Biomedical Engineering Reference
results motivate a closer look at the error sources. We have found that scaling
strongly influences the FaceAPI tracking. This observation leads to the assumption
that the FaceAPI measures enlarged translational values.
For estimating the position of the head in the webcam images a basic model for
the head geometry with well-defined distances is mandatory. For this reason, a
fixed texture for a human head is defined. This texture is the basis for any cal-
culation of the head position in the FaceAPI. Due to a fixed texture as basis for the
calculation, it is not possible to track any human head exactly with the FaceAPI.
This leads to the assumption that a human head that is more appropriate to the
fixed texture will have better tracking values than a head that differs from the
This assumption is supported by a simple experiment where we have compared
the deviations of the translational values of the data obtained by the FaceAPI and
by the Polaris Spectra for different subjects. We have found that the accuracy of
the FaceAPI strongly depends on the size of the head and the location of the facial
landmarks [ 5 ].
For neuro-navigation and robotized TMS, we have a three-dimensional (3D)
contour of the patient for navigation and treatment planning. It should therefore be
possible, to extract the landmark locations for each single patient and then to
generate an individual template as the basis texture for the FaceAPI. The presented
results lead to the assumption that a feasible texture would lead to more accurate
tracking results. Even though the FaceAPI is not suitable for robotized TMS in its
current version, an adapted version with individual textures might be a promising
9.5.2 3D Laser Scans
Our practical experiments have shown that a direct head navigation based on a 3D
laser scanner is feasible for the robotized TMS system.
As the laser scanner is focused towards the TMS-workspace (the space where
the patient's head is located during the TMS session), the limited measurement
volume of the laser scanner (compared to the measurement volume of a Polaris
tracking system) is not a restriction for use in the robotized TMS-system. During a
standard TMS session the patient's head must be in the measurement volume or
the head cannot be reached with the robot anyway.
As we have not tried to optimize the software yet, a big limiting factor is the
computation time for the Iterative Closest Point (ICP) algorithm for online or real-
time applications. With the development of faster and more parallel processors
even for standard computers, the computation time will speed up significantly.
Recent developments have shown a real-time capable ICP implementation by
using fast graphics hardware [ 11 ]. However, the accuracy of the registration with
roughly 0.3 mm is sufficient for head tracking.