Accuracy Assessment of 2D Pose Estimation with MediaPipe for Physiotherapy Exercises
Welson Lima Simões, Laurinda L. N. dos Reis, Carmen Paz Suárez Araujo, Juliana Netto Maia
Abstract
Currently, it is becoming increasingly important to provide adequate rehabilitation at home and determine strategies to prevent injuries, chronic diseases caused by lack of movement and a sedentary lifestyle. The goal is to help people and improve their quality of life. To enable faster, more practical and cost-effective recovery, monitoring and evaluating the patient's physical rehabilitation at home is crucial to providing feedback to the user. Therefore, this article proposes a system for evaluating the user's posture and performance during physical therapy exercises. The proposed model initially estimates the human pose using MediaPipe in real time. The estimated landmarks serve as input to the K-Nearest Neighbors (KNN) algorithm, which segments exercises into repetitions. Naïve Bayes organizes and classifies movements, considering the correct center point and angular tolerance margins. The developed system achieved an average accuracy of 99.22% for movements performed by the upper limbs.