Detect orthogonal lines using RANSAC #1314
Conversation
julianschuler
force-pushed
the
in-the-corner
branch
4 times, most recently
from
de0a1d2 to
cf101c5
Compare
julianschuler
changed the title
julianschuler
force-pushed
the
in-the-corner
branch
from
cf101c5 to
181f222
Compare
|
einmal mit kalibrierter Kamera testen |
julianschuler
force-pushed
the
in-the-corner
branch
from
181f222 to
5219fe5
Compare
julianschuler
force-pushed
the
in-the-corner
branch
from
35a1c74 to
a2c9d72
Compare
| let clockwise_normal_vector = Direction::Clockwise | ||
| .rotate_vector_90_degrees(direction_vector) | ||
| .normalize(); |
There was a problem hiding this comment.
Not your change, but I'd expect the rotate_vector_90_degrees to be a function of a vector taking the Direction as argument...
Resolve when read, or maybe even open an issue for that? What do you think?
| pub direction1: Vector2<Frame>, | ||
| pub direction2: Vector2<Frame>, |
There was a problem hiding this comment.
This naming has 90% overlap in the field names. This is easy to mistype...
| line1 | ||
| .squared_distance_to(point) | ||
| .min(line2.squared_distance_to(point)) |
There was a problem hiding this comment.
extract to variable to increase readability
| PathIntrospect, | ||
| PathDeserialize, | ||
| )] | ||
| pub struct TwoLines<Frame> { |
There was a problem hiding this comment.
Why isn't this called Corner if the docstring calls it corner?
| #[default] | ||
| None, | ||
| Line(Line2<Frame>), | ||
| TwoLines(TwoLines<Frame>), |
There was a problem hiding this comment.
same here: maybe call it Corner?
| if two_lines.direction2.norm() > f32::EPSILON { | ||
| Self::TwoLines(two_lines) | ||
| } else { | ||
| Self::Line(line) | ||
| } |
There was a problem hiding this comment.
What situation do you want to capture here? Having the same point twice?!
| { | ||
| unused_points | ||
| .into_iter() | ||
| .filter(|point| self.squared_distance_to(**point) <= maximum_score_distance_squared) |
There was a problem hiding this comment.
You might be able to do something like
| .filter(|point| self.squared_distance_to(**point) <= maximum_score_distance_squared) | |
| .filter(|&point| self.squared_distance_to(point) <= maximum_score_distance_squared) |
| #[derive(Clone, Debug, Default, Deserialize, Serialize, PartialEq)] | ||
| pub enum RansacFeature<Frame> { | ||
| #[default] | ||
| None, |
There was a problem hiding this comment.
Why is there a None feature type?! This sounds weird to me...
Why? What?
This PR implements detecting orthogonal lines in the field using RANSAC. This improves the placement of line segments near corners.
Depends on #1334.
Since our
image_to_groundis currently resulting in orthogonal lines not always being orthogonal, it may currently result in worse line fittings.This feature is currently gated behind the
ransac_fit_two_linesparameter, which is false by default.ToDo / Known Issues
ransac_iterationsshould be increased.Ideas for Next Iterations (Not This PR)
How to Test
Enable recording and upload to a NAO. Watch the replay and have a look at the image panel with the line detection overlay enabled. The green lines show the fitted orthogonal lines, the red ones the fitted lines. Note that only the blue lines are the extracted line segments used for localization.