near future
There are several methods by which face recognition systems work, but in general we are talking about a technology that can identify a person from a digital image or a frame from a video source.
Many smartphone owners use face recognition every day, but in mobile devices, the recognition speed is not critical, and the number of users is rarely more than one or two people. For office and street systems (with mass recognition), other technologies are used.
Recently discussed on HabrΓ© : Moscow chain coffee shops Pravda Kofe and OneBucksCoffee have begun testing facial recognition services in their establishments.
Coffee houses use our technical solution. And today we will tell you more about it. Of course, we have already talked about the technology itself, but something new has appeared - the solution has become truly cloudy. And this changes everything.
How facial recognition technology works
The first thing the system must do is to select a face in the frame and use algorithms to make sure that it is a human face.
After the initial detection, various individual traits are determined by fixed points - for example, the distance between the eyes and dozens of other parameters are taken into account.
Further, other algorithms are already searched for in various pre-created databases and give out the percentage of similarity with the desired data sample. If the percentage of similarity is high enough, the face is considered recognized.
Without going into details (the photo for analysis still needs to be normalized before being sent to the neural network that reads some descriptor), the main difficulty of the solution at the moment lies not in the technologies (algorithms) themselves, but in the implementation.
Recognition systems are developing in several directions, classified depending on the approach to information processing. Sometimes it is difficult to choose which system is best for a particular task.
Variety of systems
Data can be processed in the cloud, on local servers deployed within the enterprise security perimeter, or directly on the cameras.
In the latter case, all analysis is carried out by the camera itself, and already processed information is sent to the server. The main advantage of the system is its high accuracy and the ability to βhangβ a large number of devices on one server.
Despite the apparent simplicity and ease of scaling, this technology also has disadvantages. One of them is the high price. Plus, at the moment there is no single standard for the presentation of information that specialized cameras transmit to the server. And the data set can vary greatly from vendor to vendor.
"Simple" face recognition system from
Systems based on IP-cameras with the function of built-in video analysis are inferior in popularity to server solutions. But even in the case of using a traditional system based on a registrar and / or a local server, savings will not work.
Programs and prices* Face Recognition
*According to information from open sources.
Given the complexity of the algorithms and the high price of server hardware for video analytics modules, face recognition systems have long been an expensive pleasure.
Additionally, the cost of the solution is affected by the large network traffic generated during operation - in addition to the cost of powerful servers, we had to fork out for active network equipment and βthickβ communication channels.
Today, there are several major players on the Russian market offering high-quality algorithms for analyzing and processing video data. They are united by their interest in projects related to big business. It is very simple to explain this focus - the cost of the solution goes far beyond the capabilities of small and medium-sized businesses.
- ISS
SecurOS Face software.
The cost of a license for the face capture module is 41 rubles per channel. The software is installed on a face recognition server or on a face detection server.
The cost of a face recognition module license for 1000 people in the database is 665 rubles. Installed on the face recognition server.
- Sure
Russian developer of hardware and software for access control systems.
The cost of a license for a face verification module for one camera is 50 rubles.
The cost of a license for a face identification module for one camera is 7 rubles.
The price of a license for a database of up to 1 persons is 000 rubles.
- ITV
Intellect software for face recognition with a memory for 1 face standards in the database - 000 rubles.
The core of the system is 20 rubles. Connecting a video channel - 300 rubles.
- Macroscop
Macroscop Basic face recognition module with a database size of up to 1000 faces - 240 rubles.
License to work with one IP camera - 16 rubles.
More recently, Macroscop solutions were used to ensure the safety of only critical facilities with a large number of people: stadiums, airports, factories. But now the company supplies its product for retail. Price - 94 rubles for modules (registrars do not sell).
- TRASSIR
The software costs 79 rubles + 000 rubles for the registrar. The company's clients are mainly large firms (factories, mining companies, universities, sports complexes). But the main focus of the company is on traditional video surveillance, and not on facial recognition. Although their DVRs are great for these tasks.
- Findface
The company develops and sells only specialized facial recognition software. You will have to choose the configuration of servers for storing and processing data yourself.
- Video
A cloud-based video surveillance and video analytics service that offered services to businesses on a budget. Service works with almost any camera, the cost of connecting one device is from 3 rubles with an analysis of up to 150 unique faces per day and a basic record in the cloud archive for 100 days.
Hardware selection for Face Recognition systems
From one Full HD camera to process a video stream containing 10 faces in a frame, one processor core with a frequency of 2,8 GHz is required. If there are few faces in the frame (from 1 to 3), then one processor core can easily handle the processing of two video streams.
