Closed val-ba closed 3 weeks ago
I changed the target to the other PR, so it is easier to review as it only shows the changes from this PR now. Merge target should move to main once #468 is merged.
Also, I think that the error increases in a non-linear fashion.
If we plot the distance of the ipm intersection for a camera pitch from 0-90 deg we get the following relationship
When we now plot the impact of a 1 deg error in the camera pitch for each intersection distance, we get the following plot
This indicates a non-linear relationship between the observation distance and the noise. We need to see if it is significant enough and how we can parametrize it.
Zoom showing the relevant part:
x^2 seems to be a reasonable approximation
Tested in simulation
Summary
We want the measurement certainty to be lower if the ball is further away as these measurements are less reliable.
Proposed changes
We add two parameters: a reference distance and a noise incrementation factor. The first one is used to put the ball distance into perspective and the second one is how much the ball distance influences the measurement certainty. In the ball filter, we now calculate the measurement certainty matrix based on that.
Also ignores measurements that are too uncertain. Also general refactoring and fixes (e.g. in the tf_buffer).
Related issues
Closes #424
Checklist
colcon build