This paper proposes a robust motion extraction technology based on Deep-Learning for robot-human communication.
As AI (Artificial Intelligence) technology develops, it is necessary to understand basic human behaviors such as standing, walking, and running to understand the user's complex situation and complex situation.
The previous deep-learning-based motion recognition technology shows good performance when the user's shape, that is, the structure of the head, torso, arms, legs, etc. is clear. However, when the structure of the human body is not visible due to various causes such as a long skirt, there is a problem of a sudden decrease in the motion extraction performance.
This paper proposes a robust motion estimation model for estimating the human shape to solve the above problem. The proposed algorithm implements the robustness of the user movement model through reinforcement learning after applying a deep-learning-based user estimation model. In particular, a strong joint extraction technology for users was developed through policy-based reinforcement learning. Based on this, it was applied to an educational interface for user learning.
As a result of applying the proposed method, it was possible to accurately estimate human motion under various conditions. Moreover, we can confirm that it can be applied to various educational platforms and applications such as sports motion analysis, home training, dance, etc.

Authors:
Tae-Koo Kang, Sangmyung University, South Korea
Yoo-Seok Bang, Sangmyung University, South Korea
Dong-Min Seo, Sangmyung University, South Korea
Seo-Young Won, TDI, South Korea

About the Presenter(s)
Professor Tae-Koo Kang is a University Assistant Professor/Lecturer at Sangmyung University in South Korea

See this presentation on the full schedule – Schedule

Bookmark