Estimating the Correctness of Exercises for Flat Feet Using Semantic Segmentation

The aim of the presented project is to promote cooperation, between children with flatfoot and their physiotherapists, and to provide the opportunity for said children, to be able to execute exercises designed to correct flatfoot properly, without the guidance of a trained physiotherapist. In order to achieve this goal, an attempt is made to analyze the correctness of the exercises being carried out. Using machine learning, video footage is processed, in which only the foot and the shin of the subject carrying out the exercises are visible. The resulting task is seldom researched, and the amount of related data is limited, which necessitates the combination of machine learning with optimization algorithms. More precisely, neural networks are first used to segment the frames of the incoming footage, then previously collected data is fitted to the segmented image, using optimization algorithms. Lastly, the exercise being done on any given frame is categorized, by analyzing the relationship between it, and the fitted images, quantifying the correctness of the exercise in the process.