機器人的人物互動行為認知地圖建構

Abstract

隨著科學及技術的進步，機器人的應用會愈來愈廣泛，且需求會隨之增加。對機器人而言，為了使其更融入人類社會中，必須要理解環境並對某些抽象的規範有所認知，才能夠做出更適當的反應。透過建立環境模型來瞭解人類如何使用空間是認知機器人相當重要的一部分。此外，人會和物體有互動，不同物體含有不同的資訊，倘若機器人能夠理解這些較高階層的行為，人們與其互動會更為自然。本論文致力於人物互動行為構成的環境模型，為了辨識行為，結合彩色影像資訊與身體骨架點資訊進行分類，並提出一個被稱為動態密集連接卷積網路的神經網路架構。另外，應用拉邦動作分析與模糊積分於行為辨識，並使用物體偵測模型和深度資訊判斷有無人物互動行為，最後建立一種描述行為和環境的方法。藉由此環境模型，機器人能理解各個位置適合做哪些行為，並對做出禁止行為的人給予適當的反應。

With the progress of science and technology, the application of robots has become more extensive, and the demand for robots has increased. For robots to become more integrated into society, it is necessary for robots to understand certain abstract norms to be able to interact with people more appropriately. Understanding how humans use space by building an environmental model is one of the principal aspects of cognitive robotics. People interact with objects, and different objects contain different type of information. If robots can understand these high-level behaviors, people can interact with robots more naturally. As a result, this thesis is devoted to constructing an environmental model com-posed of human–object interaction (HOI) information. To identify actions, we combine two action recognition models and propose a novel neural network called Dynamic DenseNet. Also, Laban movement analysis and fuzzy integral are employed for action recognition. Then, object detection models with depth information are used in HOI. Finally, a method for describing behavior and spatial information is established. With the environmental model, the robot can understand what behaviors are appropriate for each location and can respond appropriately to prohibited behaviors.