Loading ...
Sorry, an error occurred while loading the content.
 

About ball detection when ball is clipped

Expand Messages
  • tencuid
    Hello, I m working with the BallDetectionGenerator module and VisionObjectEvent, in order to know the position and distance of the ball. When there is no
    Message 1 of 4 , Apr 1, 2005
      Hello,
      I'm working with the BallDetectionGenerator module and
      VisionObjectEvent, in order to know the position and distance of the
      ball. When there is no clipping of the ball (the ball is "in focus" of
      the camera), all works OK. But when there is any clipping I detect
      some problems, as BB and centroid are not always correctly detected.

      For example if the ball is clipped mostly at the right, but also at
      the top and bottom, to show only about a 25% of the total area (BB
      should be aprox. X1=0.33,X2=1,Y1=-0.769,Y2=0.769),it shows X1=0.32
      (OK), X2=0.99 (almost OK), but Y1=-0.5 (¿?) ,Y2=0.4(¿?). It's as it
      gets only the rectangle central area, and forget the arc areas.

      I wonder if there is a noise filtering algorithm at
      BallDetectionGenerator which tries to find only nearly-square BB
      regions or adjust pink regions found to a nearly-square BB form. In
      that case where I could find that algorithm in order to eliminate it
      (I consider all pink regions would be the ball's so I want to allow
      rectangle regions,and not only square regions.

      Thanks in advance for your help.
    • tencuid
      I think I ve found the answer to my own question. It seems to be a problem of ilumination and color-tuning. When the ball is too near the camera, there are
      Message 2 of 4 , Apr 4, 2005
        I think I've found the answer to my own question. It seems to be a
        problem of ilumination and color-tuning. When the ball is too near the
        camera, there are shadows over the edges of the camera, so the
        segmentation stage of the vision pipeline doesn't identify them as a
        pink region. I've noted that those areas are not identified as pink
        regions prior the BallDetectionGenerator stage.

        This makes me unable to identify when the ball is near AIBO but very
        clipped, or when it is far and near a side of the camera as, in the
        first situation the rounded edges of the ball are eliminated by
        shadows. Any suggestions?

        Somebody working with AIBO vision,..what kind of ilumination are you
        using? (I have white light fluorescent tubes at the ceiling)
      • Harm Aarts
        I have some experience with the camera of the AIBO. In our lab we have the normal cieling lighting, which are those tubes you talk about, but with only these
        Message 3 of 4 , Apr 4, 2005
          I have some experience with the camera of the AIBO. In our lab we have
          the normal cieling lighting, which are those tubes you talk about, but
          with only these lights preformance is poor or unstable. We mounted
          4x500W constuction lights on the ceiling for additional luxes.
          On your other question I can not help you. Sorry.

          gr
          Harm

          tencuid wrote:
          >
          > I think I've found the answer to my own question. It seems to be a
          > problem of ilumination and color-tuning. When the ball is too near the
          > camera, there are shadows over the edges of the camera, so the
          > segmentation stage of the vision pipeline doesn't identify them as a
          > pink region. I've noted that those areas are not identified as pink
          > regions prior the BallDetectionGenerator stage.
          >
          > This makes me unable to identify when the ball is near AIBO but very
          > clipped, or when it is far and near a side of the camera as, in the
          > first situation the rounded edges of the ball are eliminated by
          > shadows. Any suggestions?
          >
          > Somebody working with AIBO vision,..what kind of ilumination are you
          > using? (I have white light fluorescent tubes at the ceiling)
          >
          >
          >
          >
          >
          >
          >
          >
          >
          >
          > Yahoo! Groups Links
          >
          >
          >
          >
          >
          >
          >
        • Ethan Tira-Thompson
          It sounds like you ve identified the shadows as the main problem, but I wonder if the blue tinges of the corners of the images might also be contributing...
          Message 4 of 4 , Apr 4, 2005
            It sounds like you've identified the shadows as the main problem, but I
            wonder if the blue tinges of the corners of the images might also be
            contributing...

            Since this is a constant effect, it could be corrected with a vision
            pipeline stage. I know some robosoccer groups have done a bit of color
            calibration to correct for the blue tinge, but of course this costs
            precious CPU cycles, so you can debate whether it's better to spend
            time cleaning up raw data early on, or spend time doing more
            complicated interpretation of noisy processed data later.
            The German Team has the best treatment of this issue that I've seen, in
            section 3.2.3 of their 2004 team report (full version):
            http://www.germanteam.org/GT2004.pdf

            I may have to address this topic myself later this semester for my
            current vision class, but no promises...
            (If I do it, I'll set it up as an optional thing so you can turn the
            stage on or off as needed)

            -ethan


            On Apr 4, 2005, at 7:37 AM, tencuid wrote:

            >
            > I think I've found the answer to my own question. It seems to be a
            > problem of ilumination and color-tuning. When the ball is too near the
            > camera, there are shadows over the edges of the camera, so the
            > segmentation stage of the vision pipeline doesn't identify them as a
            > pink region. I've noted that those areas are not identified as pink
            > regions prior the BallDetectionGenerator stage.
            >
            > This makes me unable to identify when the ball is near AIBO but very
            > clipped, or when it is far and near a side of the camera as, in the
            > first situation the rounded edges of the ball are eliminated by
            > shadows. Any suggestions?
            >
            > Somebody working with AIBO vision,..what kind of ilumination are you
            > using? (I have white light  fluorescent tubes  at the ceiling)
          Your message has been successfully submitted and would be delivered to recipients shortly.