基於三維影像之坦克車即時姿態估測研究

Abstract

近年來，由於三維取像技術越來越成熟，以及取像工具的成本越來越低，三維影像逐漸成為重要的研究對象，在這幾年也有越來越多的三維辨識法則相繼提出，其中大多數的演算法開發，都沒有考慮到物體的形變，這些演算法如果遇到了物體外形變化，勢必會辨識失敗，所以開始有些演算法在開發時，會考慮到物體發生形變時的情況，而這類型的演算法，都必須對物件進行姿態估測才能夠繼續，而本論文的主要目標就是對坦克車達到即時且準確的姿態估測。 本論文選擇坦克車做為實驗對象，主要是考慮到，假設某一天在戰場上，發現了一台坦克車，比起辨識出該坦克車的型號，不如了解這台坦克是否具有威脅來的更有意義，所以本論文的演算法開發主要是考慮在不知道坦克車型號的前提之下，進行姿態估測，並確認是否對我方具有威脅。 本論文的貢獻有二點：第一，本論文提出針對坦克車的切割法則，可以在沒有任何坦克車的尺寸資訊下，達成切割目標；第二，本論文利用提出以網格方式進行點雲篩選，可以使姿態估測結果較其他論文來的更好。

In recent years, the 3D camera technology has become more developed and cheaper. And the research of 3D images has become more popular. Among these years, more and more 3D recognition rules are proposed. Most of these algorithms did not consider about the articulate objects. Therefore, some articulate recognition rules are noticed. These kinds of algorithms can be used only after the pose estimation. So the main purpose of this paper is to design a real-time pose estimation system. Imaging that, if one day, there is a tank appearing in the battlefield. Compared with to identify the model of the tank, it is more meaningful to understand whether the tank is threatening or not. Therefore, the algorithm of this paper is to estimate the pose of the tank to confirm the threat based on the situation without knowing the tank’s model. There are two contributions in this paper. First, we provide a segmentation rule of tank without the size information; second, we propose a method to filter the point cloud data based on grids, and that results in a better performance in pose estimation than other papers.