CVC neural network video analytics implementation in a logistics center case

On the client's side, the development department interacted with us on this case. The department consisted of core employees from various services of the company, including logistics. The task of these employees was to voice the potential problems of their departments. After that, the department's specialists were responsible for implementing the developed technology, implementing pilot projects, and, as a result, confirming that the chosen technology was capable of localizing or completely eliminating such pains.

The pilot project was launched at a site with the following characteristics:

• number of cameras: 2,400 pcs;
• number of gates for receiving shipments: 120 pcs.

The neural network robots had to determine exactly how many places were loaded into a particular vehicle, compare the data to the accounting system and identify deviations.

Imagine a situation: a car with 20 pallets has driven into a facility. However, the warehouse only needs 15. At the same time the first 5 are blocking the unloading of the required 15... This is not terrible at all, just forced to unload the first 5 places to free up the possibility of unloading the next 15.

But at this point, the following can happen: not all 5 unloaded back or unloaded back 6 of 5, grabbing something extra. These mistakes can be made without malice, but in the case of "manual" control there is a need to search for hours for this or that shipment. And it really can take hours, because the operator has to personally watch numerous videos, not knowing where and how to look for a lost pallet. And given the number of cameras and gates on the site, such a small task can take from several hours to several days.

The task of the robots in this case was not to prevent an error, but to promptly report at which gate and with which machine an incorrect backload occurred. This way the search for the load took not days, but seconds, as the robot instantly showed a video of the failure: what had been loaded incorrectly and where it had gone.

The problem appeared in situations where goods were being prepared for packaging. We were required to connect all these processes with the customer's various video surveillance systems.

In addition to the specificity of the accounting systems themselves, it is important that these systems were not one, not two or even three. Note that this is quite normal for a company that performs complicated logistics and complex processes. We would venture to guess that this was the main problem that other contractors could not handle. As a result, they inflated the price tag, and some even abandoned the project.

Note that in the warehouse, as a rule, there is already some kind of video surveillance system (what kind, for us, does not matter). In the case of this customer there was everything at once and in different parts. Different processes have different types of equipment and different types of software that work with these videos. The task for our system was to instantly (on demand) demonstrate the whole path of the goods within the logistics facility.

As a result, if a customer informs us that a product has been crushed, it is easy for our customer to find out where this problem with the product occurred.

This will not require, as before, to manually enter each system and look for that product at different points in time in the archive of video recordings. Now it is enough to enter the box number or pallet number (any entity that exists in the warehouse) and get instant access to the video history of this entity's movement through the warehouse, where any contact with the personnel took place.

We are talking about a special part of the warehouse, where, as a rule, there are small and expensive goods. Working in this section of the warehouse, employees receive a task on their PDT (company WMS), which describes in which section of the warehouse and in which cell the employee should go to perform the task.

From the point of view of the average person, there should be no problem, but in practice it is not so simple.

The specialist went to the wrong section and, accordingly, begins to perform the wrong actions. In a good way, the employee should be prompted about this by the PDT itself, which will immediately block access. But this will only happen if this system is informed in time about all manipulations (what you took out, where you put it).

A specialist receives a task in one section, but deliberately goes to another one, where he commits illegal actions (for example, puts a couple dozen iPhones under his clothes or eats a kilo of expensive nuts).

To solve this group of problems, we used our main trump card, namely neural network robots. They get information from the accounting system and understand which employees have what tasks. This instantly makes a comparison - where are such employees, are they not in sections where they do not have tasks...

Moreover, if an employee is delayed in a certain section, the robot triggers an alarm (of course, in its own language) for further analysis. Well, and then the appropriate official examines the video, calls the employee for "questioning" and, having full information, makes a management decision.

Our client, like a number of other companies of this level, uses the services of outsourcing organizations to "supply" personnel. Every day a different number of appropriate personnel come to the sites. It is reasonable to handle personnel matters in such a way, but it may cause difficulties with the fair calculation of wages.

The fact is that the cost of outsourcing company services is determined by the number of man-hours. But in fact the number of people and hours worked may be much less than was agreed upon. In practice, this happens all too often. For example, one person is registered several times (as a result, the company believes that 100 employees went out for a shift, but 90 actually work). There can also be another situation, when a person enters a facility, and then leaves on his own business. Outsourced workers are not always interested in a full day's work, and often, if possible, slack off. Tracking several thousand personnel with the help of simple video cameras is impossible in practice.

Our real-time facial recognition was used as a solution to this problem. The neural network robots recorded absolutely all incoming and outgoing people, determining the number of unique faces. Each face was linked to a specific outsourcing company. The productive time of each employee at the facility was recorded separately. Thus, if someone left the site, did not work enough, left work early - all this rightfully reduced the outsourcing company's check.

Every company has its own specific business processes. So from our side to go deep into client's processes is not only necessary but important. The client in this case was extremely ready for this. As a result, the third group of tasks, related to the outsourcing companies, arose in the course of work, when at one of the meetings we learned about this problem and offered to solve it.

Sometimes the approval may take a year, and sometimes it is fast. At any company, especially a large one, all information is confidential and usually it is very reluctant to be shared with outside companies. It is this point that always requires an explanation from our side.

In this case, the approval lasted a quarter. The main question is how the access to video will be organized on the pilot (who uploads it, what channels are available, etc.).

Working with the camera angles which the customer already had, we prepared models and conducted docking. The first docking revealed some strange things about the customer's database. It took us several weeks to figure out where some of the data was lost and why it was happening.

The pilot project was implemented with limited functionality (the program worked on two gates). This result was enough for the customer to see what we had agreed to show initially.