Christoph Leuze1, Grant Yang1, Gordon Wetzstein1, Mahendra Bhati1, Amit Etkin1, and Jennifer McNab1
1Stanford, Stanford, CA, United States
Synopsis
Many medical applications such as brain surgery or stimulation
require the clinician to identify an internal target location. Mixed reality
see-through displays that enable a holographic visualization of brain MRI superimposed
on a subject’s head can help clinicians identify internal target locations but
require tracking methods that keep the holographic brain MRI aligned with the
subject’s head as they move. We present
a method for marker-less tracking of a subject’s using a depth-sensing camera,
which tracks facial features and sends location and rotation information to a
see-through display to update the location in space of the MRI holograms.
Introduction
Many medical applications such as brain surgery or
stimulation require the clinician to identify an internal target location. This
can be achieved by locating the target location in an individual’s brain MRI and
relating the position of that brain region to externally visible landmarks. The
advent of mixed reality see-through displays such as the Microsoft Hololens has
made it possible to directly project volumetric MRI data as holograms overlaid
on the subject in the real world, facilitating targeting by providing the
clinician with the possibility to directly “look into” the patient [1]. To
align holograms with real world objects, traditionally requires estimation of the
pose of an object using well-defined markers attached to the object. Here we
present a method that measures the head pose relative to the world coordinates
of facial features. Since the Hololens room tracking depth stream does not have
sufficient resolution for accurate face tracking, we use an external RGBD
camera attached to the Hololens and transformed the coordinate system between
RGBD camera and the Hololens.Methods
An Intel RealSense SR300 RGBD (red-green-blue-depth)
camera was attached to the Microsoft Hololens and connected via USB to an
external laptop (Fig.1). The room tracking infrared lasers of the Hololens were
covered to minimize interference with the RealSense depth camera (Fig.2). We
developed custom software that uses the Intel RealSense SDK to track the world
coordinates of a subject’s facial features and uses these coordinates to update
the pose of a holographic head rendering in the Hololens display according to
the following steps:
- Calibration 1: Perform camera
calibration to find transformation from RealSense camera space to Hololens
display space (Fig.1).
The transformation was measured according
to [2], with THvRr=THvHrTHrRr where THvHr
(provided by Hololens API) is the transformation from Hololens
RGB camera to Hololens virtual camera (the user’s POV in the virtual world). THrRr
was measured by performing a camera calibration for the RealSense RGB and the
Hololens RGB camera using a checkerboard that was attached in 12 room locations
completely visible by both cameras. THrRr was then acquired by THrRr=(TCHr)-1TCRr where
TCHr and TCRr
are the Hololens and RealSense camera calibration matrices.
- Calibration 2:
Manually
align the location of the rotation axes in the holographic head to match the rotation
axes in the real head.
- Tracking: Measure head pose
of the real head with the RealSense camera in the
RealSense RGB space. This was done by tracking facial landmarks
and measuring the pose of the facial landmark point cloud. The transformation TRrRd from
RealSense depth to RGB camera, necessary to measure the landmark world
coordinates, was provided by the RealSense API.
- Rendering: Transform the
tracked pose to the Hololens display space using the transform measured in
calibration step 1. These values are then used to rotate and translate the holographic head
around the rotation axes center adjusted in calibration step 2.
Results
The transformation measured in calibration step 1 consisted of a
3cm translation from the RealSense RGB camera to the Hololens display space.
The center of the rotation axes of the holographic head was translated in calibration step
2 from the center of mass of the holographic head model 9cm to the back of the
head where the head meets the spine. Figure 3a&b show a surface rendering
of the head overlaid on the subject viewed through the Hololens and a volumetric
rendering that allows
brain
visualization.
Latency of the head hologram pose update was less
than 50ms and only noticeable for quick head movements. Our measurements in
Table 1 quantify how the head tracking accuracy for distances depends on 1) camera distance and 2) head
rotation with respect to the camera.Discussion & Conclusion:
We
have presented a method to update an MRI rendering of a subject’s head in
real-time by performing marker-less tracking using an RGBD camera. Instead of
pose estimation from the facial landmark point cloud, a probabilistic model
that creates pose hypotheses and evaluates them on the observed RGBD data could
be used if the shape of the head is known [3]. Additional measurements of head
to hologram accuracy, which depend on perception of the individual user, still
need to be performed. To improve this display accuracy an individual tracker
based calibration procedure such as presented in [4] is necessary to minimize
the perception error for each user.
A
stick PC connected to the RealSense camera could enable a completely wireless device,
increasing the freedom of movement of the user.
Our method to project and track brain MRI
holograms on a subject based on facial feature tracking will allow clinicians
in the future a high flexibility and simple localization of
internal targets by eye.
Acknowledgements
No acknowledgement found.References
[1]
Leuze C. et al., Holographic visualization of MRI data aligned to the human
subject, ISMRM 2017 [2] Garon et al. Real-time High Resolution 3D Data on the
HoloLens, ISMAR 2016 [3] Ren C. et al. STAR3D: Simultaneous Tracking And
Reconstruction of 3D Objects Using RGB-D Data, IEEE International Conference on
Computer Vision 2013 [4] Long Q. et al. Comprehensive Tracker Based Display
Calibration for Holographic Optical See-Through Head-Mounted Display, arxiv
2017