Abstract
This project investigates the use of machine learning, computer vision, and deep learning to develop an automated football analysis system using uploaded match footage. Starting
with baseline detectors such as earlier YOLO models, the study progresses to a fine-tuned YOLOv8 for detecting players, referees, and footballs, paired with ByteTrack for object tracking. A custom-trained model further improves accuracy in soccer-specific scenarios.
Players are assigned to teams using KMeans clustering based on t-shirt colors, with alternatives like color histogram matching and CNN-based classification considered for
robustness. Optical flow is used to estimate camera motion, and perspective transformation
maps player movement into real-world distances, enabling speed and distance calculations.
The system achieves over 85% precision and recall for player detection and provides visual outputs such as heatmaps and formation maps. The results demonstrate the effectiveness of AI-driven techniques in enhancing tactical analysis and highlight Sac State’s role in advancing data-informed sports evaluation.