@nguyen_thanh_tin there is a code skeleton that should get you to the point in processing the frames. The ground truth was built using homography between the rink image and the images. There are a lot of approaches but here are some examples;
There are a lot more attempts at this problem as well.