Biomedical Engineering Reference
In-Depth Information
(a)
(b)
Fig. 10.1 a 3D scatter plot with the Kinect's 3D depth sensor of our head phantom. The scan
consists of roughly 9,500 surface points. b Reconstructed head surface from the scatter plot using
the PowerCrust algorithm. The noise in the scan is clearly visible in the reconstructed surface
10.2.3.2 Customized Head Tracking with Webcams
Currently, standard webcams provide images in High Definition (HD) with a
resolution of 1920 1080 pixels. With these high resolution images also the
accuracy of the FaceApi will increase. However, the fundamental problem
remains: A standard head is used as ground truth for the tracking. The found facial
landmarks are related to the standard values and based on that the 3D pose of the
head is calculated. As human heads vary in size and shape, the difference to the
standard head can be essential. Therefore, the tracking accuracy strongly depends
on the similarity of the tracked head to the standard head.
Following the scheme by Vukadinovic and Pantic [ 20 ], facial feature points can
be estimated in a fully automatic fashion. Furthermore, face detection can be
performed in real-time [ 19 ]. Therefore, these tracking algorithms can be adapted to
relate the feature points to the ones of the individual head, e.g. obtained from
Magnetic Resonance Imaging (MRI)-scans. In this way, the 3D pose of the head
can be calculated more accurately in the webcam images.
As Microsoft's Kinect or Asus' Xtion Pro have a standard webcam integrated in
addition to the 3D depth sensor, we might consider to use the complete tracking
information to result in accurate head tracking.
Search WWH ::




Custom Search