Handcrafted Feature Assisted Light-Weight Encoder Decoder Based Classifier for Yoga Posture Recognition

The trend for the need for at-home workout regimens and the potential for technology to support the trend is rising as remote employment becomes more and more common. One such activity that has benefits like increased flexibility and stress alleviation is yoga. Presently, numerous self-guided yoga courses in the form of pictures and videos are available online. Consequently, there is an urgent requirement for a system that can recognize and evaluate the precision of yoga posture execution in these instructional materials. In the current scenario lightweight prediction models are very much essential because devices like smartphones, edge devices, and embedded systems that have limited processing power or battery life, lightweight models are the best option as they require less resources to train and operate. One particular approach of making the recognition system lightweight is the use of keypoint-based pose/action classifiers. In this paper, a lightweight classifier for yoga postures based on neural network architecture using MoveNet pose estimation technique is proposed. Our model is able to classify the yoga postures by learning the human joints information in a efficient manner.