Pose Liveness Detection

Liveness detection is the process whereby facial recognition software attempts to differentiate between genuine live faces and spoofed fake faces. (such as from a photo of a face) SAFR uses a face’s center pose quality to attempt to detect liveness.

Pose liveness detection operates as follows:

  1. State A: An unrecognized face needs to be recognized.
  2. State B: Proof of liveness will pursued as follows:
    1. The recognized face is tracked at the rate of at least 25 frames per second.
    2. Any loss of tracking (occurrence of lingering for more than 1 frame) or a detection gap > 40 ms in frame capture time results in the need to re-recognize the face and thus a return to State A.
    3. Pose quality must maintain a score of 0.5 or higher for 3 consecutive frames and at least one of the samples must have a profile pose confidence of 35% or less to trigger the transition to the next state, State C.
  3. State C: A smooth transition to profile pose will be pursued as follows:
    1. The face is tracked at the rate of at least 25 detections per second.
    2. Any loss of tracking (occurrence of lingering for more than 10 frames) or a detection gap > 40ms in frame capture time results in the need for re-recognition and thus return to State A.
    3. A momentary loss of tracking (recovered in less than 10 frames) will require a center pose quality difference from the prior frame of no less than 0.15.
    4. If change in identity is detected as part of prescribed re-recognition, State B will restarted.
    5. Pose quality must be observed to transition to score of 0.26 or lower for at least 3 consecutive frames and with 66% of at least 3 images but no more than 30 images immediately proceeding with scores observed > 0.26 and < 0.5 and in decreasing sequence to trigger transition to State D.
      • For example: 0.45, 0.37, 0.23, 0.12, 0.24
      • This algorithm can interpreted as requiring presence of descending strand of samples being at least 66% of the number of samples with min number being specified in preferences and max number being 30 (~1 second).
  4. State D: After the profile pose state has changed, a verification call is issued to obtain a similarity score to the identity obtained in State A.
    1. The verification call must indicate at least a 86% match.
    2. A response from recognition must also indicate that the face is in profile pose, based on profile pose confidence returned and threshold set.
    3. If both of above are met, liveness detection will conclude.
    4. If both aren’t true, re-recognition will continue immediately for as long as the pose quality score remains at 0.26 or lower until successful confirmation of pose and 86% identity match is confirmed.
    5. If pose score exceeds value of 0.26 for 3 consecutive frames, transition back to state B will occur.
    6. Any loss of tracking (occurrence of lingering) will result in need for re-recognition and thus return to State A.

See Also