One comrade - without a helmet, the second - without a glove.
In production, there are many cameras that are not the best, and not the most attentive grandmothers look at the squares. More precisely, they simply go crazy with monotony and do not always see incidents. Then they slowly call, and if it was an entry into a dangerous zone, then sometimes it makes no sense to call the workshop, you can immediately worker's relatives.
Progress has come to the point that the robot can see everything and give lyuli to anyone who violates. For example, by reminding by SMS, by a slight discharge of current to the siren, by vibration, by a nasty squeak, by a flash of bright light, or simply by telling the manager.
Specifically:
- It is very easy to recognize people without helmets. Even bald ones. We saw a man without a helmet - immediately an alert to the operator or the head of the shop.
- The same applies to goggles and gloves in hazardous industries, belt insurance (although we are only looking at the carbine for now), reflective vests, respirators, hair caps and other PPE. Now the system is trained to recognize 20 types of Sizov.
- You can accurately count people at the facility and take into account when and how many there were.
- It is possible to give an alarm when a person enters the danger zone, and this zone can be adjusted upon the start-stop of the machines.
And so on. The simplest example is the color differentiation of bricklayers and concrete pourers based on the color of their helmets. To help the robot. After all, to live in a society with no color differentiation is to have no purpose.
How to steal at a construction site
One type of common theft is when a contractor promised to bring 100 workers to the site, but in fact brought 40-45. And the house is being built and built. Still, no one can accurately calculate them in fact. As in a well-known joke: if a bear settles at a construction site and eats people, then no one will notice. So the general contractor has no way to control the brigades. More precisely, even if you use ACS, he will still be deceived, .
Usually there are no access control systems at construction sites, or they are only for entry.
We went to highly developed civilizations to exchange experiences and saw that each profession (more precisely, role) has its own helmet color. Here bricklayers are laying bricks - they have blue helmets, pourers are pouring concrete - they have green ones, all sorts of smart people walk nearby - they have yellow ones, so you have to do two βkuβ in front of them. And so on.
And you need all this to very easily detect each role. There are several dozens of rather cheap cameras on the site, which give something like 320x200 in color. Helmet workers are counted in real time, and a specific construction site is assigned to each camera. As a result, all this at the end of the day in the analytics is stitched into account of the schedules by zones: who, in what quantity and in what area worked.
In general, we adopted the experience. Only while we were looking closely at it, neural networks stepped forward, and many new detectors appeared. A few years ago, they were rather capricious and unstable, but now they allow you to very accurately catch the most interesting situations. Not least because of the speed of processing, detectors often make mistakes on individual frames, and on a video stream with small changes in angle, we get an excellent practical result.
What if I put a second helmet on my belt?
First we learned that a worker can get two helmets and put one of them on his butt. We have two detectors at once: searching for a skeleton and determining a color spot for correspondence with the top of this skeleton and searching for synchronously moving objects. The second one turned out to be easier to detect: for example, a person with a helmet on his ass almost never looks around with this helmet. Because you have to move your head to do it. And this movement is very easy to detect. More precisely, we donβt know what exactly is actually detected there (itβs a neural network), but it learned very quickly and catches violators, one might say, by gait.
We are building a human model.
Then we just build a real-time heatmap and reports at the end of the day.
Accordingly, according to the same principle - by training a neural network - the following are easily detected:
- Helmets.
- Bathrobes.
- Vests.
- Boots.
- Sticking hair.
- Safety carabiners.
- Respirators.
- Protective glasses.
- Proper wearing of the jacket (important for electrical equipment: it can slam in the machine room at the factory).
- Removal of large tools outside the perimeter.
In total, 29 detectors have already been tested. The only point is that since we work in hazardous industries like chemistry or mining, there are requirements for the types of gloves. For example, long and short. In this case, it is necessary that they be of different colors: it is very difficult to determine the length under the sleeve on the video camera.
And here there were often drawdowns on rats. We do not have a separate rat detector, but there is a detector of objects that interfere with the operation of the machine:
What else is being detected?
