The algorithm presented here allows the motion tracking of a human subject performing rehabilitation exercises during physiotherapy. The motion is tracked by fitting a model into the observed data, which are depth images coming from one or multiple Kinect sensors. The model consists of a surface mesh morphologically close to the patient's body and an articulated skeleton. The mesh is deformed by linear blend skinning according to the pose of the skeleton. The optimization is performed by particle filtering. Thanks to the graphics pipeline and the computing capabilities of the GPU, our algorithm reaches execution speeds close to real time. When working offline with a model very close to the patient's morphology, the joints locations and rotations are estimated with an average accuracy respectively smaller than a few millimeters and a few degrees.
[1]
Andrew W. Fitzgibbon,et al.
Real-time human pose recognition in parts from single depth images
,
2011,
CVPR 2011.
[2]
Ashley Reid,et al.
A comparison of linear skinning techniques for character animation
,
2007,
AFRIGRAPH '07.
[3]
Andrew W. Fitzgibbon,et al.
Efficient regression of general-activity human poses from depth images
,
2011,
2011 International Conference on Computer Vision.
[4]
Antonis A. Argyros,et al.
Efficient model-based 3D tracking of hand articulations using Kinect
,
2011,
BMVC.
[5]
Simon J. Godsill,et al.
On sequential Monte Carlo sampling methods for Bayesian filtering
,
2000,
Stat. Comput..