OpenPose-Based Yoga Pose Classification Using Convolutional Neural Network

Abstract

Human action recognition has various implementation, such surveillance system, elders care and construction alert, which arouse lots of interest of research in classification of still image. This paper mainly focusses on detecting the pose of Yoga. Comparing with traditional method using convolutional neural network, which is using original image as input to train the VGG network, extracting the skeleton images and feed them into Mobile net can impressively increase the accuracy. Dataset is collected from Kaggle website which contains five categories of labeled Yoga image. Openpose is an open-source API that can extract the human skeleton structure form the Yoga image based on the pose. With these skeleton image as input, the convolutional neural network will perceive everything important such as pose and angle of joints, rather than irrelevant features such as color and environment. Using Mobile net instead of common method to do classification with VGG, calculation time has been remarkably reduced and size of model is lighter which is able to be apply on single chip device. The result of model is impressive, showing high accuracy in both training data set and testing data set, which means no overfitting problem occurred in the experiment. Model size and demanding of hardware are also acceptable for a common personal computer.