+ All Categories
Home > Documents > Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2....

Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2....

Date post: 12-Jul-2020
Category:
Upload: others
View: 2 times
Download: 0 times
Share this document with a friend
40
Page 1 Koala Face Recognition Access Platform Product White Paper Document Version V3.1.0 Date of Release July 25, 2019
Transcript
Page 1: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 1

Koala Face Recognition Access

Platform

Product White Paper

Document Version V3.1.0

Date of Release July 25, 2019

Page 2: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 2

All rights reserved.

Without written permission of this company, no organization or individual is allowed to transcribe,

reproduce, or distribute in any way, in whole or in part, the contents of this document.

Note:

Any product, service or feature you purchase from Beijing Megvii Co., Ltd. is subject to the company's

commercial contracts and terms, and all or part of the products, services, or features described in this

document may not be included in the scope of your purchase or use. Unless otherwise agreed in the

contract, Beijing Megvii Co., Ltd. makes no statement or warranty, express or implied, regarding the

contents of this document.

The content of this document may be updated from time to time due to product version upgrades or

other reasons. Unless otherwise agreed, this document is provided as guidance for use only, and no

statements, information, and recommendations in this document constitute express or implied

warranties of any kind.

Page 3: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 3

Contents

1. Product Overview ................................................................................................ 4

1.1. Product Positioning .......................................................................... 4

1.2. Application Scenarios ...................................................................... 4

1.3. Product Composition ....................................................................... 7

1.4. Core Technology ............................................................................. 8

1.5. Product Highlights............................................................................ 8

2. Product Details ..................................................................................................... 9

2.1. Business Architecture ...................................................................... 9

2.2. Host Cascade ................................................................................ 10

2.3. Algorithms ...................................................................................... 10

2.3.1. Face Detection...................................................................................... 10

2.3.2. Face Analysis ....................................................................................... 10

2.3.3. Face recognition ................................................................................... 13

2.4. Software ........................................................................................ 14

2.4.1. Web Page and Functions ..................................................................... 14

2.4.2. Main Business Processes ..................................................................... 22

2.5. Hardware ....................................................................................... 24

2.5.1. Terminal Devices .................................................................................. 24

2.5.2. Host ...................................................................................................... 33

2.5.3. Optional accessories ............................................................................ 37

2.6. Authorization .................................................................................. 38

3. Successful Cases .............................................................................................. 39

3.1. Ant Financial Services ................................................................... 39

3.2. Ucommune .................................................................................... 39

3.3. More Customers ............................................................................ 40

Page 4: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 4

1. Product Overview

1.1. Product Positioning

The Koala Face Recognition Access Platform is a smart face recognition product

designed to provide the best access experience. The platform supports both new

deployment and installation based on any existing all-purpose card, and allows you to

upgrade from an all-purpose card to access based on facial recognition.

Compared with traditional products, Koala adopts computer vision algorithms such as face

detection, quality judgment, liveness detection, attribute analysis, and face recognition

based on deep learning methods. This provides the advantages of high efficiency and

precision in various application scenarios, such as 1:N face recognition gates and 1:N face

recognition access control and welcome systems.The dynamic and non-cooperative face

verification method is more convenient than the traditional card-swipe access method. In

addition, thanks to the uniqueness of facial features, this method also solves problems such

as forgetting to carry ID and fraudulent use of another person's ID.

1.2. Application Scenarios

Koala Face Recognition Access Platform can be used in commercial buildings, factories,

enterprises and institutions, college dormitories, and residential areas for a wide range of

scenarios, such as access control, attendance, and meeting sign-in.When used in

combination with the backstage management system, the platform can perform functions

such as personnel identification, VIP welcome service, staff attendance, visitor

registration, stranger alert, daily data statistics, and inquiry.

Page 5: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 5

Dome camera at building gate linked to elevator

Enterprise welcome

Page 6: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 6

College dormitory management

High-density face capture for factory attendance

Page 7: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 7

Visitor registration

1.3. Product Composition

The core of the Koala Face Recognition Access Platform consists of the cloud and terminals.

The cloud has management and computation functions and is referred to as the “host”.

