Machine Learning and Object Detection in Soccer

Abstract

In recent years, the use of artificial intelligence has been growing in a variety of fields, helping humans perform tasks with greater precision. In soccer, artificial intelligence has been gradually incorporated through systems such as goal line technology and Video Assistant Refereeing. Such technologies reduce the risk of  human errors that may prove to be unfair and/or inefficient due to refereeing tasks causing too much delay. This research explores a possible machine learning model, YOLO, that can process datasets of videos and images from soccer matches more quickly and accurately. Existing image datasets from previous soccer matches were analyzed using a neural network to process and detect where such important objects as the ball or players are. Using the YOLO object recognition algorithm, I coded a program that outputs the precision and recall of the YOLO object detection algorithm using the true positive, true negative, false positive, and false negative counts. Using these values I generated a precision versus recall curve by sweeping through different object detection confidence values. As expected, precision decreased as recall increased, but the AUC is fairly high (around 0.812), comparable in AUC to prior work using VGG16, but with much quicker completion time. Although the default training of YOLO is a good starting point, by diving deeper into training YOLO specifically for soccer images, we could compare and contrast all the possibilities and perhaps find one that proves to be the most efficient and accurate.