Center circle detection for camera calibration

[tuxbotix]

Why? What?

Provides 3 features, could be splitted but makes it harder to test all together.

1. Center Circle detection with a RANSAC implementation.
   • Canny edge detection or Segments used to get points
   • A variant is provided to handle transforming from Pixel (ellipse) to ground (circle) prior to fitting.
2. Integration into calibration measurement provider.
3. Twix panel to view the detections, a parameter in the circle detection cycler tuning_mode triggers detection on every circle regardless of CalibrationCommand.

This was originally developed on top of work on branches/ PRs (#695)

How it works: from RC2023 Bordeaux

Differences from the line-RANSAC:

1. Uses 3-point circle formula to get a sample circle
2. Early-rejection of candidates -> uses radius, minimm distance between points, etc.
3. Performing the fit only on a sample of the population (if large enough) to make the iterations reasonably fast.

Testing via #1508

Fixes #

ToDo / Known Issues

• Avoid fitting to small arcs - bottom image of the above example. (3 points selected randomly should be sufficiently spaced out)
• [ ] Alternatively, construct the circle from a minimum-sized arc (i.e. 100 degree continuous arc with radius R)
  • This can be more robust as we can check all points fitting in this region to be continuous or at least mostly continous.

Ideas for Next Iterations (Not This PR)

How to Test

• Deploy on a Nao or Webots
• Disable head-unstiff (add toggle between old calibration mode and new calibration mode #1159 ) parameter!!
• place the Nao 1.5~m facing and away from.center circle.center.
• enable calibration mode

You can subscribe to:

• calibration_command to see the state changes
• calibration_inner.measurements additional from calibration controller to see the measurements collected
• Output last_calibration_corrections

Tuning mode (vision parameters)

The parameters are tuned to a decent state already, but if needed it can be tuned further.

Parameters

• tuning_mode - keeps running the cycler even when no calibration command is recieving from the calibration controller.
• ransac_circle_minimum_circumference_percentage helps to reduce false positive.hits on T junctions, etc with very few points.
• get_edges_from_segments - uses segments instead of Canny to generate points for RANSAC Overall speed is quite decent but not as good as Canny based approach
• gaussian_sigma controlls the Gaussian blur strength. Best left untouched.
• canny_* - controlls canny edge detection. Can be tuned for lighting conditions, but these values should work already.
• Other parameters are already tuned and does not need changing.

How-to

• Deploy the branch on a Nao or Webots
• Place the nao 1.5m~ from the center circle while facing it.
• Enter calibration mode (but keep the head unstiff parameter add toggle between old calibration mode and new calibration mode #1159 to help tuning)

Twix:

• Open an image panel, there is an overlay for calibration center circle (not fully tested with latest Twix changes!!)
• subscribe to:
  • main output calibration_center_circles of the given cycler to see raw.output
  • calibration_center_circle_detection.circles_points_pixel_scores additional to see the score for each circle.
• open a parameters panel
  • Look for calibration_center_circle_detection
  • Enable tuning_mode for top or bottom (not both as this would slow things down.
  • if no circles are detected, try to reduce ransac_circle_minimum_circumference_percentage to 0.1 or something, it should

[knoellle]

LGTM

Whoops, wrong PR

[tuxbotix]

I'm cleaning up this branch to remove the testing stuff, hopefully can get it ready today.

[schmidma]

Is this still on track for RC24? If we cannot manage for RC24, I'd suggest to remove the label 👍

[tuxbotix]

Is this still on track for RC24? If we cannot manage for RC24, I'd suggest to remove the label 👍

Sorry, I didn't see this. I'm afraid there won't be enough time to properly review and merge it to main. But it is ready for "beta" testing now ;) I'll update the how-to later today.

I tested the workflow and detection seperately and with initial integration a week ago, it worked but needed cleaning up.

[pejotejo]

@tuxbotix what's the state with this pr

[tuxbotix]

Currently being tested under #1508 . The scope of this PR works as advertised and any refinement steps should be in a separate PR IMHO

I discussed with @oleflb last week and while it is easier to test calibration with an all in one branch, reviewing and merging would be annoying 😬.

This branch could be revised to match the scope of the title and merged. Then #1508 will have the integration part to the calibration pipeline and possible refinements.

I'll finish cleaning this up today.