使用Kinect感測資料進行物體辨識之研究

Case Studies on Objects Identification Using Kinect Sensor

黃正揚(Cheng - Yang Huang)

指導教授 ： 黃榮堂 ； 陳亮嘉

國立臺北科技大學/機電學院/自動化科技研究所/碩士(2013年)

https://doi.org/10.6841/NTUT.2013.00303

摘要

本研究是藉由Microsoft所開發的Kinect來研究。跟傳統攝影機或者CCD不同的地方為，Kinect所接收的資料除了二維影像陣列外，也提供了3維的立體環境，以往除了顏色外，也多了深度的資訊。
影像處理有著名的OpenCV可供使用，三維的立體環境則有Willow Garage開發的PCL( Point Cloud Library )。本篇研究使用PCL來做3D立體環境的複雜處理，包含分析區域法像量、分割平面、分別區塊。
本篇提到的物體辨識並非以建立資料庫或者是機器學習方式來辨識物體。雖然由已知資料庫的物體或者機器學方式來辨識物體有非常良好的辨識能力，但麻煩的是必須提前建立資料庫，資料庫夠豐富，才能有效的辨識物體。本篇兩個辨識使用物體既有的幾何特性來篩選不滿足的資料，並得到結果。接下來為物體表面疊合，主要是加強或者是改善碗盤辨識以及插座辨識。碗盤辨識使用表面疊合可以增加碗盤的形狀使資料更完整。辨識插座使用表面疊合可以增加插座的辨識距離，並使機器人大概知道插座的位置，使辨識時間減少。附錄是機器人避障，機器人接收藍芽訊號，從一點移動到另一位置，除了原本以記錄的固定障礙物，譬如說牆壁桌子，這些可事先借由路徑規畫的方式先預設機器人要走的路徑，不過環境中時常有隨機的障礙物，這時則要藉由Kinect的深度辨識來計算與障礙物的距離，並計算旋轉角度，機器人避障僅是初始結果，只有計算往哪方向轉所需最小角度，並無考慮與環境中的關係。

關鍵字

Kinect ； Point Cloud Library(PCL) ； OpenCV (Open Source computer vision) ； 點雲 ； 影像處理

並列摘要

In the early time, RGB camera and LIDAR are using for robot to view the world. But something makes this situation a little bit different – Kinect. Because of widely use of 3D processing, the PCL jump up and show the possibility for people to easily use in 3D world. Iteractive Closest Point (ICP) is the idea way to identify the target, but it has to reconstruction the object first. Several papers show the possibility of using pattern matching to find the outlet or machine learning, but it will face a typical problem because of the point of view and the light.
This paper shows the possibility of finding the feature of the object. For instance, the feature of the dish is the circle shape on the top, and the brighter pixels surrounding the outlet hole. Due to this feature, the step of experiment show some rule of this feature in order to remove the unwanted data and leave the target.
But detecting object by one view would possibility lack some of data, so this paper combines surface registration to make the detection more perfectly. The surface registration makes the point cloud data of dish more completely, and makes the outlet detecting range more far away.
Before the robot does the task that order by people, the robot will move from one location to another location. On the way of the robot, there would some obstacle that block the robot, so obstacle avoidance is the key point in robot navigation.

並列關鍵字

Kinect ； Point Cloud Library (PCL) ； OpenCV (Open Source computer vision) ； Point Cloud ； image processing

參考文獻

4. Besl, Paul J. and N.D. McKay, "A Method for Registration of 3-D Shapes, " IEEE Trans. on Pattern Analysis and Machine Intelligence, pp. 239–256, 1992.

6. Sean Kean, Jonathan Hall and Phoenix Perry, "Meet the Kinect : An Introduction to Programming Natural User Interfaces, " 2012.

9. Yokoya, N., Levine and M.D. "Range Image Segmentation Based on Differential Geometry: A Hybrid Approach," IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 11, No. 6, pp. 643–649, 1989.

10. Hameiri, E., Shimshoni and I. "Estimating the Principal Curvatures and the Darboux Frame from Real 3D Range Data, " IEEE Transactions on Systems, Man and Cybernetics, Vol. 33, No. 4, pp. 626–637, 2003.

11. Taylor, G., Kleeman and L. "Grasping Unknown Objects with a Humanoid Robot, " Proceedings of the Australian Conference on Robotic and Automation, pp. 191–196, 2002.