Title:	CLUSTERING AND ANALYSIS OF USER MOTIONS TO ENHANCE HUMAN LEARNING: A FIRST STUDY CASE WITH THE BOTTLE FLIP CHALLENGE
Author(s):	Quentin Couland, Ludovic Hamon and Sébastien George
ISBN:	978-989-8533-81-4
Editors:	Demetrios G. Sampson, Dirk Ifenthaler and Pedro Isaías
Year:	2018
Edition:	Single
Keywords:	Human Motion, Human Learning, Machine Learning, Clustering
Type:	Full Paper
First Page:	145
Last Page:	152
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	More and more domains such as industry, sport, medicine, Human Computer Interaction (HCI) and education analyze user motions to observe human behavior, follow and predict its action, intention and emotion, to interact with computer systems and enhance user experience in Virtual (VR) and Augmented Reality (AR). In the context of human learning of movements, existing software applications and methods rarely use 3D captured motions for pedagogical feedback. This comes from several issues related to the highly complex and dimensional nature of these data, and by the need to correlate this information with the observation needs of the teacher. Such issues could be solved by the use of machine learning techniques, which could provide efficient and complementary feedback in addition to the expert advice, from motion data. The context of the presented work is the improvement of the human learning process of a motion, based on clustering techniques. The main goal is to give advice according to the analysis of clusters representing user profiles during a learning situation. To achieve this purpose, a first step is to work on the separation of the motions into different categories according to a set of well-chosen features. In this way, allowing a better and more accurate analysis of the motion characteristics is expected. An experimentation was conducted with the Bottle Flip Challenge. Human motions were first captured and filtered, in order to compensate for hardware related errors. Descriptors related to speed and acceleration are then computed, and used in two different automatic approaches. The first one tries to separate the motions, using the computed descriptors, and the second one, compares the obtained separation with the ground truth. The results show that, while the obtained partitioning is not relevant to the degree of success of the task, the data are separable using the descriptors.
	Go Back