Closing Remarks - Robotized Transcranial Magnetic Stimulation - page 162

10.2.3 Direct Head Tracking
We have already discussed the advantages of direct head tracking for robotized
TMS ( Chap. 8 ) . However, the current three-dimensional (3D) range scanning
systems cannot be used for neuro-navigated or robotized TMS due to their inac-
curacy. Nevertheless, advanced high precision direct head tracking is a very
interesting and promising research field.
10.2.3.1 3D Depth Sensor
Recently, low budget 3D depth senors integrated, e.g., in Microsoft's Kinect
(Microsoft Corporation, Redmond, Washington, USA) and Asus' Xtion Pro
(ASUSTeK Computer Inc., Taipei, Taiwan) were introduced. With automatic
gesture detection and movement control, their major application is in the enter-
tainment sector and consumer electronics [ 10 ]. In analogy to 3D laser scanning
systems, these depth sensors also provide 3D scatter plots of the scanned surfaces.
The 3D depth sensor uses an infrared laser to measure the surface and a mono-
chrome image sensor detects the reflected incoming laser light. By estimation of
the angle and the time difference of the incoming laser light beam, a 3D position
can be computed.
Even though their resolution is limited in contrast to 3D laser scanning systems,
they provide real-time motion tracking [ 16 ]. Therefore, these systems might be an
alternative to costly 3D laser scanning systems. With further developments of high
resolution image sensors the resolution will increase in the near future. Therefore,
it might be reasonable to consider such a 3D depth sensor for direct head tracking.
As these systems are low priced, a redundant setup with multiple senors arranged
around the patient's head should be taken into account to increase the tracking
accuracy and stability. However, synchronization strategies must be considered as
multiple sensors might disturb one another due to the laser light reflections.
Figure 10.1 a shows a 3D scatter plot obtained with Microsoft's Kinect. The 3D
scan consists of roughly 9,500 surface points. However, in contrast to 3D laser
scanning systems (cf. Sect. 8.3 ) , the scan is relatively noisy. This can be clearly
seen, when reconstructing a head surface from the scan using the PowerCrust
algorithm [ 2 ], as illustrated in Fig. 10.1 b. However, by averaging throughout a set
of 3D scans, the noise might be reduced essentially. For head tracking, a point
registration algorithm, such as Iterative Closest Point (ICP), must be used
(cf. Sect. 8.3 ). These algorithms themselves compensate for the noise in the
implementation.