基於動態屬性轉移學習之人機互動

隨著科技發展，人機互動日漸頻繁，在許多工業環境下，產業也已經引進許多機器人協助人類的工作。然而，在工業環境下機器人的指令較為簡單，倘若是在公眾場合及居家環境中，服務型機器人必須理解與學習人類的社交行為才能融入人類社會並給予協助或服務。因此，如何使機器人有能力與人類共同和諧地生活已變成重要課題，所以如何使它們對於人類在社交環境的動態行為有所觀察及做出預測是發展重點。為了達成人與機器人之間的自然互動，本論文致力於結合轉移學習技術以及動作識別能力來提高機器人對於人類行為的認知能力並且提供服務。然而，我們不只提出三層式社交行為預測系統，使機器人能夠理解人類社交行為的方法，也帶來讓機器人在學習新情境時更有效率的效果。在傳統機器學習方法上，為了學習新情境常常需要重新收集大量資料及重新設定人工標籤，機器人才能有更好的表現模式。因為新方法的關係，我們將不需浪費額外成本在收集新資料和產生標籤上。

關鍵字

轉移學習；社交行為辨識；屬性；輪型機器人

並列摘要

Robots need to understand the human social behaviors to interact with them correctly and to service people in public environments and homes. The way in which robots use to live with humans has become an important issue. Therefore, estimation and prediction of dynamic human behaviors in social environment are very important for robots. In order to reach a natural and harmonious interaction between humans and robots, this thesis attempts to integrate transfer learning technology and action recognition to enhance the cognitive ability of robots when they serve humans. Moreover, the thesis not only presents a method which helps robots to understand humans’ social behavior, but also creates simpler method to help robots learn new scenario which they never learn. In the traditional machine learning methods, in order to learn new situations, developers often need to collect a lot of data and reset artificial labels again and again, to help the robot have better performance. In terms of multi-layer attribute prediction method, we need not make extra efforts on collecting new data and manually labeling and can still help robots to react accurately.