[Python] How can I predict the future landing position of a ball detected with YOLO?

I have a YOLO model that detects a baseball in a video game. For each frame, I can get the (x, y) coordinates of the ball. I want to predict the future position where the ball will cross the plate based on the first few detected positions.

What I have:

• 
  A YOLO model that outputs positions for each frame.
  
  
• 
  Sample CSV data with positions:
  

frame,x,y
1,391,91
2,377,94
3,380,87
4,381,86
5,382,86
6,382,86
7,384,85
8,386,84
9,511,215


What I tried:

• 
  Used an LSTM model with Keras and StandardScaler from scikit-learn.
  
  
• 
  The model predicts almost the same landing spot for every pitch, so it doesn’t work.
  

Problem:
I don’t know whether I should continue with a machine learning approach or switch to a physics-based approach (e.g., fitting a curve to the early positions). I want a reliable way to predict the ball’s future landing position from the first few frames.

Goal:
Given the first N frames of ball position data, predict the (x, y) coordinates where the ball will land.

Additional info:

• 
  Screenshot of YOLO detection: https://i.sstatic.net/0kIF0FCY.png
  
  
• 
  Sample dataset and previous code: Google Drive link
  

Question:

• 
  Should I use ML (like LSTM or regression) or physics-based trajectory prediction?
  
  
• 
  Can someone provide an example approach to predict the landing point from early frames?
  

Continue reading...