|
Head Nodding
The Cog Shop
MIT Artificial Intelligence Laboratory
545 Technology Square, #920
Cambridge, MA 02139
|
|
By adding a tracking mechanism to the output
of the face detector and then classifying these outputs, we have been able to have the
system mimic yes/no head nods of the caregiver (that is, when the caretaker nods yes, the
robot responds by nodding yes). The face detection module produces a stream of face
locations at 20Hz. An attentional marker is attached to the most salient face stimulus,
and the location of that marker is tracked from frame to frame. If the position of the
marker changes drastically, or if no face is determined to be salient, then the tracking
routine resets and waits for a new face to be acquired. Otherwise, the motion of the
attentional marker for a fixed-duration window is classified into one of three static
classes: the yes class, the no class, or the no-motion class. Two metrics are used to
classify the motion, the cumulative sum of the displacements between frames (the relative
displacement over the time window) and the cumulative sum of the absolute values of the
displacements (the total distance traveled by the marker). If the horizontal total trip
distance exceeds a threshold (indicating some motion), and if the horizontal cumulative
displacement is below a threshold (indicating that the motion was back and forth around a
mean), and if the horizontal total distance exceeds the vertical total distance, then we
classify the motion as part of the no class. Otherwise, if the vertical cumulative total
trip distance exceeds a threshold (indicating some motion), and if the vertical cumulative
displacement is below a threshold (indicating that the motion was up and down around a
mean), then we classify the motion as part of the yes class. All other motion types
default to the no-motion class. These simple classes then drive fixed-action patterns for
moving the head and eyes in a yes or no nodding motion. While this is a very simple form
of imitation, it is highly selective. Merely producing horizontal or vertical movement is
not sufficient for the head to mimic the action—the movement must come from a
face-like object.
|