RoboCupHumanoid

In this project, we are aiming to implement [1] and improve it using Conv + LSTM layer for detection and tracking of the soccer ball for <a href="https://www.robocuphumanoid.org/">RoboCup Humanoid League</a>.

Dataset

example_data.csv contains information about training images with the following fields:

<ul> <li><b>image_file</b> - image location,</li> <li><b>width</b> - width of the image,</li> <li><b>height</b> - height of the image,</li> <li><b>label</b> - label of the object,</li> <li><b>xmin</b> - top left x-coordinate of rectangle around object,</li> <li><b>ymin</b> - top left y-coordinate of rectangle around object,</li> <li><b>xmax</b> - bottom right x-coordinate of rectangle around object,</li> <li><b>ymax</b> - bottom right y-coordinate of rectangle around object.</li> </ul>

To extract and label images we are using <a href="https://imagetagger.bit-bots.de/images/">Image Tagger</a> We used <a href="https://pjreddie.com/darknet/yolo/">YOLO</a> for automatic ball detection and then we manually verified each the detection.

<b>Note</b>: One image file may contain multiple objects of different types.

Running the tests

• Train sweaty python train.py --batch_size=16 --alpha=1000 --model_name=alpha1000 --epochs=50
• Test Sweaty python test.py --load=pretrained_models/alpha1000_epoch_50.model --testSet=data/test/ --trainSet=data/train/

References

[1] Fabian Schnekenburger, Manuel Scharffenberg, Michael Wulker, Ulrich Hochberg, Klaus Dorer Detection and Localization of Features on a Soccer Field with Feedforward Fully Convolutional Neural Networks (FCNN) for the Adult-Size Humanoid Robot Sweaty