Imagery based Parametric Classification of Correct and Incorrect Motion for Push-up Counter Using OpenPose

Abstract

This paper presents a real-time approach to count push-ups using 2D video imagery. The proposed method uses OpenPose in each frame to extract multiple joints and links of a human body. Then, it analyzes key motion features linked to counting the push-ups. Taking in consideration the push-up rules of the Republic of Korea Army, five criteria are defined and used parametrically to discriminate both correct and incorrect push-ups. A total of 147,840 samples have been collected from 220 push-up videos each in two different viewpoints: half of the videos for modeling the proposed method and the other half for testing its performance. Finally, the results shows 90.00%, 87.82%, 97.86%, and 92.57% for accuracy, precision, recall, and F-measure, respectively, demonstrating its reliability in military physical tests.