There are multiple host configurations available for different scenarios. Terminals can be

divided into three types, video stream, capture, and recognition, and are provided in

different forms. Terminals can be combined with the host to implement all face

recognition access features. Please refer to the hardware section for item details.

Page 8: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 8

1.4. Core Technology

Leading algorithms: Accurate recognition of 200,000 people with a speed of

<0.5s; two-eye activation technology, 100% defense against electronic screen

attacks, 99% defense against photo and face model attacks; fit for all races and

all ages.

Leading ISP: Face recognition cannot be separated from the front-end optical

imaging. Megvii optimizes ISP algorithms for all kinds of complex light scenarios

such as dim light, backlighting, and direct sunlight, to achieve an optimal access

experience in all scenarios.

1.5. Product Highlights

Multimodal Integration

The actual access scenarios used by customers are highly diversified, such as passages, gates, access control devices, high density entry, and outdoor access. It is difficult to cover all scenarios with a single product format. Koala's front-end sensing devices, which include box cameras, dome cameras, access control machines, and high-density face capture, can be used outdoors and meet all customers' scenario needs with a single product.

Fast, Accurate, and Safe

Non-contact recognition in less than 0.5 seconds allows for unobstructed access.

The false recognition rate is less than 0.1%, while the pass rate is over 99.5%.

Two-eye activation technology effectively protects against attacks using photos, videos, and masks.

OPEN APIs

Koala is a face recognition access platform that provides APIs to customers. Customers may integrate Koala into third-party systems, such as any existing all-purpose card management systems, attendance check systems, or mobile apps. Integrators can use Koala to build larger solution platforms, such as park management systems, dormitory management systems, and conference management systems.

Page 9: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 9

2. Product Details

2.1. Business Architecture

Koala consists of the host, terminals, and accessories. Video streaming cameras, face

capture cameras, and recognition devices (access control devices) can be used with the

host, but different types of terminals cooperate with the host in different ways.

The host provides Web Server to implement business management functions such as

library management, access control, attendance check, and welcome. The host can also

manage network switches, TV boxes, and other accessories to open doors via network

relays and display welcomes in pop-up windows. The host also provides APIs to provide

greater convenience to customers.

Page 10: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 10

2.2. Host Cascade

Each host has a clear load capacity. When a host cannot meet a customer's needs,

multiple hosts can be cascaded, so that one host serves as the cascade master and

other hosts serve as cascade slaves.

2.3. Algorithms

2.3.1. Face Detection

1. Effect

Detection rate > 99.5%

False detection rate < 0.1%

Track interruption rate < 1%

2. Performance

Face detection < 50ms

Face tracking < 10ms

2.3.2. Face Analysis

2.3.2.1. Quality Judgment

1. Effect

Storage success rate > 99%

Horizontal angle error < 3°

Vertical angle error < 3°

Rotation angle error < 3°

Fuzzy error < 0.1

Page 11: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 11

2. Performance

Storage speed: Access control device < 300 ms

Server CPU < 200 ms, GPU < 100 ms

Face angle < 10 ms

Face fuzziness < 10 ms

2.3.2.2. Liveness Detection

The access control device supports two-eye activation, and the server supports

single-eye activation.

Algorithmic

Model Definition Definition description

Liveness Liveness Detection

Detects whether a human face is

a real person, a mobile phone photo,

or other attempts to trick the system.

Parameter

name Parameter definition Remarks

Pass rate

Positive sample set = Real

human faces

Pass rate = Number of positive

samples detected as real

humans/total number of positive

sample sets

The pass rate is affected by the

threshold.

False

recognition

rate

Negative sample set = Unreal

faces (photos, phones, etc.)

False recognition rate =

Number of negative samples

detected as real humans/total

number of negative sample sets

The false recognition rate is also

affected by the threshold.

The effect of liveness depends on a number of factors, including the picture quality, site

environment, and liveness threshold.

In particular, it should be noted that the ability of liveness recognition to prevent attacks

using different materials is different. The following table gives the liveness false

recognition rate for common materials when the pass rate of a real person is 99.95%.

