LEVERAGING POSE ESTIMATION FOR THE ANALYSIS OF HUMAN POSTURE

Abstract

Human posture plays a crucial role in healthcare, sports, and rehabilitation, serving as one of the indicators of overall health. This research aims to develop an automated tool for human posture analysis using pose estimation technology, particularly leveraging the advancements in the You Only Look Once (YOLO) algorithm series, YOLOv8. The methodology involved training a model capable of recognizing additional keypoints beyond those provided by YOLOv8 pre-trained model, focusing on the back view of the human body. The Computer Vision Annotation Tool (CVAT) was used for data annotation, and training was performed using a dataset consisting of images of the human body, the model was able to achieve an accuracy of 82.72% in detecting custom keypoints. The developed web application facilitates user interaction, allowing for image uploads, keypoints adjustments, and posture analysis through the interface built with React JS and FastAPI. The system provides valuable insights into human posture by calculating angles and deviations from a plumb line created in the process.