Good things come to those who wait
Posted on February 20, 2022
This weeks' match was canceled and will be repeated in two weeks.
Due to a bug in the referee code, the simulation has crashed shortly after the second half. The game versus the Hamburg BitBots was restarted, but crashed with the same bug again. Instead of simulating the game again, the organizers will focus on fixing the referee. Our game was rescheduled and will be shown in two weeks, along the other games of the 6th game day.
Though it is sad not to be able to cheer on our robots today but this is a good sign! During the last two games our robots basically did not do anything but dance blindly on the field looking for the balls. Our robots are actually doing something now!
We have improved our ball detection software immensely. Our software stack finds white blobs successfully using integral images. The challenge is now to classify these white blobs. We want to know if there is a ball inside the white blob, or something else (This could later also be extended to detect other field features, like line crossings, or goal poles).
In the past the light condition, the exact field and the ball we played with was announced before the game. This is why we simply took 1-3 of white blobs as ball reference samples and compared every detected white blob with these references. As a comparison metric we used the Kullback-Leiler-Divergence. This metric doesn’t work with changing environments, so we needed a new strategy.
We have implemented an algorithm similar to the Viola-Jones object detection framework using AdaBoost. As basis for AdaBoost we used 3 different weak-classifier: KL-Divergence, and the pixelwise MSE between two images.. Measuring the performance to a test set, we measured an accuracy of 90%. A 90% accuracy is not very high. Especially with many false positives, our robots target objects on the field that are not balls, wasting a lot of time.
@BitBots, as promised we went down in flames, little did we know the simulator wasn’t fire proof. We are prepared and are looking forward to our derby in two weeks. Note that we will compete with the software version of this (18th February) Friday. So no backsies.
We also are looking forward playing against our next regular opponent - MRL-HSL(Iran).
We will have two matches on the 06th March 2022. They will be played in the morning. Watch us live on: https://www.twitch.tv/robocuphumanoidfielda The official game schedule can be found here: https://humanoid.robocup.org/hl-vs2022/humanoid-league-virtual-season-2021-22/
The full table after the (partial) fifth tournament day:
|05.||01. RFC Berlin||7||2||1||1||11:12||-1|