請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74913
標題: | 點對特徵物體姿態估測結合多視角投票方案用於工業機器人隨機取物 Pose Estimation Using Point Pair Features and Combining Multi-View Voting Scheme for Industrial Robotic Random Bin-Picking Applications |
作者: | Yao-Jia Kuo 郭曜嘉 |
指導教授: | 連豊力 |
關鍵字: | 主動視覺,物體辨識,機器視覺,點對特徵,隨機取物,最佳視角預測, Active Vision,Object Recognition,Machine Vision,Point Pair Feature,Random Bin-Picking,Next Best View Prediction, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 本篇論文提出了一種基於視覺的機器人隨機取物,該系統可以對物體進行三維姿態估測。提出的辨識演算法可以整合多視角的資訊進而對估測的姿態進行修正,也提出了一個最佳視角估測的策略進而去減少感測或視角所造成的不確定性。
本篇論文提出的辨識系統是基於點對特徵的方法,此方法使用的哈希的技巧和一種有效率投票的機制達成快速的姿態估測。除此之外,此系統使用了一個驗證函數產生一個分數去檢測估測的姿態。如果估測的姿態是錯誤的,本篇論文提出了一個多視角的方案去修復錯誤的估測。這個方案利用了在點對特徵方法上和賀夫相似的的投票機制,從多視角的資訊將會被整合進而去增加系統的穩定性。 再者,為了減少感測的不確定性和環境的不確定性,本篇論文也提出了一個sensor planning的策略去決定下一個最好的相機姿態去觀測環境。這個決策定義了一個成本函數考慮了下個視角所能得到的視野,同時考量了移動的效率和機器人的工作空間,而且這個決策也有逃離區域資訊的能力。 實驗結果首先展示了當多視角的資訊被整合時,提出的多視角的方案可以得到收斂的結果,而且也可以增進驗證函數產生出的分數。再者,實驗也展示了最佳視角估測可以被用來預測下一個含資訊量最多的視角,同時考量了效率和安全性。最後,辨識系統將會被使用在機器人隨機取物的應用上,藉由一個普遍的六軸機械手臂和配備一個相機。 This thesis presents a practical vision-based robotic bin-picking system that performs three-dimensional pose estimation of the object. The proposed recognition algorithm can integrate the multi-view information to refine the estimated pose, and a sensor planning strategy is also proposed to predict the next best view to reduce the uncertain caused by sensors or viewpoints. The proposed recognition system is based on point pair features (PPF) using a hashing technique and an efficient voting scheme to achieve fast pose estimation. In addition, the system uses a verification function to generate a score to check the estimated pose. If the generated hypothesis is not correct, a multi-view scheme is proposed to fix the wrong estimation. The scheme utilizes the Hough-like voting property in PPF, and the multi-view information can be integrated to increase the robustness of the system. Moreover, a sensor planning strategy is also proposed to decide the next best pose of vision sensors to observe the environment for reducing sensing uncertainty or environmental uncertainty. A cost function is defined for the strategy which considers the expected visibility-gain in the next view in the meanwhile ensuring the efficiency in movement, and the workspace constraint of the robot. The strategy also has the ability to escape from the local information. The experimental results first demonstrate the multi-view scheme can get a convergent result to increase robustness when the information from multi-view is integrated. The scheme can also improve the score generated by the verification function. Moreover, the experiment also shows that the next best view strategy can be used to predict the next best viewpoint which contains the most information and considering the efficiency and safety simultaneously. In last, the recognition system will be used for the robotic bin-picking system by a commonly industrial six-degree-of-freedom mounted on an RGB-D camera. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74913 |
DOI: | 10.6342/NTU201904158 |
全文授權: | 有償授權 |
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-108-1.pdf 目前未授權公開取用 | 5.77 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。