應用類神經網路定位為基礎的主動視覺於人形機器人搜尋、罰踢之研究

“Search, Track and Kick to Virtual Target Point” of Humanoid Robots by a Neural-Network-Based Active Embedded Vision System

周尹喆(Yin-Che Chou)

指導教授 ： 黃志良

淡江大學/工學院/電機工程學系碩士班/碩士(2012年)

https://doi.org/10.6846/TKU.2012.00249

摘要

本研究實驗平台為一有 23 個自由度的小型人形機器人，其高度為 65 公分，重量 4 公斤，其核心系統為嵌入式單板電腦 RB-100，透過我們所設計的人機介面及整合電路控制相關馬達，使得機器人完成所指定的動作(例如，直走、轉彎、踢球)。所應用的嵌入式視覺系統為德州儀器的數位訊號處理器TMS320C6713和VM480CCD視覺模組，及其軟體系統(Code Composer Studio)，藉此以實現類神經網路為基礎的主動搜尋及定位目標物體，並導引人形機器人執行罰踢。本研究包括如下四個部份：人形機器人的機構設計與步態規劃、視覺影像的處理、類神經網路為基礎的主動搜尋及定位目標球、導引人形機器人的策略。
由於人形機器人的罰踢任務，最重要的因素之一即是目標球的定位準確性，因此本論文乃以類神經網路進行其定位，首先將影像視覺投影所形成的影像平面座標轉換到世界座標，進而導引人形機器人執行相關動作。即當視覺系統搜尋到目標球後，由視覺模組將擷取的影像輸入至TMS320C6713以進行包括二值化、中值濾波器去除雜訊、影像修正、計算目標球中心點位置的影像處理。以所訓練的類神經網路，即時並正確地計算目標球與人形機器人的方位與距離，進而導引人形機器人走向目標球所設定的位置。當機器人到達目標球前方約10公分後，視覺系統將會再搜尋球門，若存在球門則定位其虛擬中心位置，以進行機器人踢球的位置修正，當機器人踢球位置修正完成後，則執行踢球的動作；若球門不存在則往前踢球、接著追球、再搜尋球門，直至球門被搜尋到以完成罰踢之任務。最後以相關實驗證明所建議方法之有效性及可行性。

關鍵字

人形機器人 ； 罰踢 ； 搜尋目標 ； 影像定位 ； 類神經網路建模 ； 視覺導引 ； 姿態調整

並列摘要

In this thesis, the Texas Instruments TMS320C6713 digital signal processor, VM480CCD vision module, and the related software (Code Composer Studio), are employed to obtain the “Search, Track and Kick to Virtual Target Point” of humanoid robots by a neural-network-based active embedded vision system. A human machine interface is also designed by an on board computer (i.e., RB-100) to implement a variety of different actions, e.g., walking, turning and kicking, required by the task of penalty kick. The research is accomplished by the integration of the following four parts: the humanoid robot mechanism design and gait planning, the visual image processing for penalty kick, the neural network modeling for searching and positioning, and the strategy of visual navigation of a humanoid robot to execute penalty kick.
One of the most important factors for penalty kick is to accurately locate the ball. In this situation, the relation between image plane coordinate and earth coordinate of a ball is modeled by neural network with suitable learning law. First, the CCD module captures the visual image and then sends it to the TMS320C6713 for the corresponding image processing, including binary segmentation, median filter to remove noise, image correction, calculation of ball center. These coordinates of ball and its ground truths are then applied to construct a neural network model between image plane coordinate and earth coordinate. When the humanoid robot is navigated in the vicinity of the ball (e.g., about 10 cm) by the trained neural network model, the visual system starts searching the goal. If it is found, then the posture revision of HR is made for the execution of penalty kick. If it is not found, the ball is kicked, then tracked, and repeated until the goal is found. Finally, the corresponding experiments are arranged to confirm the effectiveness and efficiency of the proposed method.

並列關鍵字

Humanoid robot ； Penalty kick ； Searching Target ； Image processing for localization ； Modeling using multilayer neural network ； Strategy for visual navigation ； Posture revision

參考文獻

[24]陸念聞，“應用類神經網路定位為基礎的主動視覺於人形機器人罰踢之研究”，淡江大學電機工程學系碩士班，民國九十九年六月。

[25]黃俊豪，“人形機器人的人機介面設計與運動控制之研究”，淡江大學電機系機器人碩士班，民國九十九年六月。

[1]K. Loffler, M. Gienger, F. Pfeiffer and H. Ulbrich, “Sensors and control concept of a biped robot,” IEEE Trans. Ind. Electron., vol. 51, no. 5, pp.972-980, Oct. 2004.

[2]Q. Huang and Y. Nakamura, “Sensory reflex control for humanoid walking,” IEEE Trans. Robotics, vol. 21, no. 5, pp. 977-984, Oct. 2005.

[3]Y. Guan, E. S. Neo, K. Yokoi and K. Tanie, “Stepping over obstacles with humanoid robots,” IEEE Trans. Robotics, vol. 22, no. 5, pp. 958-973, Oct. 2006.