Torque Vision is an open-source (CC BY-NC-SA 4.0) vision module designed to be easy in three core areas: easy to obtain parts, easy to assemble, and easy on the wallet ($40 + a Pi 4B). You can see more information, including the CAD files and the complete equipment list, on our website. We recommend pairing the hardware with PhotonVision; using the equipment listed, you should achieve ~15fps AprilTag detection running at 1920x1080 with ~2-3 decimation.
I’m a programmer-at-heart using SOLIDWORKS, so if you would like to make improvements to the CAD (I know that the fan screws are missing), feel free to post the updated files here, and I can upload them to the website (with attribution).
If I can be of any use in sourcing parts, assembling, or editing the CAD, don’t hesitate to contact me. If your team doesn’t have the resources to print or source some of the parts, let me know, and we can work out a solution.
Thank you, and best of luck this season!
Jack Pittenger
Texas Torque
To be more specific - Autofocusing violates the pinhole camera model photonvision is currently using for 3d pose estimation. I’d expect at least some inaccuracies driven if your focal length or distortion is different between when you calibrate and when you detect targets.
The extent to which this change is meaningful to a particular robot or application isn’t something I’m sure of yet, but curious if you’ve measured it?
I don’t use autofocus for my build, so I haven’t measured it. The camera included in the equipment list works quite well without autofocus.
The case design should allow software to use the camera’s autofocus, but I haven’t played around to see if Photonvision is ready for that. It would be interesting to see if another camera model would be better at handling this (for example, KB4). I don’t fully understand all those intricacies, so I can’t comment further.
