Identification of the person on a regular security camera: adventures on one pass

Photos from the object can not be. This is about a similar place on the example of CRIC.

It all started with a cat terminator. At the beginning of the operation, we knew the following:

During the construction of a building with a large grocery, the working team started a cat to catch rats.
After acceptance, the cat is unknown where it has been hid, and for three or four years no one has seen it.
Some bastard showed him a direct relationship between an open bag of cat food from the trading floor and the appearance of food.
He appeared only on the cameras - he came at night to hunt bags, and he did not touch children and females like a real experienced hunter, but took only fat males, that is, chose the largest bags, unexpectedly jumped on them and opened their belly.
They tried to impose a separate bowl with food for the night, but he was far from home, and could not refuse hunting.
In addition, the cat loved extremely expensive alcohol: the staff guessed to write off bottles for the cat. Like, dropped, the cleaning lady has already removed the fragments.
He brought the total loss by about 50–100 thousand a month (yes, it will be better than in the sensational story about a one-time cat’s dinner for $ 1000 at the Vladivostok airport).
For the capture of the cat for 4 months was a reward of 5 thousand rubles.

The idea of alcohol led us to the idea of what it is worth to check for starters how honest the employees are. So, word for word, we put our camera on the checkpoint at the back of the store and began to look at those who enter the building.
')
Here and us and the operating director were waiting for several discoveries, consistent with the Russian mentality. When we first learned what was happening, the main mate simply had no limits.

Disposition

So, the service entrance of the store, on which there are two cameras, - one looks at all incoming, the other at all leaving. Turnstile there can not be put on the requirements of fire safety, so just a corridor, where there is a rack for access control. Employees on the logic should pass by the security booth, put a pass to the rack, and then follow inside.

It was interesting for the shopkeeper to count smoke breaks, check any abuses and record the exact time of arrival at work. From this moment began the quest.

So, we put our camera (usual security camera) and supplied it with a long-focus lens. They set it down the corridor so as not to embarrass and allow it to keep the experiment clean. The staff were not surprised by the two other cameras and did not even attach importance to the event, which was clearly to our advantage.

The lighting inside was stable, but at first the door was very disturbing to us in recognition (it was very bright) - we were asked to glue it with a dark film.

And they were going to shoot all those who went inside. The hunt has begun!

Wonderful discoveries

After 2 weeks, we sat down with the personnel manager of the store and began to sort through the steps of who, how much, and how. Our task was to compare the passes with persons, so that we could calmly recognize people. That is, we have collected a database of photographs, and the personnel officer should have named everyone by name in order to enrich this base.

It went like this:

“I know that, Diamonds ... I know that, Bokayev.” Soooo, what the fuck is it? Who let him go? Mark, it is necessary to say protection. Oppa! Well, this is Ivanov, he has not been working with us for three months. What did he forget here? And who is this?
- Here are the data on access control. Wait, there are four people in general.
- Yes, this is Petrov. And who are the three with him?

In general, it turned out that:

Left people pass to the object every day, sometimes several people.
Employees go out for lunch sometimes with a pack of 4–8 people, and then one of them (the youngest), returning from lunch, puts the whole pack of passes to check in back to work. The brigade arrives after an hour 2–3 later.
These same young people sometimes in the morning (especially on Mondays) did the same operation with a packet of passes.
Already fired people calmly walked past the guards, taking advantage of what the guards remembered them (I don’t need to tell them what they did in their former place of work).
At the same time, the guard itself ineffectively sifted out the left — 300 people of their own personnel, contractors, suppliers (drivers, freight forwarders, etc.), persons who are not always “local” in appearance, that is, difficult to recognize by our domestic neocortex.

The average for an employee is the company’s cheating for 2 hours a day. 2 hours per person per day ... 2 hours, Karl! Our record holder for two weeks - one guy who cheated the company for 26 hours. He was checked in by his colleague for the first 16 - he just didn’t come for two days, and then he was late for dinner.

The result of our gatherings was such a report (for example, statistics for the week by employee):

You can click on each entry and get such a report per day:

The director looked at all this, paused, pondered, and then suddenly counted the amount and began to curse for a long time, ornately and very loudly. As a result, the system, even with all the seemingly expensive price, pays off extremely quickly (maximum in 1.5 years), and this, despite the fact that we will reduce the time loss by 70% (100% would be too ambitious statement), without any consideration of outsiders and other things.

Equipment

Cisco 6000P cameras (although we could have done with much more “simple” cameras, but tsiska were already in place). All this connects to the server, where there is a software for recognition - VisionLabs Luna. It was important to set a high shutter speed (slow shutter speed) so that the picture does not blur.

From the manual mode, we are now moving to automatic, that is, we are integrating with the access control system and their SAP personnel subsystem in order to immediately enter the faces. Now the faces are recorded on the first pass of one person on one new pass and are checked once a week manually by Eychar. In the new process, they will photograph people on acceptance in the state.

Empirically, we found that for good facial recognition you need at least 40 pixels between the eyebrows. Software doesn't care whom to recognize, it perfectly distinguishes twins, Chinese and other non-Russian faces (as opposed to a guard imprisoned for a European appearance).

False positives 0.5 percent. This is more than in banks on the webcams on their racks (there one millionth is considered the norm). But we have a real object and far from hothouse conditions. It is possible to reduce time in two due to selection of expensive equipment, but it is uncritical.

For correct recognition, you also need to take into account that our passengers love to wave their hands, hang out and generally welcome each other warmly in the passing area, so you need to put a small delay, otherwise there will be false recognitions like the situation when a person almost came out, then he saw a friend turned around. I took a few steps to him and said hello, and then I left.

General errors of comparing manual control with the personnel manager after another 2 weeks and automatic recognition - 6.5%. About 2-3% of our mistakes are there, the rest of the situations are purely domestic, which the security guards should record: for example, a person passed, but did not attach a card (we see two entrances, one exit). 2.6% comes from cases where there is no face, but there is an ACS - these are ours (except for emissions like guys in respirators and covering the face one way or another accidentally, plus a couple with huge fagals after the weekend).

Up to 2 thousand passes a day.

Why can not biometrics on hand

About how the fingerprint is transferred with scotch or mimic, probably, it is not necessary to tell. The jelly bear allows even the pulse to be transmitted and has the necessary structure for tricky sensors.

Biometrics in the palm of the hand is not applicable even in large offices - if a person eats at least something fat, then the whole venous pattern will change until the body wins the food (this is 7–8 hours of the other hand).

Total

We introduce in combat. Now on the left people are doing the alarm for the guard. The next step is probably the recognition of well-known shoplifters at the entrance, in order to recognize them all across the entire network, since it is quite easy to collect pictures after the fact.

To our some displeasure, the store owner showed the system to his European colleagues (unprepared), and now they are asking to consider the implementation of such a system in them. We wanted to show something after 2-3 months of combat tests - but no. By the way, they have a funny problem: they do not distinguish their migrants from the Middle East, who have the main cheap labor force.