Page 12: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 12

Attack

material

False

recognition

rate under

dim light

False

recognition

rate under

normal light

Printed paper 0.05% 0.10%

Coated paper 0.20% 0.18%

Human image

matting paper 0.59% 0.18%

Non-human

mask 0.97% 0.93%

Photo 0.33% 0.40%

3D head model 2.01% 1.45%

Two-eye access control devices can achieve 100% defense against attacks using

photos and videos on mobile phones, computers, and other devices, i.e. 0 false

recognitions.

Page 13: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 13

2.3.3. Face recognition

Algorithmic

Model Definition Definition description

1:N

Recognition Face recognition

Find the most similar person

from the N people in the

library and give the

corresponding score.

Parameter

name Parameter definition Remarks

Pass rate

Positive sample set = Sample of target

personnel within the library

Pass rate = Number of positive samples

of the target personnel passing by

positive recognition/total number of

positive sample set

Positive recognition = First

person matching the target

person in the library with

a score higher than the

threshold.

Therefore the pass rate is

affected by the threshold.

False

recognition

rate

Negative sample set = Sample of target

personnel not within the library

False recognition rate = Number of

negative samples of target personnel

passing by false recognition/total number

of negative sample set

False recognition = First

person matching other

personnel in the library with

a score higher than the

threshold.

Therefore the false

recognition rate is also

affected by the threshold.

Page 14: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 14

The table below provides the pass rate and false recognition rate when using different

library sizes in a general scenario for access control device and server platforms.

Platform Library

size Test set

Pass

rate

False

recognition

rate

Access

Control Device

10,000

people

Simple set 99.9% 0.5%

Complex set 99.5% 0.5%

20,000

people

Simple set 99.6% 0.5%

Complex set 98.7% 0.5%

Server

100,000

people

Simple set 99.9% 0.5%

Complex set 99.5% 0.5%

200,000

people

Simple set 99.5% 0.5%

Complex set 98.5% 0.5%

Test set ratio:

Age coverage: 15-60; ages of 15-30 account for 85% and those of 31-60 account for 15%

Gender ratio: 1:1 for men and women

Lighting, makeup, and accessories:

o Simple set: normal light, no accessories, and no makeup

o Complex set: Four kinds of evenly distributed light (normal

light/backlight/dim light/overexposure) and four kinds of evenly distributed

accessories (no accessories/ordinary glasses/hat/bangs) with a female

makeup rate of 5%

Speed of single face recognition (including detection, liveness, and recognition):

- Server CPU < 400 ms, GPU < 200 ms

- Access control device (RK3288) < 500 ms

2.4. Software

2.4.1. Web Page and Functions

2.4.1.1. Koala Customer Management System

Website: http://192.168.1.50/admin2 (factory IP)

Page 15: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 15

This system is used by the technical support team or IT O&M team to provide admin

backend management for Koala customers. The website provides information such as

the customer directory, directory of hosts bound by customers, host firmware versions,

and upgrade management information. When a new customer is connected to the

customer management system, the technical support team will add the customer

information and its bound host and firmware version to the system.

2.4.1.2. Koala User Management System

Website: http://192.168.1.50 (factory IP)

This system is used by Koala users to carry out access-related management. The

functions of the website include:

1. Library entry

Page 16: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 16

2. Recognition record

Page 17: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 17

3. Access control and welcome settings

4. Attendance statistics

Page 18: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 18

5. Stranger pop-up

There are three options for the “stranger pop-up selection” function in welcome

management of the user management system: no pop-up (default), pop-up without

prompt tone, pop-up and prompt tone.

When the subject is recognized as a stranger, the TV app pops up a stranger pop-up

window and sounds an alarm at the same time.

“Alert when stranger identified” in the Web -> Access Control Settings has a higher

priority than the TV app and welcome management, that is, if the stranger function is

disabled in the access control settings, no stranger pop-up will appear in the TV app.

Page 19: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 19

2.4.1.3. Koala Host Configuration

Website:http://192.168.1.50:8866 (factory IP)

Koala host configuration is used to view host-related configuration and information and

provides the following major functions:

1. Display the host status: IP, parameters, free storage space, log, binding status,

access control information, identification callback settings, NTP settings, etc.

