• Categories
  • Recent
  • Tags
  • Popular
  • Users
  • Groups
  • Register
  • Login
ModalAI Forum
  • Categories
  • Recent
  • Tags
  • Popular
  • Users
  • Groups
    • Register
    • Login

    Stereo Cameras are highly inaccurate

    Ask your questions right here!
    3
    11
    647
    Loading More Posts
    • Oldest to Newest
    • Newest to Oldest
    • Most Votes
    Reply
    • Reply as topic
    Log in to reply
    This topic has been deleted. Only users with topic management privileges can see it.
    • D
      Daniel Ryu
      last edited by 26 Aug 2021, 23:01

      I calibrated my stereo cameras twice according to these instructions. https://docs.modalai.com/calibrate-cameras/#calibrate-stereo-cameras. However, the output I receive from running

      voxl-vision-px4 -s
      

      is very inaccurate. How can I improve the stereo camera's object detection accuracy? I used a 4x11 acircles board with 59.5 mm distance separating each circle.

      Testing environment
      Screenshot from 2021-08-26 15-43-28.png

      First calibration results
      Screenshot from 2021-08-26 15-11-56.png

      Second calibration results
      Screenshot from 2021-08-26 15-37-28.png

      voxl-vision-px4.conf
      Screenshot from 2021-08-26 15-43-56.png

      1 Reply Last reply Reply Quote 0
      • D
        Daniel Ryu
        last edited by 27 Aug 2021, 18:44

        Clarifying notes:

        Same problem seen on another m500 drone.
        Calibration card used has a spacing of 98.8 mm between the center of each circles (acircles card). 4x11

        1 Reply Last reply Reply Quote 0
        • D
          Daniel Ryu
          last edited by 27 Aug 2021, 18:46

          Image from iOS (2).jpg

          1 Reply Last reply Reply Quote 0
          • C
            Chad Sweet ModalAI Team
            last edited by 29 Aug 2021, 19:17

            you need to get the calibration target professionally printed. All of those inaccuracies in what you have put together are going to create issues with calibration. Any office printing place can print a card like that from PDF on stiff backing. I don't think what you have is going to work that well

            D 2 Replies Last reply 30 Aug 2021, 17:31 Reply Quote 0
            • D
              Daniel Ryu @Chad Sweet
              last edited by 30 Aug 2021, 17:31

              @Chad-Sweet I'll try again with a card professionally printed card and get back to you

              1 Reply Last reply Reply Quote 0
              • D
                Daniel Ryu @Chad Sweet
                last edited by Daniel Ryu 3 Sept 2021, 20:30 3 Sept 2021, 20:29

                @Chad-Sweet Here are the results with the pictured calibration board we used (spacing 117.5 mm center to center). As you can see, we still have a detection at 0 degrees at 0.2 meters, when there should be nothing at all.

                Screenshot from 2021-09-03 13-22-39.png

                Screenshot from 2021-09-03 13-24-55.png
                Image from iOS.jpg

                1 Reply Last reply Reply Quote 0
                • D
                  Daniel Ryu
                  last edited by 3 Sept 2021, 22:42

                  Note: Every other angle appears to be somewhat accurate, everything except the measurements at 0 degrees

                  1 Reply Last reply Reply Quote 0
                  • J
                    jaredjohansen
                    last edited by 8 Sept 2021, 23:24

                    An update on our efforts in this regard.

                    We have now attempted to calibrate 4 different VOXLs. In one case, we calibrated the VOXL board twice (with similar results).

                    We have tried with three different circles grids (two of which were professionally printed).

                    In all cases, at the zero degree angle, there is an obstacle constantly detected at 0.2 meters. We do not know why this is. We presume this is a problem. Is it? If not, why?

                    In three of the four cases, the detections at all other angles are inaccurate. In these three cases, in general, they "sense" when an obstacle is within their bin range, but this is not always the cause: sometimes miss detecting obstacles. In the three cases where the detection are inaccurate, they can be off by several meters. We expect that autonomous navigation that relies on the output of these bins would cause problems. Do you have any additional guidance on how to get this working more accurately?

                    1 Reply Last reply Reply Quote 0
                    • C
                      Chad Sweet ModalAI Team
                      last edited by 13 Sept 2021, 23:52

                      This looks like it could be a bug, we're looking into it. In the meantime you can just comment out using the center bin

                      1 Reply Last reply Reply Quote 0
                      • C
                        Chad Sweet ModalAI Team
                        last edited by 15 Sept 2021, 17:34

                        A potential fix for this has been pushed to the dev branch

                        change

                        D 1 Reply Last reply 17 Sept 2021, 00:14 Reply Quote 0
                        • D
                          Daniel Ryu @Chad Sweet
                          last edited by 17 Sept 2021, 00:14

                          @Chad-Sweet The fix in the dev branch successfully populates the 0th degree bin for obstacle detection. We are now able to run VOA mostly successfully, with one observable bug with the voxl-camera-server/voxl-dfs-server. Please see this post for details: https://forum.modalai.com/topic/450/intermittent-crashing-of-voxl-camera-server-during-voa-testing?_=1631837674205

                          1 Reply Last reply Reply Quote 0
                          1 out of 11
                          • First post
                            1/11
                            Last post
                          Powered by NodeBB | Contributors