Swimming Stroke Analysis and Feedback System using Machine Learning

Abstract

Swimming is a sport which offers numerous health benefits which however improper and poor techniques can affect the swimmer’s health as well as their performance by leading to significant injuries particularly affecting knees, back and shoulders. The proposed system utilizes machine learning and computer vision techniques to initially analyse freestyle swimming stroke videos and categorizes swimming postures as “good” and “bad” poses and provide feedback to swimmers using a classification machine learning model. The development process includes the segmentation of swimmer images, pose estimation and feature extraction and the training the machine learning model. The model is trained on a large set of swimming footage with annotations. Pose estimation accuracy is increased by the segmentation procedure, which makes sure that only relevant body movements are examined. Once trained, the machine learning model uses a user-friendly web application to recognise poses and provide corrective feedback. The expected outcomes of the proposed system are accurate pose estimation with a mean absolute error of less than 5 pixels and the machine learning model achieving over 90% accuracy in pose classification. Variable video quality and the requirement for model changes throughout time are the challenges of the proposed system as future research will be focused on evaluating for the other strokes and offering more detailed feedback thorough analysis.