An automatic tool for yoga pose grading using skeleton representation
Thanh Nam Nguyen, Thanh-Hai Tran, Hai An Vu · 2022
Automatic grading of yoga poses may help yoga self-practitioners to rectify their poses to follow correctly the teacher. Some existing wearable or depth sensor based methods require the user to equip additional devices. This paper presents a framework for automatically grading yoga poses from images taken from a conventional RGB camera/phone camera. Our framework consists of three main phases. First, we estimate human joints from RGB images using BlazePose model. Second, we investigate various deep models and select VGG-16 as a model for yoga pose recognition. Finally, we define a score that takes the difference of angles at important joints and yoga pose label into account. Our solution is low-cost, easy, and light to implement on mobile devices. It gives a confident score on the YogaPose dataset and our self-collected dataset.