KickVision

This repository features a comprehensive computer vision and machine learning project dedicated to football analysis. It employs various cutting-edge techniques to provide detailed insights into player movements and actions.

Check out this video!🎥

Features

1. Object Detection with YOLOv8

   • Utilizes ultralytics and YOLOv8 to detect objects in images and videos.
   • Fine-tune and train YOLO on a custom dataset for tailored detection.

2. Pixel Segmentation with KMeans

   • Clusters pixels to segment players from the background.
   • Accurate t-shirt color detection using advanced K-means clustering.

3. Motion Tracking with Optical Flow

   • Measures camera movement through optical flow analysis.
   • Tracks player motion seamlessly across frames.

4. Perspective Transformation with OpenCV

   • Uses CV2's perspective transformation to represent depth and perspective of the scene.
   • Enhances visual analysis by adjusting for perspective changes.

5. Player Movement Analysis

   • Measures player speed and distance covered in the image.
   • Provides quantitative data on player performance and movement.

Datasets

• 30s football clips: https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/data?select=clips
• Fine tuning and training of YOLO: https://universe.roboflow.com/roboflow-jvuqo/football-players-detection-3zvbc/dataset/1

Iteration 1

• Utilised YOLOv8 for object detection

Iteration 2

• Fine tuned YOLOv8 to accurately detect players, referees and ball while filtering out noise factors

Iteration 3

• Updated players, referees and ball bounding box for easier tracking

Iteration 4

• Performed KMeans Clustering to separate players into their separate teams

Iteration 5

• Implemented interpolation and back filling using Pandas to address occasional missing data for the ball tracker
• Integrated a tracking mechanism (Red Triangle) to identify the player in possession of the ball

Iteration 6

• Added ball control statistics for respective teams (White = Team 1, Green = Team 2)

Iteration 7

• Estimated camera movement using Optical Flow
• Adjusted players' positions in relation to camera movement

Iteration 8 (Final Iteration)

• Integrated CV2's perspective transformation to accurately represent depth of each frame
• Provided quantitative data (speed and distance) for each player