We tested detectors in chemical industries, in the mining sector, in the nuclear industry and at construction sites. It turned out that with little effort, you can close a few more requirements that were previously solved by the same grandmothers, who were crazy trying to see something in the picture through poor resolution and with a bad frame rate. Specifically:
- Since we are still building a skeletal model of each worker, falls can be detected. By falling, you can immediately stop the machine next to which it is located (in pilot implementations, there was no such integration, there were just alarms). Well, if you have IioT.
- Of course, being in dangerous areas. It's very easy, very accurate and very useful to everyone. At metallurgical enterprises, people work next to vats of boiling steel, it is useful to harden steel, but sometimes it is dangerous to stand a little on the wrong side. Taking into account the operation of different components and equipment, these dangerous zones can be changed, set a schedule for them, and so on.
- Another very useful PPE detector monitors the responsibility of employees and checks that they are not in danger. Here the grandmother approaches the task of accounting very responsibly and wears all the PPE assigned to her. Commendable!
Behavior control was implemented very easily - whether the employee is sleeping or not. While we were testing all this, the rules evolved from "There must be a person in a green helmet in this area" to "A person in a green helmet must move in this area." So far, there was only one smart guy who cut through the chip and turned on the fan, but this also turned out to be easy to fix.
It was very important for chemists to fix all sorts of jets of steam and smoke. In the oil industry - the integrity of the pipes. Fire is generally a standard detector. And there is also a check of closed hatches.
In the same way, forgotten things are detected. We ran it in one of the stations a couple of years ago, there it almost does not make sense due to the large number of events. But in industries, especially chemical ones, it is very convenient to keep track of things in a clean area.
Interestingly, right from the video analytics, we can read the readings of devices in the camera area. This is relevant for the same chemists whose production complexes have a high hazard class. Any change, such as replacing a sensor, is a re-negotiation of the project. It's long, expensive and painful. More precisely, LONG, EXPENSIVE and PAINFUL. Therefore, their Internet of Things will come late. Now they want video surveillance on the meters and read the data, quickly respond to them and reduce losses due to unexpectedly and unnoticed equipment failure. Based on the actual data of the counters, it is possible to build a digital twin of the enterprise, implement predictive maintenance and repair, but that's a completely different story ... Control is already there: we are now writing proactive analytics based on the totality of data. And separately - a battery replacement prediction module.
Another incredible thing - it turned out that in granaries and in storage facilities for materials such as crushed stone, you can shoot a pile from 3-4 angles and determine its edges. And having determined the edges, give the volume of grain or material with an error of up to 1%.
The last detector that we wrote was the control of driver fatigue, such as βnose peckingβ, yawning and blinking frequency. This is for HD cameras where you can see the eyes. Most likely, it will be placed in the control room. But the main need is for BelAZ trucks, KamAZ trucks for quarries. There, sometimes, cars fall, so now they are forced to come up with something in order to control the driver. Robot is better than grandma.
About cars. For example, the topic of fatigue control is actively used by automakers not only for BelAZ, KamAZ and other MAZs. Already in ordinary ordinary cars, manufacturers are integrating driver fatigue warning systems, but so far they have fairly simple solutions that analyze only the position of the car relative to the markings and the nature of the steering wheel movement. We went further and detect human behavior, which is much more complicated.
Another case of tracking a driver is the detection of incorrect behavior when using car sharing machines. They can not talk on the phone without hands free, eat, drink, smoke and much more.
Oh, and the last one. For several years now, we have been able to do object tracking between cameras - when, for example, something was stolen, you need to check which way and how. If there are 100 cameras at the facility, then you are tormented to lift the material. And then the system will automatically generate an action-packed thriller about Ocean and his friends.
What is the difference from the system two years ago? Now this is not just a recognition like βa bald man in an orange jacket left one cell and almost immediately went into anotherβ, but a mathematical model of the room is being built, and hypotheses about the movement of the object are based on it. That is, all this began to work in areas with overlaps and places with blind spots, and sometimes extensive ones. Yes, and detectors are now much better, because there are libraries that determine age by face. On HD cameras, you can set orientations like βa man of 30 years old with a woman of 35 years oldβ.
So, maybe in 5-7 years we will finish production and go to your house. For safety. It's in your best interest, citizen!
references
- My mail - [email protected]
Source: habr.com