⚽ Football Player Tracking

Detecting and segmenting football (soccer) players from broadcast frames using three complementary computer-vision approaches.

Author: Mohammad Javad Maheronnaghsh Dataset: Football Player Segmentation (Kaggle)

---

📋 Overview

This project explores player detection and segmentation on football match images. After a short literature review (Google Scholar, Kaggle, GitHub), three different methods were implemented and compared, ranging from off-the-shelf object detectors to trained segmentation networks and foundation models.

#	Method	Type	Framework	Idea
1	YOLOv8 (Ultralytics)	Object detection	PyTorch	Bounding boxes around each player + jersey-color / grass-mask analysis
2	U-Net & U²-Net	Semantic segmentation	TensorFlow / Keras	Pixel-level player masks, trained on the dataset annotations
3	Segment Anything (SAM)	Promptable segmentation	PyTorch	Zero-shot automatic mask generation from a foundation model

---

🔄 Process Pipeline

flowchart TD
    A[Football match frames] --> B[Preprocessing<br/>resize · BGR→RGB · COCO masks]
    B --> P1
    B --> P2
    B --> P3

    subgraph P1 [Project 1 · YOLOv8 Detection]
        C1[YOLOv8x pretrained] --> C2[Bounding boxes<br/>conf = 0.7]
        C2 --> C3[Per-player crop<br/>+ jersey color]
        C3 --> C4[Grass mask removal]
    end

    subgraph P2 [Project 2 · U-Net / U²-Net]
        D1[Encoder–Decoder<br/>+ skip connections] --> D2[Train on masks]
        D2 --> D3[Pixel-wise prediction]
    end

    subgraph P3 [Project 3 · Segment Anything]
        E1[SAM ViT-H checkpoint] --> E2[Automatic mask generator]
        E2 --> E3[Filter masks by area]
    end

    P1 --> R[Compare results]
    P2 --> R
    P3 --> R

Loading

---

🧩 Methods in Detail

1. YOLOv8 — Detection & Analysis

Uses the pretrained yolov8x model from Ultralytics to detect players (COCO person class) with a confidence threshold of 0.7. Each detection returns two corner points (x, y) and (x2, y2) plus a class and confidence score. The detected crops are further analysed to estimate jersey colors and to build a grass mask (background subtraction) so player pixels can be isolated from the pitch.

2. U-Net & U²-Net — Segmentation

• U-Net: a U-shaped encoder–decoder with skip connections (contracting path → expansive path) producing pixel-level masks.
• U²-Net: a nested U-structure built from Residual U-Blocks (RSU-L) instead of plain convolutional blocks — a six-stage encoder, five-stage decoder, and a saliency-map fusion module.

COCO-style polygon annotations are converted to binary masks (via imantics), images resized to 512×512, and both models are trained and then compared side-by-side against the ground truth.

3. Segment Anything (SAM)

Uses Meta AI's Segment Anything Model with the ViT-H checkpoint and the SamAutomaticMaskGenerator to produce zero-shot masks, then filters them by area to keep player-sized regions.

---

🖼️ Sample Results

---

🚀 Getting Started

All code lives in football-tracking-final_release.ipynb. The notebook was developed on Kaggle (with the dataset mounted under /kaggle/input).

# clone
git clone https://github.com/mjmaher987/Football-Player-Tracking.git
cd Football-Player-Tracking

# key dependencies
pip install ultralytics tensorflow opencv-python imantics
pip install git+https://github.com/facebookresearch/segment-anything.git

Then open the notebook and run the section for the approach you want to explore (each project is self-contained).

Standalone Kaggle notebooks

• YOLOv8 player detection
• Segmentation with U-Net & U²-Net
• Segment Anything — how to

---

📄 License

See LICENSE.