PDA

View Full Version : SUMMARY: Best practices for 3D reconstruction in SportsBiomechanics



nflyger83
09-07-2006, 05:07 PM
Below is a summary of replies to my post regarding 'Best practices for 3D
reconstruction in Sports Biomechanics'.

Regards
Nick Flyger
Biomechanics Unit,
Institut Sukan Negara Malaysia,
Komplex Sukan Negara,
Bukit Jalil, Sri Petaling
PO Box 10440, 50174 Kuala Lumpur,
MALAYSIA

ORIGINAL QUESTIONS:

What would be the 'Golden Camera Calibration Algorithm' for a sports
biomechanist who was happy with reconstruction error of around From the Biomech-L archives Tomislav Pribanic suggested the following
however given time constraints and my knowledge of programming and maths I
wondered if this might be overkill for the desired applications.

> 1. Carry out DLT calibration
> 2. extract camera parameters with so-called qr decomposition of camera
> matrix (there is a paper explaining qr decomposition for the purpose of
> camera calibration, but I cannot think of it now; if you fail finding it
> also I'll try look for it)
> 3. Include non-linear distortion parameters and along with camera
> parameters from previous step start non-linear minimization procedure (I
recommend
> Levenberg-Marquardt algorithm, but there are many others as well) and
> refine your camera parameters through certain number of iterations until
> convergent set of camera parameters solution is reached.
I also asked, given it is recommended when setting up the cameras for DLT
that they are roughly perpendicular or at least not seperated by 120
degrees, what happens if 4 cameras are used such that 2 pairs of camera axes
are parallel?

Andy Dianis relied...
Regarding your parallel optical axis question: A good 3D reconstruciton
algorithm can utilize data from cameras with essentially paralled optical
axes without introducing errors, but you need at least one other camera with
a non-parallel axis. However, cameras with non-parallel axes will always
provide more reliable results.

*************************

SUMMARY OF POSTS

Hi Nick,

One of the reasons for the more complicated techniques is that they
allow you to use cheap cameras and lenses. Machine vision style
cameras are very cheap and easy to use especially when you want to
synchronise several at once but they generally have really dreadful
distortion around the edges of the field of view. This can be
corrected for if you have enough reference points and the best way of
doing that is to use a moving reference object that you can move
around the whole field of view of the camera. This is what happens
when you use the calibration wand with Vicon and Qualisys systems -
over 10 seconds or so they can collect thousands of calibration
points which is enough to calibrate both the spacial and optical
parameters of the cameras over the whole volume of interest. They can
then use much cheaper optics and nowadays computing power is much
cheaper than decent lenses. If you plan to use camcorders or spend
serious money on good quality video lenses then the benefit of all
this extra information is much smaller. However dynamic calibration
is still worthwhile because it is much easier to build a accurate
small reference object that you move around than it is to build a big
one that covers the whole volume of interest.

Cheers
Bill Sellers
Faculty of Life Sciences
The University of Manchester

******************************

Dear Nick,

To my experience, the most practical way is
http://www.vision.caltech.edu/bouguetj/calib_doc/

All the best,
Victor A. Sholukha
Universite Libre De Bruxelles (ULB)

******************************
Hi Nick,

The best method for camera system calibration depends more upon the ray
identification method to be used and the complexity of the marker
distribution than the type of motion that you are measuring.

If you identify the marker images manually in each camera record then most
calibration techniques will provide adequate accuracy for biomechanical
sports applications. However, if you wish to use automatic ray
identification techniques to do the 3D reconstruction, especially in
applications where there are many markers in close physical proximity to
each other, then highly accurate (and more sophisticated) camera
calibrations are essential.

Of course there are other confounding factors, a principal one being the
ability of video system cameras to resolve finite sized markers in the
images.

Regarding your parallel optical axis question: A good 3D reconstruciton
algorithm can utilize data from cameras with essentially paralled optical
axes without introducing errors, but you need at least one other camera with
a non-parallel axis. However, cameras with non-parallel axes will always
provide more reliable results.

Hope this helps,
Andrew Dainis

******************************

Dear Nick,

Reconstruction has come a long way since the DLT. Unfortunately, most of the
work is undocumented as it has been done in-house by the big companies, e.g.
Vicon, Motion Analysis Corp. For a long time these companies used Andy
Dainis' algorithm, AMAS, but many have since gone their own way to avoid
paying royalties to him. I suspect his is still the state of the art,
though.

Chris Kirtly
**********************************