Hi,
I have a great deal of experience in deep learning and detection. The task you are describing is an action recognition from series of images (a video) which is a very interesting concept in machine learning.
Do you have any training/testing data available so we can use that in the development or is finding it a part of the project?
Do you need any state detection? Like do you need to know at which percent is the throw right now, or just the fact that the throw happened?
Hope we will work together!