Baseball Throw Detection

Hi and thanks for a great tool. I’m looking for guidance on detecting a baseball throwing motion on a user wearing a wrist based device. I’ve been recording gyro and accelerometer. So, velocity and rotation would be key. I’m wanting some guidance on recording the data. What I’ve been doing is record single throw motions and variations of that. So, throw, underhand, sidehand and neutral. These are short single throw motions with accel and gyro (something like 100 hz for 4 seconds).

  1. Is this the best way to approach it?
  2. what’s the best way to split training and testing data.
  3. I’m using ingestion api. what are the unit types for accelerometer and gyro?


Hi @davidloew!

  1. Yes, I think 4 second samples are good, probably slice them up in 3 second windows if that still gives a good indication of the throw with maybe 300 ms. window overlap.
  2. 80% training / 20% testing would be a good split, and make sure that the dataset is balanced (you have the same amount of data in every class).
  3. We use m/s2 for accelerometer data, but have stepped away from actually using the units to render anything in the UI right now. So just use NA for gyro data.

Thank you for the quick response. I’ll set it up that way and let you know how it goes!

Awesome! Yes, would be super interested to see how this pans out.