Squat Assessment using Pose Estimation

Abstract

This Thesis offers a new perspective on how to assess squat performance without the help of personal trainers. Media Pipe model allowed me to extract keypoints for various squat positions from pre-recorded videos. The data was then saved into a file with respective categorizations. Each file and a different parameter being valued. Squat data was categorized based on side (front, left, right), width (wide, narrow, neutral), range (low, medium, parallel, high) and stance (up, down). Thresholds were defined so that with the help of the Pose Estimation model keypoints were automatically categorized and saved. The CSV files were then loaded into data frames to be used to train each category’s model. With all the models trained, combining them in real-time feedback demonstration where all the categories were showcased and then saved as well into a csv file so that the set could be analyze. Ultimately, it was decided that this last file was also a good base to try and classify a squat as correct or incorrect depending on the users preference. the study aims to provide a customizable framework for squat classification, accommodating various target ranges and widths, thereby offering flexibility based on different research parameters and individual goals. The resulting system increases the accuracy of squat performance feedback, highlighting the potential for automated, real-time assessment in fitness applications.