Rugby Injury Detection System

Project: AI for Player Safety: A Computer Vision System to Detect Rugby Injuries 🏉

The Challenge
For my dissertation, I tackled a critical issue in amateur rugby: the significant gap in medical support compared to the professional game. Grassroots players are exposed to high-impact collisions but often lack the immediate medical attention, advanced technology (like smart mouthguards), or television match officials (TMOs) that professionals rely on. This disparity puts them at a much higher risk of unrecognised and potentially serious injuries. The challenge was to devise a low-cost, non-invasive solution that could help bridge this safety gap.

My Solution
I researched, designed, and built a proof-of-concept machine learning (ML) system that uses computer vision (CV) to automatically detect potential injuries from a video feed. My approach covered the full project lifecycle, from research to a working prototype.

Primary Research: I began by conducting a quantitative survey of 61 amateur players to gather direct insights into their attitudes towards injury, recovery protocols, and their willingness to adopt new technology.

Creative Data Acquisition: Due to the ethical constraints of filming live injuries, I created a robust and diverse dataset from scratch. This involved a multi-faceted approach, using green screen photography, motion capture, 3D modelling, and generative AI to produce thousands of simulated images of players in various injured or vulnerable positions.

Model Training: I trained a YOLO (You Only Look Once) object detection model on this custom dataset. The model was specifically taught to distinguish between players who are 'uninjured' and those who are 'injured' based on their posture (e.g., lying on the ground, kneeling, holding a limb).

Prototyping: I developed a functional web application that runs the model in real-time using a live camera feed, demonstrating how the system could alert touchline staff to a potential injury the moment it occurs.

The Result
The project successfully demonstrated the viability of using an accessible MLCV system to enhance player safety in amateur rugby. The final trained model achieved an outstanding mean Average Precision (mAP) of 93.4%, with 94.4% precision and 92.4% recall, proving its high degree of accuracy despite being trained on simulated data.

This dissertation provides a strong foundation for a real-world, low-cost tool that could be deployed by grassroots clubs to provide a crucial extra layer of protection for their players, ensuring fewer injuries go unnoticed.

Key Skills Showcased: Machine Learning, Computer Vision, Object Detection (YOLO), Data Synthesis, Academic Research & Analysis, Prototyping, Unity, Python