Comprehensive Analysis of Deep Learning Approaches for Yoga Pose Detection

Yoga, a leading ancient practice with its roots in mindfulness and physical activity, has had new-found popularity globally in the modern world due to profound benefits in both physical and mental fitness. It is a holistic discipline that combines a diverse set of yoga postures, i.e., yoga asanas, breath control, and meditation. However, maintaining correct body posture during any yoga asana is essential. Convolutional Neural Networks have made it possible to correctly detect and classify yoga poses along with real-time feedback for yoga pose correction. The paper presents a comprehensive analysis of three CNN architectures, namely, MobileNet, VGG19, and EfficientNet, trained and tested on Yoga Pose data set using six optimizers, namely, Adam, AdaGrad, AdaDelta, RMSProp, SGD and SGD with Momentum. The experimental results show that EfficientNet architecture with AdaDelta optimizer outperformed for yoga pose detection with the accuracy as well as with F1 score of 97.45% and 0.98, respectively.