A lightweight real-time system for tracking soccer ball movement in broadcast footage using a Kalman Filter and YOLOv11. Designed for handling partial occlusions, missed detections, and camera motion using confidence-aware prediction.

⚽ Built with Football Love on highly Caffeinated Code Sessions ☕

• Detect soccer ball, players, goalkeepers, and referees in match footage.
• Predict the ball’s position even when occluded or not detected.
• Compare predicted vs. actual positions using error metrics.

• YOLOv11 (via Ultralytics) for detection
• Kalman Filter (pykalman) for trajectory tracking
• Supervision for annotated video rendering
• Roboflow for model hosting
• ONNX Runtime GPU acceleration on Colab

1. 🔍 Detection
   YOLOv11 detects objects in each video frame (ball, players, etc.).

2. 📈 Tracking
   A Kalman Filter smooths and predicts ball positions based on confidence scores.

3. 🎥 Visualization
   Annotated video is generated with actual vs. predicted positions.

4. 📊 Evaluation
   Errors (mean, RMSE, max) between prediction and ground truth are visualized and stored.

Tracks [x, y, vx, vy, ax, ay] using a constant acceleration model:

state = [position_x, position_y, velocity_x, velocity_y, acceleration_x, acceleration_y]

Object-Tracking-EKF-YOLO/
├── track_ball.py
├── kalman_filter.py
├── yolov11_training.ipynb
├── inference.ipynb
├── vids/
│   ├── 121364_0.mp4
│   └── 121364_0_result_1.mp4
└── positions.csv

• YOLOv11 via Ultralytics
• Roboflow dataset: football-players-detection-3zvbc
• Supervision library for annotations