[D] How do I get certain number of frames before every point-of-impact in a tennis game video?

I am trying to follow the steps of database creation that was followed in [this research paper](https://www.researchgate.net/publication/336659904_Prediction_of_Future_Shot_Direction_using_Pose_and_Position_of_Tennis_Player).

In the paper, they have stated they use sequential data as input for the model they are developing, wherein 20 frames are taken before every point-of-impact in a tennis match.

How can I go about doing that? Anyone could explain in as easy words as possible or guide me to an appropriate medium?

Quoting the paper:

>For this dataset, we used a tennis …

game impact machinelearning tennis video