Clustering intersections
TypingCat opened this issue · 3 comments
Test #13 showed the possibility of the clustering approach. However, the current structure trains a new clustering model every time for the last 30 samples. Even there are no intersection angle. It need to be improved.
Principal Component Analysis(PCA) results are visualized by ellipse in commit 01d010d.
All clusters store last 300 samples. For each cluster, mean and standard deviation are calculated by PCA. It represented by center and size of an ellipse. In the following, intersections are represented by 6 ellipses. Distance errors from the robot tend to be larger than angle errors.
The clustering module was tested on real robot.
This robot moves at 0.2m/s, and its path is exactly same as #17 (comment). Intersection cluster covariance were represented by yellow cylinder. All intersections were detected. The problem is the misdetection. The size of covariance or the number of samples can be a clue.
