gemalto_redacted
Biometrics on the Move
Overview of contactless biometrics Biometric Advantage Disadvantage Pre- Use Case enrolment Voice • Ideal for • Uniqueness? • Always • Pre-travel mobile Fingerprint • Very unique • Need a free hand • Only if not • ABC with • Present in within touch distance in eID or eID or visa some eIDs • Incompatibility with eID is not contact FP? available Iris • Very unique • Sunglasses • Always • ABC + on- • Crowds • Light/angles the-move • Distance Face • Always in eID • Uniqueness? • Only if eID • Pre-travel • Crowds • Sunglasses not • ABC • Distance • Light/angles available • Gateless • Subject can be • Land unaware borders Others • ? • ? • ? • ?
Contactless video based m:n Face Recognition ⃰ How does it work? ⃰ How powerful is it? ⃰ Restrictions and challenges ⃰ Application for borders Gemalto Live Face Identification System 3
How does m:n work? STEP 1: An IP camera takes let’s say 25 frames per second STEP 2: System saves say 1 in 5 frames and makes a scene from them STEP 3: These several (m) images of 1 person are formed to make a model and are matched against models of many (n) people in the hot list Model Database Matching MATCH 4
Demo of system (colleagues in our London Office) 5
How powerful is it? More than 1,000 cameras Template creation in Capture to <500ms match 1 within Hot + White lists 1 second of more than 1,000,000 faces Accuracy >98% (<2% FRR at 0.1% FAR) 12 faces Fully centralised recognisable from system each camera view management All performance data here was measured in controlled conditions using ideal equipment 6
Capture to Match in under 1 second Live Videos Alerts Watchlist Video Face Face Template Model Score Record H.264 Matching Analysis History Tracking Image Modeling MPG4 MJEPG … 10-40 ms 10-100 ms 300-600 ms 10-300 ms 10 ms GPU 0.5M/S Total ~500-1000 ms Data here was measured in controlled conditions using ideal equipment 7
Accuracy Biometric accuracy is expressed as x% False Reject Rate (FRR) at y% False Acceptance Rate (FAR) Pass score In our internal tests Gemalto’s LFIS system (v4.5) produces 1.95% FRR at 0.1% FAR Means: FRR FAR ⃰ Using a pass score of 2,900 points where 1 person in 1,000 is falsely accepted across the border, ~2 people in 100 will need to be referred to a border guard for checking ⃰ Using a pass score of 3,000 points where 1 person in 10,000 is falsely accepted across the border, ~4 people in 100 will need to be referred to a border guard for checking In tests, human border guards miss ~5-10%, depending on ethnicity of traveller 8
Restrictions and Challenges Restriction Solution Face needs to be visible Need to remove helmets/sunglasses Lighting Darkfighter/Lightfighter* cameras to handle poor illumination Face at a long distance 48 pixels between the eyes is achievable with multiple high res cameras Angle of pitch and pose from camera to face Position cameras carefully Need an algorithm not restricted at 15 degree angles Fast moving faces Faster shutter speed Density of crowd Good algorithm Using existing cameras HD or better, ONVIF compliant cameras are generally needed • Darkfighter and Lightfighter are camera brand names from HIKVision. Other brands are available 9
Scenario 1: Border desk or e-kiosk/e-gate 1:1:1 match 10
