Human Body Posture Detection in Context: The Case of Teaching and Learning Environments

摘要

This paper describes an approach to detect and classify human posture in an individual context, more precisely in a classroom ambience. The posture can be divided into two main groups: "Confident/Not Confident", aiming for the teacher's posture evaluation, and "Interested/Not Interested", targeted for the students. We present some relevant concepts about these postures and how can they be effectively detected using the OpenPose library. The library returns the main key points of a human posture. Next, with TensorFlow, an open-source software library for machine learning, a deep learning algorithm has been developed and trained to classify a given posture. Lastly, the neural network is put to the test, classifying the human posture from a video input, labeling each frame. The experimental results presented in this paper confirm the effectiveness of the proposed approach.