Capturing the Motion of Ski Jumpers using Multiple Stationary Cameras

1
Capturing the Motion of Ski Jumpers using
Multiple Stationary Cameras

• Atle Nes
• atle.nes_at_hist.no

Faculty of Informatics and e-Learning Trondheim
University College
Department of Computer and Information
ScienceNorwegian University of Science and
Technology
2
Project description

• Task Build a cheap and portable video camera
  system that can be used to capture and study the
  3D motion of ski jumps during take-off.
• Goal Use it to give reliable feedback to the ski
  jumpers and their trainers that can help improve
  the jumping skills.

3
Solution / How ?

• Multiple video cameras are placed strategically
  around in a ski jumping hill capturing image
  sequences from different views synchronously (10
  m before 30 m after edge).
• Using calibrated cameras it is then possible to
  reconstruct 3D coordinates if the same physical
  point is detected in at least two views.

4
(No Transcript)
5
Camera equipment

• 3 x AVT Marlin CCD based cameras
• Firewire (no frame grabber card needed)
• 640x480 x 30 fps
• 8-bit grayscale (color cameras not chosen because
  of interpolating bayer patterns)
• Exchangeable lenses (fixed and zoom)

6
Camera equipment (cont.)

• Video data (3 x 9MB/s 27 MB/s)
• 2 GB RAM (sequences buffered to memory)
• 2 x WD Raptor 10.000 rpm in RAID-0 (enables
  continuous capture)
• Extended range
• 3 x 400 m optical fibre (full duplex firewire)
• Power from outlets around the hill
• 400 m BNC synchronization cable

7
Camera setup
Synch pulse
Video data Control signals
8
Video processing

• Points must be automatically detected, identified
  and tracked over time and accross different
  views.
• Reflective markers are placed on the ski jumpers
  suit, helmet and skies.

9
Video processing (cont.)

• Blur caused by fast moving jumpers (80 km/h) is
  avoided by tuning aperture and integration time.
• Three cameras gives a redundancy in case of
  occluded/undetected points (epipolar lines).
• Also possible to use information about the
  structure of human body to identify relative
  marker positions.

10
Camera calibration

• Direct Linear Transformation used to give a quick
  estimate of the 2D?3D mapping.
• Unconstrained Bundle Adjustment is used to refine
  the 3D geometry iteratively.
• Intrinsic parameters precomputed (focal length,
  principal point, lens distortion)
• Extrinsic parameters computed on-site (camera
  position direction)

11
Direct Linear Transformation
12
Direct Linear Transformation

• 6 visible calibration points minimum for camera
  calibration. More points will in general increase
  calibration accuracy.
• 2 different views minimum for 3D point
  reconstruction. More views will in general
  increase triangulation accuracy.
• Direct solution using Least Squares Method
  (linear equations)

13
Lens distortion / Optical errors

• Imperfect lenses result in nonlinear terms
  (straight lines are no longer straight)

14
Bundle Adjustment

• Adds lens distortion, skew and affinity
• Iterative solution using Levenberg Marquardt
  Method (unlinear equations)
• Calibration points with the largest errors are
  removed automatically resulting in a more stable
  geometry.

15
Calibration frame

• Was used for finding estimates of theintrinsic
  parameters.
• Exact coordinates in the hill was measured using
  differential GPS and a land survey robot station.
• Points made visible in the camera views using
  white marker spheres.

16
Visualization

• Moving feature points are connected back onto a
  dynamic 3D model of a ski jumper.
• Model is allowed to be moved and controlled in a
  large static model of the ski jump arena.

17
Granåsen ski jump arena
18
Conclusion

• I have presented a 3D video system that can be
  used in a large scale environment like a ski
  jumping hill.
• It remains to be seen how well the ski jumpers
  will perform based on this kind of feedback.

19
Questions?