Detection of criminal groups stealing from stores - Data Mining

A group of 3-4 persons, by prior agreement, is able to take out goods from the big store of the “Auchan” or “Crossroads” type up to 400 thousand rubles per month. If ordinary shoplifters just peacefully steal sausage, dragging it under clothes or elsewhere, then these guys are at times more impudent and destructive.

That's the difference. First, they have the ability to confuse the entire surveillance system built to search for single thieves. The simplest bundle - one takes the goods, passes imperceptibly to another, he takes them into the blind zone of the cameras, where the third takes them.
')
Secondly, they do not steal to the administrative limit of 1000 rubles, but immediately to the maximum for resale.

Thirdly, when they try to stop and call the police, they use force to the guard and leave. Force measures are from banal driving off guards from the suspect to weapon threats.

We were approached by the PSC, which asked to make a video analyst on those whom they already know. So that at the entrance to the store of such a comrade, the police would immediately be called up and they would be taken warm.

Further in the examples and descriptions some details are changed upon the request of the security men. In the photo for an example interface, I do not need me to face the floor.

Where are the blind zones from the cameras?

In many large chain stores, cameras are arranged along a planogram in the necessary and sufficient quantities. That is, the most optimal coverage, butt on the absence of blind spots. Then, if the display changes or the rack moves, a blind zone is formed. In her would buy more cameras, but the budget for this from the store is local, and the process can take 2-3 months.

Where do the criminals know where the cameras are?

Because in the Russian Federation it is forbidden to mount hidden cameras for such purposes. They just see them.

Why can't they just be arrested at the exit?

Firstly, it is difficult to prove that it was the guy we are holding that committed the crime - or rather, it is difficult to calculate it, because you want to arrest someone who took the goods from the shelf. And it is empty. Secondly, they know how to use store spikes. Thirdly, presumably, sometimes they have distracting members of the organized criminal group, who clearly steal a little (at the administrative level), but they take on protection. Fourth, and most importantly, the guard has no right to detain. They can only block the exit with their bodies and call the police. Touched man - sat down. The police travels for a long time, and during this time accomplices manage to save their colleagues. Well, about the threat of weapons, I did not just write.

In general, continuous attack. Especially if you remember that in grocery somewhere on the outskirts there may not be a very big change of security at night. 2-3 guards are easily neutralized by two or three armed men.

What about personal data?

The secret is in a one-way algorithm. It is necessary to keep such a perceptual hash from the face, so that it is impossible to restore a person according to the hash, but it was possible to get the same hash for a person. There are a couple of legal tricks and details about the fact that you can not store a number of identifiers. As a result, the laws on PD are met. The FSTEC confirmed that the video recognition system does not store personal data.

How long does it take to steal?

OPG work very quickly. Come and spend on theft is not hours, but from 5 to 20 minutes. Judging by the records, they act directly according to the algorithm: where and when to go, what to take. Probably plan, as in the films before the bank robbery.

If you see an organized criminal group in advance, you can have time to call the police and prevent it.

The second option is psychological pressure. You have a group of people, which is potentially an organized criminal group. The guards go after them and crush psychologically - there were facts that people discovered observation, turned around and left.

Usually such organized crime groups work with backpacks, therefore, when they try to cheat you in the store, please treat with understanding. I used to be very offended and sent guards to three letters, now I understand why they are so nervous.

Video analytics

Video analytics allows you to mark our heroes and give an alarm when they appear. The important thing is:

When we see an already known character (he is impersonal in the system due to the processing of personal data, for example, in the form of the profile “Damn 72” or “Ocean's Friend 7”, a comment and a list of atrocities), we can guess that he is not alone , and came from all OPG.
Since the model of such theft is to enclose about five stores per day (guys drive to different points), after the first affected point, profiles of people can be entered into all network surveillance systems.
Enough ordinary HD cameras.

Following several months of operation, for certain reasons I cannot tell the details. But let's try to tell you what I have the right to:

There is an intuitive assumption that such organized crime groups are beginning to keep clear of the stores of the protected network. That is, they go to steal to those who are less protected.
There is an intuitive assumption that those who go insolently will remain. We can crush them by observation or by calling the police. Then they get tired of walking in vain.
There are those who believe in the incredible memory of the guards and collect a new jazz band, and then begin to tour the other shopping centers. Since the base is distributed and they do not change the entire composition at once, they are visible.

Hold on, bastards!

How does it work?

Source - the source of data for video processing. The source of the video stream can be IP-cameras, video surveillance server, video file.
Video stream processing server — processes the video stream and receives from the stream frames of persons of visitors of optimal quality in accordance with the specified conditions.
Query server — receives a frame of the visitor’s face and generates requests to the recognition server to create a descriptor (extract) and compare descriptors in the specified list (matching). Query server provides work station, as well as integration with external systems.
Video Recognition Server - performs processing of requests for extracting a descriptor from a frame of a visitor’s face and comparing the received descriptor with the person descriptors in the database using the specified lists.
Automated workplace of an employee is a component that provides user interaction with the system, the ability to configure it and obtain the result of work.

Funny

While setting up the system, calculate at the same time a couple of bright comrades. For example, they found a man who had come to the store for a few months to drink whiskey. He takes a bottle, then a good cheese, goes to the blind zone and reinforces his strength there. The guards called him “Sir.” He simply rolled up the bottles under the shelves, then found them and put the battery in the office of the head of the security guards. For the 10th time he was caught. He drank a little, but got everyone. He inflicted damages in retail prices by about 7 thousand rubles, but it took the security officers a total of 3 days to analyze the situation. And the last three times he came, already when they began to analyze his case, they spent 4 hours a day. That's what lack of automation does.

One 14-year-old boy brought out 7 laptops as part of an organized crime group. We expanded the sample of recognition by childish faces and achieved very high accuracy in this segment.