ISBS 2022
40th Conference of the International Society of Biomechanics in Sports
In the context of free body physical exercise, the variety of human movements make traditional mechanical measurement systems ill-suited to assess human motion (Chan & Liu, 2009). Moreover, such approaches are often calibrated on the athlete to produce reliable results. However, vision-based Deep Learning (DL) techniques allow model-based estimation of the body movement to measure anthropometrics quantities (i) without the need to calibrate the model on the athlete, and (ii) without forcing the athlete to wear measurement devices during the exercise (Cao et al., 2019). The method adopted in this work is a DL skeletonization algorithm named MediaPipe (Bazarevsky & Grishchenko, 2020) applied on color images that estimates the position of the athlete’s body joints (Figure 1). These joints could be used to produce a smart mirror of the exercise in real-time. Their position along with their estimation uncertainty could also be used to compute intuitive feedback indicators for the athlete. Therefore, this study presents a comparison of embedded hardware to determine the best performing one and a brief study on the feedback indicator computed for an example exercise biceps curl… (scarica qui sotto l’articolo completo)
Fig. 1
