs where the system analyzes whether a person is wearing gloves, the right shoes, a hard hat, whether his uniform (jacket and pants) is zipped or unzipped, etc.

Why are companies so concerned about their staff using safety precautions? Any accident at work starts with minor violations, which can eventually lead to death and large insurance payments, as well as the loss of corporate assets.

“Once we say we are watching them and will know if they wear helmets and gloves or not, the number of serious accidents declines – just like in the Heinrich pyramid,” Vasily Dolgov says.

(According to the Heinrich model of accident causation, a catastrophe is usually preceded by a series of smaller incidents no one gives any importance to).

The VizorLabs solution is used to control the wearing of helmets and gloves at the Gazprom Neft warehouse facility and at the Kola nuclear power plant. NPP workers are reminded “of the need to use personal protective equipment with a sound or light signal.” The employee is identified by the company’s photo database, and a supervisor gets notified of the violation of safety regulations. It is a pilot project so far.

Rosneft uses drones to control the use of personal protective equipment on its rigs. Dolgov believes that it is too expensive. An indefinite license for a VizorLabs detector costs from RUR 20K ($320), and there is a platform as a service option available.

In addition to people, VizorLabs can monitor equipment, and the way people use it. The system can watch the passage of heavy vehicles, a crane lifting loads (the angle between the slings, and how securely the load is attached), based on video data and cameras installed on the cab. The cameras can also ensure that employees do not enter danger zones.

Over two years, the founders have invested RUR 12 mio ($190,620) of their own funds in the VizorLabs project. In 2018 – the first year of operation – the company made RUR 9 mio ($142,950); its 2019 revenue is expected to be almost 4 times that amount.

There are several object recognition solutions for industrial enterprises now available on the market. Russian developer Center2M offers an IBM-based product, which guarantees 98% accuracy. But this solution involves the purchase of IBM software as well as the use of their servers, so overall it costs $1 mio, which is much more expensive than VizorLabs. Another important competitor is software from NVI Solutions.

By Natalia Kuznetsova

Previous ArticleNext Article