非等時空影像回授之混合視覺伺服應用於機器手臂對等向性目標的追蹤

Abstract

在本篇研究中，我們提出使用混和視覺伺服演算法於追蹤等向性物體的控制架構。等向性物體，舉例來說圓球，可將複雜的模型精簡化，以減輕控制的負擔。然而，過於簡單的模型將導致相機能取得的特徵點數量過少，使得只依靠視覺去追蹤目標有困難。傳統的方法有基於位置，和基於影像的視覺伺服，但兩者皆無法分別地有效處理所遇到的困境，尤其是系統存在一些不確定性時。因此，我們提出了混和視覺伺服的控制器，並試圖解決該問題。 首先機器人會針對物體的半徑大小做不同的處理。接著我們透過來自不同相機(非等空間)，或是不同時間點(非等時間)的影像回授，藉由混和視覺伺服控制器的處理，只要大於三個自由度的機器手臂，便能完成視覺伺服的任務。此外，透過誤差累積補償演算法，系統所需迭代的數量可被減少，進而促使所需時間減少。利用李亞普諾夫系統分析，可確保提出的系統在一定系統增益範圍內，有漸進穩定的性質。 實驗中，我們使用兩種機器人型態進行測試：單相機配置與雙相機配置。視覺伺服的目標為採收不同大小的作物，分別為體積較大的哈密瓜與體積較小的番茄。實驗場所為溫室，針對不同的場景和參數配置進行實驗。 最終測得結果包含，對大體積物體有90%的半徑預測正確率，並且在採收大小物體時分別有80%和68.4%的成功率。結合理論計算與實際驗證，我們找到系統的增益，在數值等於0.2時有最佳解。平均來說，在最佳系統增益下需要三次迭代，便可使誤差收斂，同時耗時約21.2秒。然而透過誤差累積補償，可更進一步到只要一次迭代，即可到平衡點，耗時降低到只要6.26秒。這些成果展現了，我們的系統不僅成功解決問題，在兩種機器人型態和不同目標物上的實驗，均展示該控制演算法能適用於不同的場合。

In this study, we propose a hybrid visual servoing method to track an isotropic object. Isotropic objects, such as spheres, are ideal models that simplify complex systems. However, they pose difficulties in visual tracking due to the insufficient number of features. The current visual servoing techniques, position-based and image-based algorithms, cannot handle the task individually, especially when model uncertainties are included. As a result, we designed a hybrid visual servo control to conquer this problem. The robot handles objects of varying sizes differently. By using image feedback from different cameras (spatial) or from different time steps (temporal), a manipulator with at least 3 degrees of freedom can accomplish the task. Furthermore, cumulative error compensation reduces the number of iterations, decreasing the operating time. Through Lyapunov analysis, the system is proven to be asymptotically stable. During the experiments, two robot configurations are presented: a single-camera setup and a dual-camera setup. The objective is to harvest large and small isotropic objects, specifically melons and tomatoes. The experiments were conducted in a greenhouse and tested under various parameter selections. The results include a 90% accuracy rate in radius estimation for large objects and successful harvesting rates of 80% and 68.4% for large and small objects, respectively. The optimal gain of the system matches both theoretical and experimental values, with an optimal solution at 0.2. The average iteration needed is 3, with a time cost of 21.2 seconds. However, with cumulative error compensation, the time can be reduced to 6.26 seconds with only one iteration. These results showcase the success of the tasks and demonstrate the robustness of the system.