Marathon million people, face recognition system how to quickly and accurately complete the check?

Author | Ali entertainment technology expert Yin ink

Produced | AI technology base camp (ID: rgznai100)

Barley face gates at the 2019 Hangzhou International Marathon completed a successful debut brush face admission function, compared to the traditional marathon on admission admission verification program and experience admission efficiency has been greatly improved the following describes how barley Face recognition is supported marathon.

Marathon process description

Marathon's main process in three steps:

The first step, participants to the marathon official website registration, players will be notified after successful registration;

The second step, successfully registered players need to carry an ID card to receive an official site equipment;

The third step is to carry Directory race day admission ticket to compete.

Our challenge is:

1, how to complete within a specified time participants receive equipment work: it is necessary to ensure that the backlog quickly receive no harm to people, but also to ensure the equipment does not get the wrong;

2, how to ensure that tens of thousands of race day to complete the admission verification work in a short time.

Barley Face Recognition Solution

Before introduction barley recognition scheme, several commonly used terms first introducer face alignment:

1 to 1: 1 ratio of 1 means to compare with photos with people, judging by the photos and whether the algorithm is the same person, simple understanding is to prove "you are you."

1 ratio of N: 1 ratio N N refers to look in the personal photo library (bottom library), algorithmically determine whether the person in these photographs inside, is popular in terms of "who I am."

Clear process after the marathon, here is what barley Support site to receive equipment and verification of game day admission, admission in our verification program which is the most stressful day of the competition, the highest risk, but this is dependent on the first two step.

Players will not carry mobile phones and documents when admission game, that recount how it?

Barley traditional practices: communication with the organizer in advance, place a radio frequency chip, the chip number on the number-one correspondence with the directory in the address book, so by scanning equipment to recount the chip will check this information to the participating players, including photos (from step 1 and step 2), Why do we need pictures?

And then only to get photos to compare himself to confirm for himself or is not running, which uses face recognition, which is mentioned above than 11 comparison.

This program is feasible, but faced with two problems:

1, the Directory is responsible for third-party companies, often encounter the phone book there is no number on the chip or chip number is inconsistent with the Directory (the actual incidence is still relatively high), which will result in the player can not directly verify admission, require manual processing;

2, this program requires admission to verify the chip, and then face comparison, there are some powerless in the marathon, as tens of thousands of players need to complete the admission in just half an hour, the pressure can be imagined .

How to optimize athletes admission, both to avoid the problem of the Directory of the chip, but also fast and accurate verification?

It uses face recognition than 1 N, we get ahead of photo all players, players direct brush face for comparison, rapid verification of admission.

This optimization program needs to do two things:

1, the safety of people face alignment algorithm, the people who go through the release, should not go in person to refuse floodgates;

2, ahead of all the players get photos and photo quality needs to be good enough, as a substrate libraries face alignment

Face alignment algorithm this requirement is more stringent, marathon every year for a lot of people running, we face the algorithm must be able to find out for runners, and tickets barley recount scenes like the concert at every turn thousands, is never allowed to make without a ticket of admission to anyone, of course, and let people be able to buy a ticket through the normal, such a scene and a scene more like a brush face payment, so we use a sophisticated level of financial face algorithm.

Get photos of all players, we are led through the link registration and equipment described previously addressed:

We will ask the players to submit a photograph of himself at the time of registration, but this does not guarantee all submitted sufficient quality photo, and the photo is my submission. This requires time to resolve Receiving equipment in the field, as this could affect the normal smooth admission time to the game.

Here's how we get to photograph players receive at the completion of the equipment at the same time:

We all know that when the ticket is required for admission tickets, the players in fact, his identity is a ticket when using the equipment ID collar, players use barley gates brush ID card system, the system will be based on identity cards number of players to read the information, and then will print out a small ticket through the gates of the printing system, the players holding a small ticket window to receive the corresponding entry equipment.

But this does not meet the requirements to host, you must require players to hold my ID card to the site to receive the equipment, of course, can not meet our own needs - to get photos of all players.

So we have him for a ratio of 11 players when brush ID card, and then confirm that I have to collect a user's site photos, which completed the acquisition identification and photographs.

In this scenario, the receive equipment, this process can actually be optimized, user-submitted photos in the first step registration there are a lot of good quality can be directly used, there is no need to live to make another acquisition, for this part of the user we then have their pictures will guide them through the brush directly than 1 N face to face than the small print admission tickets were receiving equipment, the advantages of this program are:

1, direct brush face, no need to brush ID + 1 to 1 ratio of + collection, let receive higher efficiency and avoid queuing site

2, solves the problem many people face the uploaded users forget to bring ID card or ID card is lost due to not receive equipment

Summary: We are guided by the time registration of players to submit photos through the field guide allows players to submit photos of direct brush face Receiving equipment for brush face recognition fails and the players were not submitted Face conducted himself confirmed by more than 11 and face collection, thus ensuring that all players have the photos to our face end database, it allows us to make players on game day can directly use than 1 N ratio of direct brush face admission.

