Automated Posture Analysis for the Assessment of Sports Exercises

Numerous studies in the medical field correlate the maintenance of human posture in static and dynamic situations with the muscle-skeletal health. One of the most widely used methods for assessing human posture is through visual inspection by professionals. However, this observational assessment process requires the presence of a field expert performing a time-consuming manual analysis. Hence, a reliable automatic posture evaluation system would be of great help for professionals to detect postural misalignments. In the recent years, significant progress has been achieved in pose estimation through state-of-the-art deep learning techniques, competent to estimate human body landmarks fast and accurately from RGB images. In this paper, we describe a methodology scheme to estimate human posture and detect postural misalignments in static and dynamic exercises in real-time. The MediaPipe Pose algorithm is employed to detect human pose and the vector geometry of the pose is evaluated to detect postural misalignments. Furthermore, in order to not limit the applications of this work by preselecting rule parameters for only a certain set of exercises, the rule parameters of any form of exercise are automatically extracted from an example of a single correct execution through machine learning. The datasets of the videos utilized in this work were provided and annotated by a clinical exercise physiologist.