2. View recognition result and check whether the recognition function of the host is

normal through commissioning

3. Display information about various devices such as cameras and switches bound to

the host

4. Restart the host remotely

Page 20: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 20

2.4.1.4. TV App

The TV App is mounted on the TV box to display the recognition interface. It has the

following major functions:

1. Display the welcome interface (time/weather/video stream)

2. Display pop-up windows for recognition results

3. Change the interface theme

4. Link host(s) and camera(s)

2.4.1.5. Access Control Device App

The Access Control Device app contains a standby interface, a recognition interface,

and an engineering interface.

Page 21: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 21

Standby interface:

If no one passes within the set time, it will enter the standby interface. If someone

approaches, it will automatically jump to the recognition interface.

Recognition interface:

The recognition interface displays relevant prompts to guide users.

Page 22: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 22

Engineering interface:

The engineering mode includes four parts: access control device settings, recognition

settings, account settings, and other information. Users can modify the default settings.

2.4.2. Main Business Processes

2.4.2.1. Add company

In Koala customer management system, select Create new company on the customer

management tab.

2.4.2.2. Add host

(1) Connect to the host.

(2) View the host directory in the Koala customer management system. After the host

is properly connected to the network, it will automatically identify the host's IP and

token in the host directory and show whether the host's connection status is

normal.Find the configured host and record the IP and token.

(3) In the Koala customer management system, click the new company name to enter

the customer's host list. Select Add host, and select the host whose token was

Page 23: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 23

recorded.Multiple hosts can be added here.

(4) Go back to the host directory and view the host just added. You can see the

bound company. Click View to see the host configurations. Here, the host version

is the version currently used by the host. The optional version is used to when the

host firmware is changed. The list is the firmware list in the current system, which

can be added in the host firmware version. After selecting the version, click OK

and the host will update the version.

2.4.2.3. Select camera device

(1) Use the system account added to log into the Koala User Management System.

(2) Select Access Control Management -> Access Control Device -> Create. You can

add cameras and control switches. The host drop-down list includes all hosts

bound to the company, allowing you to determine the host corresponding to the

camera and switch.

2.4.2.4. Add library

(1) In the Koala User Management System, select User Management -> Employee

Management -> Create New Employee and upload the employee photo and name.

(2) You can batch import information, and import zip packages, pictures, or tables.When

batch importing pictures, you can choose whether they are automatically associated

with employee numbers, names, mobile numbers, and other information.

2.4.2.5. Access control permission settings

(1) Go to the Koala Access Control Management -> Permissions Management, click

Add new access control rule, and enter the name of the access control.

(2) Configure the corresponding personnel, access control group, period, and

holidays for custom access control rules.

2.4.2.6. Add attendance

(1) Go to the Koala User Management System -> Attendance management and

select Attendance Statistics. The list will show all the employees added.

(2) Go to Attendance Settings and set the attendance time.

Page 24: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 24

2.5. Hardware

2.5.1. Terminal Devices

2.5.1.1. Dedicated face recognition cameras

Dedicated face recognition cameras, using highly optimized image parameters and face

recognition algorithms, provide high imaging quality and a compact and beautiful appearance.

Model MegEye C2

Target size 1/2.8in

Dimensions 62x62x88 (mm)

Maximum resolution

1920x1080

Maximum frame rate

1920x1080@25fps

Shutter speed 1/10 sec to 1/1000 sec

Output video bit rate

100 Kbps to 8 Mbps

Minimum illumination

[email protected]

Image settings Auto exposure, auto white balance, color correction, brightness, contrast, saturation, sharpness

Network protocols

TCP/IP, ICMP, HTTP, FTP, DHCP, DNS, DDNS, RTP, RTSP

Network management

Remote setting via client or browser

General functions

Restore to factory default, anti-flicker, heartbeat, mirroring, password protection, video masking, watermarks, IP address filtering

Power supply DC12V±10%, supporting POE

Power consumption

4W MAX

Operating temperature

-30°C to 60°C

Operating < 90%

Page 25: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 25

humidity

2.5.1.2. Backlight dome camera for face recognition

These dedicated face recognition dome cameras have an ultra-wide dynamic range and

can adapt to various complex light scenarios.

