Multi-agent trajectory prediction is a common problem seen in a variety of settings. We focus on the game of basketball in which the players act as agents and the ball moves according to actions taken by the players. We implement a transformer encoder-decoder model to predict ball trajectory given an input sequence of player and ball position data. We propose two experiments, physics ball 3D movement prediction, and game-level ball 2D movement prediction. Our model is able to capture 2D ball movement over 2 seconds reasonably well and some simple physics ball movement over 0.6 seconds.
Results: Colab Link
- Blue: Input Trajectory Sequence
- Red: Ground Truth Output Trajectory Sequence
- Green: Predicted Output Trajectory Sequence
