Add a feature to blur someones face immediately. You'll get better traction if there are no faces in your dataset. You can use the code from what was done to the summer riots pictures.
I didn't mention it earlier, but I'm processing the video locally on the user's device. The video and derived data are also stored locally and I'm not using any user data to construct a dataset.
So far, the only time data leaves the user's phone is to allow them to share the video and data to others (ex. for social media or to coaches).
I'll definitely look into face blurring if I try to build a dataset in the future.
