Assessing Bicep Curl Exercises by Human Pose Application: A Preliminary Study

Abstract

The goal of human pose estimation is to locate human body parts and create a human body representation (e.g., a body skeleton) from input data such as images and videos. It has received a lot of attention in the last decade, and it's been used in a variety of applications like human-computer interface, motion analysis, augmented reality, and virtual reality. Even though newly developed deep learning-based algorithms have achieved great performance in human pose estimation, insufficient training data, depth ambiguities, and occlusion remain problems. Another challenge with human pose estimation was implementing evaluations during workouts and physical treatment. Evaluation aids in determining the most appropriate and correct ways to undertake physical workouts. To leverage the human pose estimation neural network to identify human joints and give users instructions on how to exercise properly, this study proposes to analyze bicep curls by measuring elbow flexion angle and identifying key points at the shoulder, elbow, and hand. Thereby, compare with the standard angle to determine if the user has reached the correct amplitude of the exercise or not. This evaluation is essential when the user is doing a bicep curls exercise. For this problem, two different solutions, which are using OpenPose open source and using MediaPipe open source, are presented. After testing on the COCO dataset and our dataset, results show that the MediaPipe method provides better results for bicep curls workout evaluation. In the future, MediaPipe will be used for developing a new application software on a mobile phone to support humans in training.