Kinect perspective correction?

    Dec 04 2011 | 3:44 pm
    I am dealing with same this problem.
    Kinect is on the ground and want to correct the perspective like seeing in the front of person. Here is dtr's thankful patch contribution.
    but the shadows won't make the depth visual look normal as a front view.
    So If I use 2 Kinect (on the ground-look up, and over the head-look down), Can I make the corrected perspective like seeing in front of the person?
    Any suggestion please! thanks

    • Dec 05 2011 | 2:36 pm
      that should work. calibrating the 2 cams so they overlap nicely might be a pain though.
      i'm doing a similar thing but with kinect skeleton tracking. in my motion tracked performance i need 360° freedom of movement so i need to prevent limbs being obscured to the camera. i use a 2nd kinect at 120° angle. both kinects perform skeleton tracking and i merge the 2 results together. it's not super precise but good enough for my needs right now.
    • Dec 05 2011 | 8:28 pm
      How do you merge the 2 skeletons together ?
    • Dec 06 2011 | 9:58 am
      I've got 2 computers each running a Processing sketch which uses the SimpleOpenNI library for Kinect skeleton analysis. They both send their joint coordinates over OSC to Max running on one of the computers.
      I draw the joints to a 3d GL context so I can manually rotate and translate the 2 skeletons till they overlap pretty good (compensating for the 120° horizontal angle, vertical angle, etc). I do this visually though I'm sure there's more scientific methods to do this but this works for me for now. (Any tips on this?)
      Then I have some logics going on to merge the 2 skeletons in 1 resulting skeleton. I iterate through all of the joints. If I have a valid reading (according to the confidence parameter) in both skeletons I take the average coordinate. If one is valid and one is invalid I take only the valid one. If none are valid I don't update the joint.
      Lastly I apply smoothing to the resulting skeleton.
      (Btw, I hope that at some point the OpenNi/NITE libraries will have this built in, as well as fixing the 2 cam's on 1 computer skeleton tracking issue. Doesn't work as it should for now.)
    • Dec 06 2011 | 1:59 pm
      thanks for the info! I never thought of matching of the skeletons by hand. I don't have the math skills either to do this by calculus. I suppose this involves solving a system of linear equations..might get nasty..
      (@djtoshi182: sorry for hijacking your thread!)