Ball Detection
Just like line detection, our ball detection algorithm is composed of two stages: region proposal and classification.
For the region proposal stage, we use a version of the Hough transform for circles. Due to the high similarity between the limbs of the robots and the ball, ball proposals which significantly overlap with robot detections are filtered out.
The generated ball proposals are then filtered using a CNN for image classification. As the network was trained on 32x32 Y-channel image patches from different datasets, for each ball proposal a square patch is cropped around the center and resized to match the input dimensions of the classifier.
The model was trained using Keras and deployed using Tensorflow Lite.