Good morning,
Most probably due to a too low errorpoints threshold my monsterBorg went into the avoidance(Backup) mode all the time during yesterdays challenge.
Can you give some hint, maybe on how high the lower threshold should be?
That would be great!!
regards,
xStatic
Tuning the error points is very difficult, the standard number was a compromise between removing false detentions and seeing robots early enough.
The best way to improve the detection would really be to replace the error points method with something else a bit more reliable, such as matching a robot shape in the image.
If you do want to try and improve the existing detection there are a few things you can try:
Changing
overtakeThreshold
to tune the detectionChanging
cropY1
to see further aheadSlowing down to allow more time to detect robots
Averaging error points over a few frames
This would need some additional code, but you could change the detection so it needs to see the detection limit for a few frames in a row.
As all of these things affect each other you will probably need to play with the settings to see what happens. The best hing to do is test your code with some of the images taken from the MonsterBorg driving around the track and checking when the code decides their is a robot to overtake against the picture it is being shown.
I already implemented a template matching function and thought of the overtake logic from the standard code rather as a backup to fastly trigger the "Crashed" logic if the template matching fails.
Since my monsterBorg was like nearly all the time in "CRASHED" mode, and I only allow the TM code to run while the imageMode == FOLLOW_TRACK ... My TM code didnt even have a chance to proof its avoidance skills... too bad, since its matching rate was quite good during testing.
I think I won't directly abandon the idea of the standard overtaking logic as a kind of backup behavior, but I will set the threshold significantly higher!
Thanks for your extensive and informative answer!! =)
Kind regards,
xStatic
Add new comment