In this experiment, a person's head was tracked using both the vision
algorithm and the Polhemus magnetic sensor 6 simultaneously.
Figure 9 shows the vision estimate and Polhemus
measurements (after an absolute orientation [29] was
performed to align the estimates properly). The RMS difference in
translation is 0.11 units and the RMS difference in rotation is
.
(The scale of translation is, of course, unknown, but
is approximately 10-12cm per unit, yielding a RMS tracking error of
approximately 1 cm.) This yields accuracy on the order of the
observed accuracy of the Polhemus sensor, indicating that the vision
estimate is at least as accurate as the Polhemus sensor.
This example is identical to the example presented in our earlier work
on vision-based head tracking [5], except here
we recover focal length and structure simultaneously with motion. The
previous work relied on a rough, a priori structural model and
calibration of focal length. The RMS errors between vision and
Polhemus estimates for this example were slightly better than those in
the previous study, (1cm versus 1.67cm and 2.35 degrees versus
2.4 degrees).