Model MegEye-C3V-32J-X05

Camera

Sensor type 1/1.8” Progressive Scan CMOS

Electronic

shutter 1/25s to 1/10000s

Minimum

illumination

Color [email protected];

Black and white [email protected];

Signal to

noise ratio ≥50dB(AGC OFF)

Wide dynamic

range ≥120dB

Noise reduction 3D digital noise reduction

Lens type 5mm fixed-focus lens with M12-mount

Compression standard

Video

compression

standard

H.265+/H.265/H.264/MJPEG

H.265

coding type Main Profile

Video bit rate 32 Kbps to 16 Mbps

Audio G.711a/G.711u/G.726

Page 26: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 26

compression

format

Image

Image settings

Adjustable brightness, contrast, saturation, sharpness, white

balance; AWB (Automatic White Balance), AGC (Automatic Gain

Control), AE (Automatic Exposure), AIC (Automatic Iris

Correction); Supports private area occlusion; OSD overlay;

Image rotation: Normal 90 ° /270 ° automatic, manual or

automatic color to black and other switching methods

Maximum image

size 1920x1080

Video resolution

Main

stream 1080P(1920x1080)/720P(1280x720)

Sub stream D1(704x576)/VGA(640x480)/QVGA(320x240)

Video frame rate 50Hz: 25fps (1920x1080, 1280x720)

60Hz: 30fps (1920x1080, 1280x720)

Character

overlay Supported

Smart functions

Smart coding Supports low bit rate, low latency, enhanced coding of ROI area,

up to 4 ROI areas

Smart detection

Supports occlusion detection, defocus detection, brightness

detection, color detection, sound detection, forgotten object

detection, movement detection, virtual cordon, area intrusion

Smart control Supports intelligent control of alarm on/off, troubleshooting, and

intelligent noise reduction

Network functions

Network

protocols

IPv4/IPv6, 802.1x, HTTPS, HTTP, TCP/IP, UDP, RTP, RTCP,

UPNP, RTSP, SMTP, NTP, DHCP, DNS, PPPOE, DDNS, FTP

Access standard Supports the latest ONVIF and GB28181

Number of users Up to 8 users can log in at the same time

Mobile phones Supports iPhone, iPad, and Android platforms

Auxiliary interfaces

Network

interface 1 RJ45 10M/100M adaptive Ethernet port

General parameters

Housing Silver aluminum alloy shell

Power supply DC 12V, POE

Power

consumption < 5W

Page 27: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 27

Operating

temperature and

humidity

-30°C to 60°C, humidity less than 95% (no condensation)

Dimensions Φ107.1x85.5 mm

Protection level IP66

Weight 430g

Installation

method Wall or ceiling mounting

2.5.1.3. Dedicated face recognition dome cameras

These dedicated face recognition dome cameras have a compact and beautiful

appearance.

Model MegEye-C2V-320-P-X05

Target size 1/2.8in

Dimensions 108x70 (diameter/height) (mm)

Maximum

resolution 1920x1080

Maximum

frame rate 1920x1080@25fps

Shutter speed 1/10 sec to 1/1000 sec

Output video

bit rate 100 Kbps to 8 Mbps

Minimum

illumination [email protected]

Image settings Auto exposure, auto white balance, color correction,

Page 28: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 28

brightness, contrast, saturation, sharpness

Network

protocols

TCP/IP, ICMP, HTTP, FTP, DHCP, DNS, DDNS,

RTP, RTSP

Network

management Remote setting via client or browser

General

functions

Restore to factory default, anti-flicker, heartbeat,

mirroring, password protection, video masking,

watermarks, IP address filtering

Power supply DC12V±10%, supporting POE

Power

consumption 4W MAX

Operating

temperature -30°C to 60°C

Operating

humidity < 90%

2.5.1.4. Face recognition access control devices

With an outstanding appearance and friendly interface, these devices can be applied in

various indoor scenarios.