This example shows that even in a simple system, it is necessary to have a certain margin of hardware. After all, if not 10, but 15 people visit the object at the same time, then a second core with a similar performance will be required.
Therefore, for the operation of a traditional system, taking into account peak loads, it is required to keep double reserve capacity.
To make it easier for you to imagine how much a traditional face recognition system costs, we will take a point of sale as an example and calculate the cost of a traditional and cloud-based face recognition system.
Cost Calculation: The Cost of a Traditional Facial Recognition System
Let's say we are deploying a face recognition system in a pharmacy network consisting of 16 points. On average, 500 customers visit each pharmacy per day.
To fully recognize faces, one PTZ camera or a camera with a mechanized lens can be installed on each object of observation.
In the case of using the traditional system, the costs will be as follows:
- Each pharmacy will require at least one specialized video recorder. Its retail price is approximately 40 rubles.
- Each recorder will additionally require a special hard drive (not to be confused with a regular PC HDD) with a capacity of at least 4 TB in order to record a video stream in a resolution of 1920x1080 at high traffic intensity. The average retail price is 10 rubles.
- The budget should include the cost of maintenance of the video surveillance system (for example, the installer's visit to fix errors, update software or replace the HDD). The cost of such work is 12 rubles / year (once a quarter) for each object (in accordance with the price list of one of the installation organizations).
- The minimum cost of full-featured facial recognition software is an average of 120 rubles per camera (time-limited license).
- According to Backblaze, about 50% of all hard drives require replacement by the age of 6. Thus, after 5 years of continuous operation, about 8 disks will fail, and provided that such a system does not provide for redundancy, on average, additional costs of 1,6 disks per year, or 16 rubles / year.
Capital costs (excluding the cost of cameras) will amount to 2 rubles/year.
Cloud system costs
In the case of a cloud system, the cost of a video surveillance tariff with 500 face recognition/day will be 4 rubles/month (750 rubles/year) per camera, or 57 rubles/year for 000 cameras.
Recall that the owner of the network will not have to purchase any additional hardware. Maintenance costs are also not needed, because all cloud servers are serviced by a cloud service provider in the data center.
There is a saving of more than 3 times during the first year of operation of the system.
Subtotal and additional "buns"
There is an important nuance in the calculations above: after 3 years of operation, the traditional system in terms of total costs will become cheaper than cloud-based face recognition. There are two factors to consider here.
At first, the equipment that the owner of the network will buy will become obsolete in 3 years of operation. But for sure there will be new, more advanced technologies and face recognition algorithms that work on more powerful hardware. And in 3 years, most likely, it will be necessary to completely replace the equipment at the points.
You donβt need to do this with a cloud system - the service is constantly being improved and updated due to the development of algorithms and the growth of the computing power of data centers. Support for security standards is also not tied to hardware.
Secondly, saving money in the first years will allow you to wrap this money several times, bringing additional profit to the business.
The past, present and future of cloud-based facial recognition
The evolution of recognition systems has accelerated in recent years. Not so long ago, instead of complex algorithms and neural networks, an ordinary security officer using a computer simply compared the faces recorded by the program with the databases and noted who all these people were.
In addition, the systems worked through local servers. Accordingly, for the service to work, the user had to install a dedicated PC or a special DVR. And these are extra costs for equipment and overhead costs for its operation.
Cloud-based face recognition does not require the purchase and configuration of any other equipment other than cameras, and will work with those cameras that are already installed on the site.
No need to keep a staff of specialists to maintain the operation of the equipment. The problems of the technical condition of the equipment are solved by the service provider itself (and it does it more efficiently than non-specialized companies).
Cloud recognition turns a cumbersome and vulnerable system from local analytical servers into a flexible, fault-tolerant cloud structure. In practice, this means that the recognition system no longer depends on the capabilities of a particular server purchased and installed in the client's office, as well as the IT infrastructure that this client has. There is no need to purchase new equipment and long-term coordination with the supplier of configuration issues and the possibility of its expansion.
The cloud automatically distributes the load across the entire available infrastructure with powerful servers. The client does not need to keep infrequently used capacity in reserve for work during periods of unexpected load spikes (holidays, weekends). For more information about the capabilities of the system, see with us.
True Coffee and OneBucksCoffee have now caused a storm of discussion, but very soon there will be practically no companies in the offline business without video analytics. Players in the consumer market have an urgent need to recognize their buyer in person: personalize service and offers, analyze the mood of the guest, reduce costs and return customers, and not just buy technological solutions for the sake of reporting.
Source: habr.com