for a university project I want to analyze soccer tracking data. The input of the system will be 5-10 scenes which describe a certain, not yet defined, technical maneuver (play through the center, …). The model should return similar scenes in other matches.
All is based on tracking data. It includes x,y coordinates of each player and the ball every 40ms of a game (frame). Flags like ball possession and a few others are also included.
Since I’m not super deep into machine learning but got a general understanding I just wanted to verify my idea of a possible approach and maybe get a few tips.
* Tracking data of multiple (100-500) games
* Input: 5-10 example scenes (tracking data, no video)
* Output: similar scenes on new tracking data
Since I don’t have classified data I want to use unsupervised learning to cluster (pre defined) scenes of a game and query with the given examples to get similar scenes from the model.
Approach in my head:
* Slice all the games in scenes (300-500 scenes per halftime)
* Aggregate the given example scenes into one
* Query model with aggregated scene
In my opinion, the most challenging task will be to find appropriate features. Since I only have theoretical knowledge of ML from university, it would be great to know if this would work in reality. I would appreciate hints to promising tools/libraries (sci kit learn, tensorflow, ..) or algorithms which I should look into