3D Scanning
We develop a contour-coherence based registration algorithm to align wide baseline range scans. The traditional
registration algorithms, e.g., ICP, fail in this case, as many closest point-to-point corresponding pairs are incorrect in presense of limited overlap.
The contour-coherence, on the other hand, still serves as a strong clue, as no matter the amount of overlap only the 2D contour points are used for
registration. An example is given when registering two range scans of the stanford bunny with an overlap of approximately 40% (top left figure). The ICP style algorithms fail
as most correspondences are incorrect. On the other hand, we successfully register these two range scans using contour-coherence (top right figure).
We apply contour-coherence to address the problem of multi-view rigid registration, and further extend it to solve the problem of multi-view
piece-wise rigid registration. This allows us to reconstruct rigid as well as articulated objects from as few as 4 range scans, i.e., front, back, and two profiles.
On top of this, we build a 3D Body Scanner, using a low-cost Kinect device, to accurately capture the 3D shape of human bodies. Our scanning system robustly provides globally accurate results
with preserved small details. The complete system is easy to set-up (i.e., put the Kinect on table or tripod) and the instructions are easy to follow (i.e., turn when asked).
In less than 2 minutes, including the data acquisition, an accuate and complete model is generated.
Point of Care Monitoring (POCM)
For patients with Parkinson's Disease, the current patient-clinician evaluation mode is very ineffective (top left figure). As such, we propose a reliable and accurate home monitoring and evaluation system as an
attractive alternative (top right figure). While traditional home monitoring systems heavily rely on the use of invasive sensors, e.g., accelerometer and gyroscrope, we offer a solution using a single non-invasive
3D sensor. The subject is asked to perform standardized medical tests in front of the 3D sensor, while the depth sensor monitors his/her motion by producing a skeletal stream. We propose a Temporal
Alignment and Spatial Summarization (TASS) algorithm to process this noisy high-dimensional periodic time-series data, and produce a denoised skeletal sequence which captures the subject's most consistent
motion pattern. Important biomedical indicators are further extracted from this robust representation.
