motion capture laboratory
TRANSCRIPT
Introduction• Sampling the pose and location of a subject over time.
• Optical motion capture
Markers
Multiple cameras
3D position reconstruction ofmarkers
Rigid body orskeleton information
Marker extraction – Cameras
Markers Cameras Pixels Image processingMarker
positions
• Web cameras
640x480 @ 30 FPS
• PlayStation 3 Eye
640x480 @ 60FPS
• uEye cameras
Industrial camera
1024x768 @ 30FPS
• Basler cameras
Industrial camera
658x492 @ 120 FPS
• On the GPU
CUDA
1-2 ms processing
• Realtime processingat 120 FPS
Marker extraction –OptiTrack
Markers 18 x OptiTrack Flex 13Marker
positions
• 1280x1024 @ 120 FPS
• IR Leds, passive markers
• On-camera image processing
Reconstruction• Point correspondence
Matching points on different camera views
Epi-polar geometry
• Marker position reconstruction
• Model fitting
Rigid body
Skeleton
Calibration• Camera parameters
Intrinsic parameters (focal length, principal point, distortion)
Extrinsic parameters (position, orientation)
• Two steps
Starting estimation for each camera
Built-in methods in OpenCV for both intrinsic and extrinsic
Bundle adjustment for whole camera system
Wanding
Levenberg-Marquard algorithm
• Average error: 0.233cm
Future work• Automatic marker labeling for skeleton
• Robust handling of missing and reappearing markers
• Skeleton calibrationshoulder
elbow
wrist
Conclusions• Different types of cameras
GPU based real-time image processing at 120 FPS
Scalable
OptiTrack cameras
• Modular design, easy extension
• Good calibration <0.5 cm accuracy
• Improving skeleton tracking
EEA Financial
Mechanism 2009-2014 -
HU08 Scholarship
Program
BUTE, Department of
Mechatronics, Optics and
Mechanical Engineering
Informatics
Narvik University College
Energy Agency Public
Nonprofit Ltd.
COLLECTIVE DEVELOPMENT OF MODERN EDUCATIONAL METHODOLOGIES IN THE
FIELD OF ONLINE MEASUREMENT, -CONTROL AND REMOTE MONITORING SYSTEMS