No announcement yet.

Filtering multi-view 2D, then resulting 3D kinematic data ?

  • Filter
  • Time
  • Show
Clear All
new posts

  • Filtering multi-view 2D, then resulting 3D kinematic data ?

    Hi all,

    I'm sorry, it may have been asked already but I couldn't find it.
    I've got very noisy and low quality 2D kinematic data that I want to triangulate in order to get 3D motion.
    I'm interpolating the missing 2D data and filtering them, which gives me smoother curves. The triangulated 3D is still quite noisy, so I filter it too.

    - Are there any guidelines about how to process kinematic data?
    - And especially, do people and softwares usually filter both their 2D and their 3D?
    - Edit: When is it advised to filter the speed or the acceleration instead of the position?

    I'm looking for inspiration mostly, but also for justification from the literature if you have any.

    Thanks !
    Last edited by David Pagnon; February 10th, 2021, 01:34 PM.

  • #2
    Back in the early days of 3D motion capture, the cameras recorded 2D locations of the markers as a series of 2D points in each frame that showed the outline of each marker in terms of a measurement on the lines in the video image. This might be a dozen points on half a dozen lines if the marker was close to the camera, or just two points on one line if the marker was a long way away. The 2D processing used its guess of the outline to calculate the center of the marker so while this was not filtering, it was a variably degree of processing that had a "filtering" effect. The vectors were then combined to create 3D locations which were simply stored - I never saw any filtering of the data during the collection process. Once the data was stored as 3D locations (I'm describing the world before C3D files existed) then applications like the Boston Software and the Helen Hayes Software would read the original 3D locations from a collection of files and perform interpolation and filtering as part of the data processing.

    The way that everyone looked at this was that the data collection systems were recording what they "saw" so if there were problems you would look at the data - because it was never filtered or interpolated initially, debugging was easy - you might not like what you saw but you knew that it was the data, not something that had been manipulated and you could perform your own processing to "fix" the problem. Filtering and interpolation were usually separate processes that output results but preserved the original data samples.
    Last edited by Edmund Cramp; February 17th, 2021, 12:27 PM.


    • #3
      Originally posted by David Pagnon View Post
      When is it advised to filter the speed or the acceleration instead of the position?
      I suspect the answer depends on the data but since the acceleration is calculated from the change in position over time, I think you should filter the position data first. Once you are confident that your positions are accurate then the acceleration should be accurate.