Model MegID-W2K

Main functions Face recognition, intelligent access control, liveness detection

Product dimensions

248.5x130.5x28.1 mm (length, width, and height)

Product weight About 1.1 kg

Page 29: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 29

Core architecture

CPU ARM-A17 quad core

GPU ARM Mali-T764

Memory RAM 2GB dual channel DDR3

ROM 16GB eEMMC high speed memory

Operating system

Android Android 5.1.1

Camera

RGB camera 2 megapixel HD camera (1080P)

Infrared camera 1.3 megapixel HD camera (720P)

White fill light Supported

Infrared fill light Supported

Screen Display screen 8-inch IPS HD screen (resolution 1280x800)

Touch screen 5-point capacitive multi-touch screen

Network WIFI Complies with IEEE802.11 a/b/g/n standard (2.4G/5G)

Ethernet 10/100/1000 Mbps Ethernet, RJ45 interface

Speaker Built-in speaker Power 2W, sound pressure 82dB±3dB

Buttons

On/Off button Supported, with hidden design

Reset button Supported, with hidden design

Anti-disassembly button

Supported

Sensors

Photosensitive module

Photosensitive detection method

Object sensing module

Laser detection method

Interfaces

USB DEVICE interface

Supported, mini USB, slave device

USB HOST interface

Supported, master device

Relay interface Supports NO, NC, COM

Wiegand output interface

Supports 26bit and 34bit

GPIO interface Support three-way extension

Power supply

Adapter DC18V-3A

POE POE+ power supply

Power Rated power: 12W Peak power: 20W

Temperature and humidity

Operating temperature

-15°C to 55°C

Operating humidity 20%-95% non-condensing

Storage temperature

-30°C to +65°C

Storage humidity 20%-93% non-condensing

Page 30: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 30

2.5.1.5. Face recognition outdoor devices

Designed for outdoor use only, these devices can be applied in outdoor and

semi-outdoor scenarios.

Model MegID-W2K-E

CPU Quad-core high performance processor

Screen 8-inch touch screen, full view, resolution 800x1200

Camera Binocular, 2 megapixel, 1080P, supports wide dynamic range

Storage capacity RAM: 2GB,ROM: 8GB

Infrared fill light Supported

White fill light Supported; white, green, and red are controllable

Ethernet 10/100/1000 Mbps

Microwave module Supported, response at 4 meters

Light sensing module

Supported

Wi-Fi Supported

Audio 2-channel audio output (line out)

USB interface 1 USB2.0 interface

Exit switch input 2.5 mm terminal 1P

Gate magnetic input 2.5 mm terminal 1P

Serial communication

interface RS232 serial port, 2.5 mm terminal 2P

Relay output 2.5 mm terminal 3P

Wiegand interface 2.5 mm terminal 3P, supporting Wiegand 26/34

Page 31: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 31

Wired network interface

RJ45 Gigabit Network

Protection level IP55

Power supply DC12V (±10%)

Operating temperature

-30°C to 60°C

Operating humidity 10% to 90%

Power consumption 18W MAX

2.5.1.6. Face recognition high-density capture cameras

Designed for high-density scenarios, such as factory attendance or school roll call, these

cameras can capture 30-50 on the same screen.

Camera

Model MegEye-C3S-123

Pixel 2 megapixel, 1/2.8in target size

Image resolution

Maximum resolution of main stream: 1920x1080, 1280x720, 704x576, 352x288

Maximum resolution of sub stream: D1(704x576)

Frame rate PAL: 1080p@25fps

NTSC: 1080p@30fps

Wide dynamic 120dB

Glare inhibition Supported

Digital noise reduction

3D digital noise reduction

Page 32: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 32

Electronic shutter

1/100000S0s to 1s

Minimum illumination

0.0001 Lux(F1.2,AGC ON), full color night vision

Color to Black Forced color mode

Video parameters

Adjustable image brightness, contrast, saturation, and chromaticity

Character overlay

Supported

Image processing

Supports image enhancement, backlight compensation, and image style options

Lens interface

Lens interface type

C/CS

Lens drive DC

Interface performance

Alarm interface 1-channel alarm input and 2-channel alarm output

Control interface 1-way RS485 interface, 1-way RS232 interface, supporting transparent channel and protocol mode

Memory interface 1 TFC, 1 USB

Reset interface 1

Network interface

1 10M/100M adaptive Ethernet port

Extension protocol

Onvif, GB/T28181

Coding mode

Video compression

standard H.265, H.264 HP/MP/BP, M-JPEG

Video compression rate

32Kbps to 16Mbps

Transmission mode

Dual code stream

Other functions

IE access Built-in web server

Clients ≤7

Local storage 1 USB (supporting WIFI/4G extension), 1 SD card up to 128 GB

Network protocols

TCP/UDP/HTTP/MULTICAST/UPnP/DHCP/PPPoE/DDNS/NFS/FTP/NTP/RTP/RTSP/IPv6/SNMP/SMTP/802.1X/QoS

IP address Supports static and dynamic IP addresses, MTU customization, and network card customization

