Direct Head Tracking - Robotized Transcranial Magnetic Stimulation - page 139

Fig. 8.6 Results of the laser calibration with n ¼ 50 datapoints, 25 datapoints were used for
calibration and the remaining 25 for testing. The left graph shows the translation errors of the
laser calibration for the Tsai-Lenz, Dual Quaternion, QR24, QR24 M and QR15 algorithms,
respectively, using 25 points for calibration and 25 points for testing. The right graph shows the
corresponding rotation errors. The corresponding numbers are given in Table 8.1
8.3.3.2 Head Registration for Head Tracking
We have scanned a human head phantom to analyze the performance of the head
registration. Therefore, we have registered the laser scans to a reference image
generated from an MRI-scan. For registration, we have applied the ICP algorithm
[ 3 , 5 ].
The computation time for ICP using a low resolution laser scan image and a
high resolution MRI reference image has been in the range of 30 s on a standard
desktop PC. The mean error found is 0.29 mm with an root mean square (RMS)
error of 0.36 mm. The maximum error is 0.98 mm. Roughly, 10-15 % of the data
points have been excluded due to noise and deflections, resulting in 6,000-8,000
data points for matching. Figure 8.7 illustrates a laser scan overlaying the MRI
head contour.
8.3.3.3 Head Tracking Based on 3D Laser Scans
To estimate the accuracy of head tracking based on 3D laser scans, we have
mounted a head phantom to a robot. At 50 positions, we have recorded a 3D laser
scan which we have registered to a 3D laser scan at the initial starting position.
Subsequently, we have used the robot motion as ground truth in order to evaluate
the translational accuracy of the head tracking.
On average, the absolute translational error of the head tracking is 4.84 mm
with an SD of 2.99 mm. Figure 8.8 visualizes the estimated absolute translational
errors as a boxplot.