Facial recognition - overcoming the barriers
September 2001, Access Control
Facial recognition was first introduced onto the South African market about four years and has been vigorously marketed by a number of companies over this period.
Initially the barriers to acceptance included:
* TCP/IP or serial port communication protocol software integration that needed to be implemented if this type of biometrics technology was to be used in conjunction with card based access control systems.
* The unknown or unproven performance factors of false acceptance and rejection rates.
* The establishment of environmental lighting conditions for enrolment and verification positions.
* The reliability of any facial recognition systems where the majority of persons being verified has a dark facial complexion.
* The price per control point, which was made up of the software and the hardware required.
Over the past year Modular Communications has had a system running in a live application of access control with T&A. This specific application was chosen for ongoing evaluation because market indications show that this application will be the most common use of this type of technology followed by high security access control applications.
Modular Communications' Clifford Rose says Modular has methodically worked through the barriers facing biometric facial recognition and has four years of experience in this field.
Protocol system integration
Raw facial recognition systems are available for integrators to develop into a Windows-based application and then integrate this to other systems. On the other hand, a system such as FaceVACS available from Modular is a developed solution, which has an open architecture design.
Here integration to other access control systems can be simply accommodated via a TCP/IP protocol where 'hooks' and 'sockets' are used to transfer messages between different software packages and their database structures.
This concept of integration is different to providing an access control system with a standard 'Wiegand' interface (like from a card reader) and has been the root cause of confusion in the past with access control system vendors as well as consultants. The TCP/IP communication between these biometrics systems and access control systems as well as other peripherals is established in a LAN/WAN type communication network. New development readers and cameras configurations will also sit on this type of network and be IP addressable and intelligent.
Performance factors (false acceptance and rejection rate)
In understanding the performance factors one needs to understand the conditions that apply to this measurement. In this instance, the performance figures are based upon a facial recognition system being integrated into a card access system.
An important fact to remember is that for any biometric system there is a time delay to scan a large database for comparison to one feature set (the computer summation after analysing a body part and reducing this to a 500-2000 byte size string of computer information). In addition, there is an exponential increase in error rate as the database size increases. These result in systems giving operators a variety of options from best to worst of corresponding people who may fit the biometrics match.
So why the card system?
When the card is presented it draws the corresponding feature set applicable to the authorised cardholder from the database for comparison. This method automatically reduces the false acceptance probability. Furthermore, there is no feature set comparison against the database and the effective verification time of the person can be less than 1 second.
Under these conditions sample tests of 1064 and 1125 transactions have established the following results:
* No card fraud, which means that out of the 50 odd people clocking through the live site in PE, no fraudulent cards were presented. This was confirmed by checking all transactions, which were accepted or rejected.
* All rejections were because of staff not facing the camera during the verification process. This is a disciplinary problem and would probably disappear if the person had to go through a booth or turnstile to gain entrance. In our operational environment the attendance clocking is in a passage against the wall with unrestricted traffic flow.
* These rejections cannot be called false rejection (rejecting the correct cardholder) because the rejection is recorded as no eyes or face found, which is a human problem and not a machine problem.
* There was no false acceptance.
Environment lighting conditions
The success or failure of any facial recognition system is based upon the image quality which is a function of the quality of the camera, the environmental lighting and reflective conditions. There have been various approaches to the controlling of environment conditions for a facial verification point. The concept of flooding an area with as much light in order to stabilise light conditions is wrong. Understanding CCTV cameras, lenses and a bit of physics is what is needed to create a good repeatable photographic environment in which modern digital CCTV CCD cameras can operate.
The basic considerations are:
* A reasonable light level diffused onto the image, in this case the person's face.
* Blocking off interfering external light sources from behind the image.
* Creating a low reflectivity background behind the person or face area.
* Programmable and auto iris control of video signal level depending upon the colour of the subject in front of the camera.
* The use of a high resolution monochrome CCD digital camera.
Naturally it is important to enroll people under similar environment, camera and lens conditions as the verification environment. Having different conditions will increase the false rejection rate, which will cause basic operational problems.
Reliabilty of the recognition process
A great deal of enquiries in respect of facial recognition are whether or not the system works with dark faces. This is a very good question as most CCTV applications today include colour cameras and colour monitors. Trying to identify dark complexion people using a colour camera is sometimes quite difficult. Normally colour CCTV systems help to identify the colour of clothes, cars and goods and on this basis a probable identification can be made.
According to Cliff Rose where dark complexion faces need to be identified, the use of a monochrome high resolution camera better captures contours and features when compared to colour cameras. Colour temperature changes are also easier accommodated using monochrome cameras compared with colour cameras. Whether looking at a dark complexion or a light complexion face the recognition area on the face does not have much of a colour component.
For many years the price of facial recognition software was high. Added to this, it was recommended that the system should only be sold with the top of the range digital cameras. This meant that the hardware and software were very expensive. Added to this one still needed additional computers and have an integration cost in excess of R10 000 to bear.
Besides price, access control vendors were unsure of the TCP/IP integration, the environmental conditions, reliability figures and whether this in fact is the right solution. However, over the past few years systems such as FaceVACS have been further refined enabling the use of cheaper monochrome cameras and the cost of the software itself has reduced.
For further details contact Cliff Rose, Modular Communications on tel: (041) 364 2653, e-mail: firstname.lastname@example.org