基於RGB-D感測器之物件姿態估測

Object Posture Estimation Based on RGB-D Sensor

許恆銓(Heng-Chuan Hsu)

指導教授 ： 翁慶昌

淡江大學/工學院/電機工程學系碩士班/碩士(2015年)

https://doi.org/10.6846/TKU.2015.00464

摘要

本論文提出一個基於具有深度影像資訊的RGB-D感測器之物件姿態估測的方法。為了讓機械手臂能夠自主夾取一個在測試平台上的物件，必須依靠機器視覺的處理與分析技術，透過姿態估測演算法來計算位於測試平台上之物件的可用資訊才可以讓機械手臂自主的夾取物件。本論文所使用的RGB-D感測器為微軟公司所生產的Kinect，而欲估測的物件為直徑三公分的灰色直通水管。主要有兩個步驟：(1) 影像處理以及(2) 姿態估測。在影像處理上，本論文提出一個可以定位感測器以及修復深度資訊的方法。當感測器裝置於測試平台上，所提出的方法可以對感測器進行定位來降低感測器與測試平台間的距離誤差，並且用色彩與深度的回授資訊來改善與修復破損的深度影像。在姿態估測上，本論文提出一個物件姿態估測方法，其先合併感測器所獲得的色彩資訊與深度資訊，然後以點雲的方式呈現，所提出的方法可以有效率的估測出物件位於測試平台上的相對關係。最後，本論文將物件放置於測量紙上來測量它的相對角度與相對向量，從實驗結果可以發現所提出的方法確實可以對測試平台上的物件做正確的姿態估測。

關鍵字

物件姿態估測 ； RGB-D感測器 ； 影像處理 ； 點雲

並列摘要

An object posture estimation method based on a RGB-D sensor is proposed in this thesis. The RGB-D sensor could provide a depth information of object. In order to enable a robot manipulator to autonomously grasp an object on a test bench, the processing and analysis technologies of robot vision are necessary. Some posture estimation algorithms can be used to obtain some available information of object on the test bench so that the robot manipulator can autonomously grip this object. In this thesis, the RGB-D sensor is the Kinect produced by Microsoft and the estimated object is a gray water pipe whose diameter is 3 centimeters. There are two main steps: (1) image processing and (2) posture estimation. In the image processing, a method is proposed to locate the sensor and refine the depth information. When the sensor is installed on the table, the proposed method could reduce the distance error between the sensor and the test bench based on the location of the sensor. Moreover, it can repair the damaged depth image based on the feedback of the color and depth information. In the posture estimation, an object posture estimation approach is proposed. It first combines the color and depth information obtained by the sensor, then all the combined results are represented by the way of point clouds. The relation between the object and the test bench could be obtained efficiently by the proposed method. Finally, the object is placed on a measuring paper to measure its relative angle and relative vector. Some experimental results are presented to illustrate that the proposed method can indeed do the right posture estimation of the object on the test bench.

並列關鍵字

Object Posture Estimatio ； RGB-D Sensor ； Image Process ； Point Clouds

參考文獻

[17] K. Kim, J. Kim, S. Kang, and J. Kim, “Vision-Based Bin Picking System for Industrial Robotics Applications,” Proc. of 2012 9th International Conference on Ubiquitous Robots and Ambient Intelligent (URAI), Daejeon, Korea, pp. 515-516, 2012.

[5] 張彥中、林欣儀，機器人容器內取料技術介紹與應用，機械工業雜誌365期，2013年。

[4] R. T. Chin and C. R. Dyer, “Model-based recognition in robot vision,” ACM Computing Surveys, Vol. 18, pp. 67-108, 1986.

[11] A. Berner, J. Li, D. Holz, J. Stückler, S. Behnke, and R. Klein, “Combining Contour and Shape Primitives for Object Detection and Pose Estimation of Prefabricated Parts,” Proc. of IEEE International Conference on Image Processing (ICIP), Melbourne, Australia, pp. 3326-3330, 2013.

[12] D. Holz, M. Nieuwenhuisen, D. Droeschel, J. Stückler, A. Berner, J. Li, R. Klein, and S. Behnke, “Active Recognition and Manipulation for Mobile Robot Bin Picking,” Springer Tracts in Advanced Robotics, Vol. 94, pp 133-153, 2014.