請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79037
標題: | 以多層環境可供性地圖達成強健室內定位、人類活動事件偵測以及社交友善導航 Robust Indoor Localization, Human Event Detection and Socially Friendly Navigation Based on Multi-Layer Environmental Affordance Map |
作者: | Ping-Tsang Wu 吳秉蒼 |
指導教授: | 傅立成(Li-Chen Fu) |
關鍵字: | 同時建圖及定位系統,人類活動事件偵測,可供性地圖,社交友善導航, Simultaneous Localization And Mapping,Human Event Detection,Affordance Map,Socially Friendly Navigation, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 近年來,社交型與服務型機器人的研究逐漸嶄露出其重要性,隨著老年人口的增加,陪伴型與服務型機器人有著龐大的應用潛能。然而這些機器人能夠順利完成任務的首要前提必須是機器人擁有非常強健的定位能力、強大的感知能力以及安全導航的能力;另一方面,為因應實際使用的需求,所有的計算必須講求效率。能夠雙方面都兼顧才能使機器人更邁向實際應用的階段。
本研究提出一個新穎的系統架構,其中分別包含了「定位系統」、「感知系統」、「推論系統」以及「導航系統」。在「定位系統」中,本研究提出了一個創新的方法來結合兩種不同性質的「同時建圖及定位系統」演算法,藉由來自不同性質的兩種演算法可以互補各自的不足以達到機器人強健的定位。在感知方面,近年來的研究有許多利用類神經網路機器學習解決物體辨識、場景辨識、動作識別等等之問題;然而同時使用過多的類神經網路卻是不切實際的作法,其一,類神經網路會使用過多的電腦暫存空間,其二,同時且實時的執行多個類神經網路模型在同一台電腦上並非容易的事情,其三,過度的浪費運算資源將會使系統架構的未來擴展性受限制。因此在本研究中,僅物體辨識與人體姿態感測採用類神經網路的方式做運算,其餘如場景辨識以及動作辨識將會基於所提供的物體「可供性能力」使機器人推論出最佳的結果。在「感知系統」中,機器人會利用類神經網路實現物體辨識的能力,並且結合深度影像確認出該物體在空間中相對於「定位系統」地圖中的空間位置。在「推論系統」中,本研究提出了一個結合人類知識機率模型,此機率模型會依據使用者所提供的「可供性能力」以及此環境的資料來做出最佳的推論,其優點是不需要大量的訓練資料以及訓練時間便能使機器人做出足夠準確的推論,並且推論結果會依照使用者所提供的資料而能有不同表現,此方法能夠體現出依環境客制化的優點。在「導航系統」的部份,本研究則加強改進現有開源程式包之系統架構來達到更穩健且社交友善的導航表現。 In recent years, the research of social and service robots has gradually shown its importance. With the increase of the elderly population, companion and service robots have enormous potential applications. The first prerequisite for these robots to successfully complete their tasks must be that the robot has a very strong localization capability, a strong sensing ability, and the ability to navigate safely; on the other hand, in order to meet the needs of practical uses, efficiency of calculations should be emphasized. It is necessary to take both sides into consideration in order to make the robot more practical. This present thesis proposes a novel system architecture that includes a 'localization subsystem', a 'perception subsystem', an 'inference subsystem' and a 'navigation subsystem'. In the 'localization subsystem', we propose an innovative method to fuse two different kinds of Simultaneous Localization And Mapping (SLAM) algorithms, which can achieve more robust localization ability. As for perception, many researches in recent years have used neural network-based machine learning methods to solve the problems such as object recognition, scene recognition, and action recognition; however, it is unrealistic to use too many neural network methods simultaneously. First, neural network methods will cost too much computer temporary storage space. Second, it is not easy to execute multiple neural network tasks simultaneously and especially in real time on the same computer. Third, excessive waste of computing resources will make the system architecture be limited in the future development. Therefore, in this present thesis, only object recognition and human skeleton detection are calculated using neural network approaches. For other information such as scene recognition and action recognition, we will let the robot infer the best solution based on the “Affordance” of the objects provided by the user. In the 'perception subsystem', the robot uses the neural network method to realize object recognition, which is combined with the depth image to confirm the spatial position of the object in the space related to the map in 'localization subsystem'. In the “inference subsystem”, we propose a probability model. This probability model will make the best inference based on the “Affordance” provided by the user and the data of the environment. The advantage is that the robot is able to make good inferences without a large amount of training data and training time, and the inference results will behave differently according to the information provided by the user. This behavior can reflect the advantages of optimizing with respect to different environments. In the 'navigation subsystem', we have strengthened the system architecture of the existing open source package to achieve a more robust navigation and socially friendly navigation. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79037 |
DOI: | 10.6342/NTU201803236 |
全文授權: | 有償授權 |
電子全文公開日期: | 2023-08-21 |
顯示於系所單位: | 電機工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-R05921013-1.pdf 目前未授權公開取用 | 4.42 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。