3D Transparent Object Recognition for Service Robotics

隨著科技的進步，讓生活變得更自動化的需求是擋不住的浪潮。屆時，將有許多的服務型機器人會深入到人類的環境中進行各式各樣的任務，例如在家中幫忙倒牛奶、在餐廳中幫忙端水等等。在我們的生活中，用到了許多透明的物體，包括玻璃杯、寶特瓶、甚至是玻璃門，若機器人沒有能力辨認透明物體，將會造成許多問題，這些問題包含機器人容易毀損玻璃杯、容易撞到玻璃門或窗戶等等，不僅僅會造成機器人工作上的不便利、損壞的玻璃更可能造成人類的危險。因此，在此篇論文中，我們提出了一個透明物體的姿態辨識系統，其中我們將討論的重心放在透明物體的辨識上，輔以討論姿態辨識的模組以及抓取的模組。之所以將重心放在透明物體的辨識上，是因為姿態辨識以及抓取的功能在非透明物體上已經有相當成熟的研究。然而，辨識透明物體的研究是近十幾年來才漸漸發展起來，而且論文數量相當稀少，我們若能發展出有效的透明物體辨識演算法，將場景中透明物體所在的位置標示出來，接下來的姿態辨識和抓取的方法就可以參考適用於非透明物體的技術了。 故關於辨識透明物體，我們討論了三種方法，第一種使用RGBD感測器來感測場景、利用感測器的特性加以辨認出透明物體的所在位置。第二種及第三種方法都使用一般的相機當作感測器，分別使用Latent Dirichlet Analysis以及Deep Learning的機器學習方法來學習辨識透明物體。雖然探討了三種方法，我們主要使用第一種方法辨識到的透明物體輪廓當作姿態辨認模組的輸入。 於是，我們可以使用已經儲存在資料庫裡的透明物體3D模型，配合前述方法所找到的透明物體輪廓，可以利用配準的方式進行姿態的估計，進而得到物體姿態的估計值。With the advancement of technology, the trend to make our lives more convenient by robot technology is unstoppable. In the future, many service robots will enter our living environments to do all kind of tasks from pouring milk for us in our home to serve water in restaurants. In our living environment, there are lots of transparent objects including cups made of glass, PET bottles and glass doors. If a robot who serve in our environment cannot recognize transparent objects, it might easily broke the transparent objects made by glass, it might not be able to open the door made of glass, it might bump into and broke glass windows and cause danger. As a result, we propose algorithms that make a robot be able to recognize and estimate the pose of transparent objects in this thesis. We emphasize on transparent object recognition because pose estimation and manipulation for non-transparent objects are relatively mature, while research on transparent object recognition just starts from a decade ago with a few papers discussing this problem. If we can develop effective algorithm for recognizing transparent object, we can take advantage of pose estimation and grasping for non-transparent object to build a complete system for grasping transparent objects. For recognizing transparent object, we discuss three methods in this thesis. The first method which uses RGBD sensor to detect the transparent object is mainly used because the result is suitable for pose estimation. With the stored 3D model of transparent object and the silhouettes of transparent object, we can estimate the pose by matching the model and the silhouette. Experiments show that our method can be used to detect and estimate the pose of transparent objects