System permissions

Level 3, 20 users with customized usernames and passwords

Page 33: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 33

System clock Built-in clock supporting external synchronization, NTP timing, time zone setting, and daylight saving time

Local management

Restore factory settings button

Remote management

Network remote upgrade, configuration import and export

Face Detection

The minimum and maximum pixels for face detection can be set, and the detection area can be set.

The maximum number of faces simultaneously detected in the picture is 50.

The detection frame rate can reach 25fps.

Face attribute detection supports gender, age, and race detection.

General functions

One-click recovery, password protection, black and white list filtering of IP addresses

General specifications

Protection level Electrostatic discharge Class 3B

Operating temperature

-30°C to 65°C

Operating humidity

0-95% (no condensation)

Power supply DC12V±10%

Power Max 9.0W

Dimensions 160mm (length) x 88mm (width) x 64mm (height)

Weight 810g

2.5.2. Host

Single host load capacity:

Host C2 M3002 M5 online

M5 offline (10,000 library)

C3S-123

Number of routes

Index Version Number of routes

Index Version Index Version Index Version Number of routes

Index Version

NUC 1

Access by 3

persons/second

V2.6.0 1 Gate

access V2.6.0

Total access of 6

persons/ second

V2.8.0

Total access of 30

persons/ second

V2.9.0 1

Access by 5

persons/second

V2.9.1

Industrial Personal Computer

3

Total access

of 5 persons/second

V2.6.0 3 Gate

access V2.6.0

Total access of 8

persons/ second

V2.8.0

Total access of 30

persons/ second

V2.9.0 3

Access by 8

persons/second

V2.9.1

TX1 2

Total access

of 3 persons/second

V2.8.1 2 Gate

access V2.8.1

Total access of 5

persons/ second

V2.8.1 Not supported Not supported

H4/H5 10

Total access of 20

persons/second

V2.6.0 10 Gate

access V2.6.0

Total access of 25

persons/ second

V2.8.0

Total access of 50

persons/ second

V2.9.0 10

Access by 30

persons/second

V2.9.1

Page 34: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 34

Description:

C2\M3002 has a clear limit on the number of routes. For example, if an industrial

personal computer can take only three routes for C2, and allow 5 persons to pass

per second, it means that the system can perform recognition for a maximum of 5

persons from three routes per second. If more than 5 persons pass together, one or

more persons may not be recognized.

M5: The number of routes varies according to the scenario. For example, NUC

allows 6 persons to pass per second in total, which means it can recognize 6

persons per second. In the gate scenario, assuming an average number of 30

people pass through each channel per minute at peak times, NUC can support 12

access channels at peak times. If two-way gates have one way open at the same

time, one NUC can support 24 routes for M5. In the hotel scenario, assuming that 4%

of the rooms are opened at the same time, NUC can support 150 routes for M5.

M5S: This device is mainly limited by record upload concurrency and library

synchronization concurrency. Similar to the M5 calculation method, NUC can

support 120 routes for M5S in the two-way gate scenario, and 750 routes in the hotel

scenario.

The above indexes are all for single products, such as when an IPC only has C2 or

M5. If the IPC is equipped with C2 and M5/M5S, it can be estimated proportionally,

but it cannot be measured due to the many possible combinations. We recommend

that only one type of front-end device be used for each host in actual use.

2.5.2.1. Standard version (IPC)

Page 35: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 35

The configuration of the standard IPC host is as follows.

Model MegServer-S1C-L3

Processing

capacity

Supports up to 3 channels of 2 megapixel video

streams

Dimensions 210x225x75 (mm)

CPU Intel i7 6700

Memory 8G DDR4

Storage 120GB SSD

Network

connection Dual Gigabit network card

Power adapter 120W

Operating

temperature 0°C to 50°C

2.5.2.2. Mini version (NUC)

The configuration of the mini host is as follows.