Protection of user privacy and data security

Our program is based on face recognition admission of the user's photo, which involves user privacy, we will notify users in advance when marathon registration, authorized users, we need to do more is how to ensure that users photos of safety. Here's our plan:

Gates need to show identification to the user after the user's photo, in order to protect the user's photos will not be leaked, we do read the judgment of the authorized user's photo interface, unauthorized devices can not obtain photo data, only authorized devices to get to photo data, and that has been encrypted to ensure that the data will not leak from the end.

When user photo collection site, we will collect the photos once again after the encryption operation are encoded after encryption will be deleted immediately after local photos uploaded to the server, uploaded successfully. The server will automatically delete all photos after the end of the game.

By the above operation, ensure that the user's photo data security, it protects the user's privacy.

Face gates Software Architecture Evolution

We finished our face recognition program, again below explain how our software architecture is to support our face recognition admission:

Our gates are used in the Android system, developed gates face verification system with common Android application development is slightly different, we need to be adapted for the hardware: the movement (swing gates, roller gates), a printer, two-dimensional code module, the RFID module, a module ID.

That software architecture throughout the gates of it is natural to look like this:

Such a scheme is no problem with it, but a long time will find several questions:

Whether it is business code changes or hardware driver code is required for the entire application upgrades
With the increase in gates models, the program will have various types of drive, leading to hardware-driven code bloated, difficult to maintain
Adaptation on the hardware, since the code together with the business, we can only do it by yourself, not to the gates of vendors

So on this basis we have optimized the software architecture: separate business processes and drivers, using the idea to program to interfaces, business processes commands to the driver through AIDL way.

Such benefits are clear:

Business and driver code is no longer coupled together, independent, business applications only care about business logic, hardware drivers drive applications do things
Business processes without concern for how the driver, the driver can achieve handed over to manufacturers to achieve, we set standards like
Business programs and drivers independent upgrades, on-demand upgrade
Each type of device uses a unified service program, only install its own drivers no longer need all the different types of device driver code together

Face recognition capability optimized

We use a secure face recognition algorithm, but this does not mean that the algorithm can solve all the problems of our field, barley field environment than the face brushes pay more complicated scenario, here we are how to optimize and improve our face recognition ability:

1, the library is increased to solve the bottom && reduced error rate

Research extraordinary face algorithm students should be clear, then the perfect face there will be a case of mistaken identification algorithm, that is, error rate, and with the bottom of the library becomes large, the error rate rises will follow. Marathon scene tens of thousands of people is normal, it is necessary to reduce the error rate but also to meet the bottom becomes large library, a seemingly contradictory problems, and algorithms can not solve the current level, but must solve business problems, our idea is : Since the algorithm can not be resolved, it would only resolve to go through the business, we know that the marathon is also broken down into full-horse, half-horse, lovers run, mini-run, home run, such as different species, then we can use these different categories will be separated from the crowd, so that you do reduce the bottom of the library, it will naturally reduce the error rate.

2, face data to get through the whole process

We said above through the registration site or site collection face brush ID cards were collected at the scene, complete facial capture, but also the case that neither the user to upload photos at the time of registration, there is no carry an ID card, the user only through barley APP acquisition, how to ensure that after the acquisition of APP can receive admission equipment immediately? It would need to get through the entire process, as follows:

3, the dynamic remote tuning

Marathon scene recognition scene actually paid complicated than the human face, most people face is paid in the room, the light will affect a relatively small number, but the marathon is necessary to consider the indoor scenes (receiving equipment) will have to consider outdoor scenes (open run admission), excessive impact faces, side faces the backlight and the like of an exposure distance, and therefore need to be tuned for specific environments, we impact face by dynamically modifying the algorithm ends at various management parameters, and whether to adopt downgrade programs, remote issued to all field equipment to ensure smooth and efficient players admission, does not cause the problem of long queues.

to sum up

According to our marathon actual business scenarios, the face recognition function 1 to 1 and 1 to N applied to the marathon, opened a new mode of admission Marathon industry, but we have a long way to go, face the algorithm needs to continue lift: the increase recognition rate and reduce the error rate, the optimal allocation algorithm face different environments servant, indoor and outdoor different ticket program and has begun commercial 5G can give our business to bring new breakthroughs, we will the face recognition function applied to more verification scene, continuing to enhance traffic efficiency and experience admission site users.

【End】