python-opencv — Football Tracking & Computer Vision

A collection of computer vision scripts built with OpenCV, YOLOv8, SAM2, and EasyOCR. The main project is a real-time football player tracking pipeline with team assignment, jersey number detection, and roster lookup.

---

Football Tracking (football_tracking.py)

Full pipeline for analysing football footage:

• Player & ball detection — YOLOv8 trained on soccer footage
• Multi-object tracking — ByteTrack via Ultralytics
• Segmentation masks — SAM2 (Segment Anything Model 2) per player
• Team classification — colour-based clustering with sports library
• Jersey number detection — dedicated YOLOv8 detector trained on FCBerdi footage + EasyOCR fallback
• Roster lookup — maps jersey numbers to player names

Setup

conda activate /path/to/your/env
pip install ultralytics supervision easyocr sports
pip install torch==2.4.0 torchvision==0.19.0

Clone SAM2 locally (required):

git clone https://github.com/facebookresearch/sam2 /path/to/segment-anything-model-2

Run

python football_tracking.py

Output is saved to /Users/pascal-maker/Desktop/fcberdi_tracked.mp4.

To process a different video, update these lines in the script:

VIDEO_PATH  = "/path/to/your/video.MOV"
OUTPUT_PATH = "/path/to/output.mp4"
MAX_FRAMES  = None   # or set a number for a short test run

---

Jersey Number Detector (train_jersey_detector.py)

Trains a YOLOv8n model to detect jersey number regions on players. Used by football_tracking.py to give EasyOCR a tight, clean crop instead of guessing the chest region.

Dataset

Annotated via Roboflow — COCO format export goes in jerseydetection/. Classes: people, ball, jerseynumber, logo

Train

python train_jersey_detector.py

Converts COCO → YOLO format, trains for 60 epochs, saves best weights to:

jerseydetection_yolo/runs/jersey_detector/weights/best.pt

Current model performance (41 training images):

• mAP50: 0.42 — decent for a small dataset, improves with more annotations

To improve: upload more frames from your match footage to Roboflow, re-export, replace the jerseydetection/ folder, and re-run training.

---

Other Scripts

Script	Description
regiondetection.py	Count objects crossing user-defined regions
segmentit.py	Instance segmentation on video
voetbal.py	Early football detection prototype
berdiboys.py	Player detection on FCBerdi footage
motion-detection-security-camera.py	Motion detection for security use
boxingvideo.py	Pose estimation on boxing footage
inference.py / videoinference.py	General YOLO inference helpers

---

Models

File	Description
yolov8-soccer.pt	Main player/ball/referee detector
yolov8n-football.pt	Lightweight football detector
best.pt	General purpose fine-tuned model
jerseydetection_yolo/runs/.../best.pt	Jersey number region detector

---

Environment Notes

• Python 3.10, PyTorch 2.4.0, Apple M2 Max (MPS)
• Set PYTORCH_ENABLE_MPS_FALLBACK=1 for SAM2 (already set in script)
• Conda env: /Users/pascal-maker/pascalworkspace/env