Model MegBox-B1D-122

Processing

capacity Supports 1 channel of 2 megapixel video streams

Dimensions 115×111×48 (mm)

CPU Dual core, 1.9 GHz main frequency

Memory DDR3 1600 8G(4Gx2)

Graphics card Integrated graphics

Hard disk 128GB SSD

Network Gigabit network card

Page 36: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 36

2.5.2.3. High performance server (Management + Recognition)

Model KSS-DR1000-H4

Processing

capacity Supports 10 channels of 2 megapixel video streams

Height 2U

CPU Intel XEON E5-2630 V4x2

Memory 64G

Graphics card Geforce GTX 1070/1080

Hard disk 1T 7.2K 3.5” SATA 6G x4 + Intel SSD

S3520/Series/(240GB/2.5in/SATA/6Gb/s/16nm/3D NAND)

RAID RAID:LR382A/8 ports/SAS 12Gb/half height/PCIe 3.0

x8/1GB cache/support RAID 0,1,5,6,10,50,60,JBOD

Network 1000Mx2

Power supply Redundant dual power 800W

2.5.2.4. High performance server (Recognition)

Model KSS-DR1000-H5

Processing capacity

Supports 10 channels of 2 megapixel video streams

Height 2U

CPU Intel XEON E5-2630 V4x2

Memory 64G

Graphics card Geforce GTX 1070/1080

Hard disk Intel SSD S3520/Series/(240GB/2.5in/SATA/6Gb/s/

16nm/3D NAND)

Network 1000Mx2

Power supply Redundant dual power 800W

Page 37: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 37

2.5.3. Optional accessories

2.5.3.1. TV box

A self-developed app program is run in the TV box to project the face recognition

welcome system on a TV or display screen. The specific configuration is as follows.

Model MegBox-B1F-122

Output

resolution 4K(3840x2160), 1080P(1920x1080), 720P

CPU RK3368

RAM 2GB

ROM 16GB eMMC

Video output 1xHDMI, 1xVGA

Other interfaces 4xUSB, 1xMicro SD

Wired network RJ45/WLAN

Wireless

network a/b/g/n +ac(2.4G+5G)

Remote control Infrared

Page 38: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 38

2.5.3.2. Network switch

A network switch is connected to the gate lock. When a subject passes face recognition

at the front end of the camera, the host issues a gate opening command and the network

switch opens the door.

Model TCP-KP-C2

Interfaces COM/NO/NC 3.81 mm terminal

Relay AC0-277V 10A; DC0-30V 10A

Operating

voltage DC5-28V

Maximum power

consumption 3W

Operating

temperature -40°C to 85°C

2.6. Authorization

All hardware products have been authorized before delivery.

After an access control device is restored to its factory settings, it must be connected to

the network and automatically authorized once.

Page 39: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 39

3. Successful Cases

3.1. Ant Financial Services

In 2015, the Megvii intelligent access control and welcome system was officially put into

operation at Ant Financial Services, serving its headquarters and thousands Ant

Financial Services and Alipay employees. Through the dynamic and non-cooperative

face recognition access control and intelligent welcome system, the experience of

enterprise employees and visitors was greatly improved. This helped Ant Financial

Services build an intelligent enterprise office environment from inside to outside and

solve a series of problems with traditional access control, such as complexity, poor

experience, and security issues.

3.2. Ucommune

For this well-known Chinese start-up incubator, the intelligent management of numerous

incubators has become a difficulty. In early 2015, Megvii provided an intelligent

enterprise solution and partnered with Ucommune to build an end-to-end face

recognition solution, including dynamic and non-cooperative face recognition access

control, intelligent welcome, visitor invitation and management, meeting room facial

check-in. This does away with traditional access control, all-purpose cards, and

complicated meeting room reservation and use procedures. It makes “face swiping”

Ant Financial/Face Detection

Intelligent Access Control

Page 40: Koala Face Recognition Access Platform Product White Paper · Face fuzziness < 10 ms 2.3.2.2. Liveness Detection The access control device supports two-eye activation, and the server

Page 40

a reality and provides further data support for big data operations and emerging

business models in incubator enterprise buildings.

3.3. More Customers

In no particular order:


Recommended