利用可視標誌與稠密光流法之姿態估測及定位

Pose Estimation and Localization of Using Visual Markers and Their Dense Optical Flow

虞登翔(Teng-Hsiang Yu)

指導教授 ： 連豊力

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2015年)

https://doi.org/10.6342/NTU.2015.00433

摘要

相機定位應用在空中無人載具是近期一項熱門且應用廣泛的議題，諸如室內環境探勘、搜救以及空中抓取。由於相機僅提供色彩資訊，一般可以用兩個步驟來取得空間資訊以定位相機。首先在空間中放置已知相對距離的標誌來估算座標之間的轉換。另外可以藉由不同時間的影像來描述相機移動的過程，可以將其分為三個階段來討論，分別是關聯性估測、運動量估算以及定位最佳化。關聯性估測是利用搜尋環境中具有相關性的特徵資料來匹配影像，且使用匹配的結果來估算相機的運動達到定位的目的，並在最後使用多個估算的位置計算出最佳的相對位置。
首先在本文中實現的第一個步驟是設計特殊顏色的標誌以利影像上的特徵辨別，並將其設置在空間中已知的相對距離，從影像與空間之間的對應關係來估算出相機與空間之間的座標轉換關係，另外也針對相機定位及校正的結果做討論，並證實將鏡片的主點假設在影像平面的中心可以得到較佳的定位結果，再者，運用敏感度分析可以得知特徵點的像素誤差對定位結果所產生的影響。
另外一個提出的步驟是利用影像之間的稠密光流來描述不同時刻的相機位置。在本文中利用光流法具有不同方向的特性來分群，並設計一個邊界濾波器來取得可靠的光流群組，由於靜止物體的在影像中位移方向與相機移動方向相反，此時可以觀察到錯誤匹配的光流群組會先出現在影像上相機移動的方向，因此運用這個現象來估算相機垂直移動的變化量，並從理想的實驗場景中得知有使用邊界濾波器的估算結果較為準確。在空中無人載具的運動模型中，翻轉會造成光流圖有旋轉的軌跡，因此我們提出使用霍氏圓型轉換方法來獲得圓型子集合，並用每個子集合的資料特性設計線性及指數權重函數，經由理想實驗得知指數型權重函數可以獲得較正確的相機旋轉量。為了將這兩個步驟運用在真實飛行場景，我們另外對於光流圖做有效性的分析來確定取得的影像是否可靠，從分析的結果可以有效的判別出哪些光流向量是來自缺乏特徵的區域。最後從真實飛行實驗中得知錯誤匹配的光流出現位置不如預期，以及飛機傾斜所造成的水平移動並不在光流的旋轉軌跡假設當中，導致高度及旋轉角度變化估測不準確。

關鍵字

相機定位 ； 相機校正 ； 敏感度分析 ； 稠密光流法 ； 有效性分析 ； 動態估測

並列摘要

Using camera to localize the Unmanned Aerial Vehicle (UAV) is a key technology that has been widely researched in recent decade and has many applications such as indoor exploration, search and rescue, and aerial gripping. Due to the camera only provides color information, the spatial information can be obtained by two steps to localize the camera. The first step is to put the markers in real world with known distances to estimate the transformation between coordinates. Besides, the camera motion can be executed by using the images in different time steps. It can be divided into three stages: correspondence evaluation, motion estimation, local optimization. First, correspondence evaluation is used to find some distinct features with their special characteristics to match images. Second, the corresponding pairs are used to estimate the camera motion. Third, the optimization method is used to correct the localization result.
In this these, the first localization step is implemented by designing the markers with particular colors for the detection method in the image plane, and setting them with known distance in the real world. The transformation between image and spatial coordinates is used to evaluate the relationship between the camera and the spatial coordinates. Otherwise, by discussing the localization and calibration results, the setting of principle point at the center of the image plane can make the localization result is more accurate. Furthermore, by the sensitivity analysis, the localization result is affected by the detection error of features.
The other step in this thesis is using the dense optical flow to describe the correspondence between two images in different time steps. To reduce the information, the flow vectors are segmented into non-overlapping blocks. Each block can be classified into the group according to the mean angle of flow vectors. Moreover, an edge filter is proposed to obtain the reliable blocks. Due to the displacement of the static scene in the image is reverse to the motion direction of the camera, then the outliers of optical flow reveal in the image plane about the direction of the camera motion. The distribution of each group is used to estimate the vertical motion of the camera. In the ideal experiment, the motion estimation result with the edge filter is more accurate. In addition, the roll rotation of UAV makes the spiral pattern of optical flow. Then, the Hough circle transformation is used to obtain the circular subsets. The linear and exponential weighting functions are designed by the standard deviations of the subsets. The estimation result of roll by using the exponential weighting function is more accurate than the result by using the linear weighting function. In order to implement these two steps in real-flight scenario, the validation analysis is used to evaluate the flow map is reliable or not. The analysis result shows that the method can effectively judge the textureless region in the image plane. However, the estimation results in real-flight experiment show that the outliers of optical flow do not reveal as expectation and the pattern of optical flow does not involve the effect of horizontal movement of the quadrotor.

並列關鍵字

Camera localization ； camera calibration ； sensitivity analysis ； dense optical flow ； validation analysis ； motion estimation

參考文獻

[1: Sahin & Guvenc 2007]

H. Sahin and L. Guvenc, “Household robotics: autonomous devices for vacuuming and lawn mowing [Applications of control],” IEEE Control System, Vol. 27, No. 2, pp. 20-96, 2007.

[2: Merino et al. 2011]

L. Merino, F. Caballero, J. R. Martinez-de-Dios, I. Maza, and A. Ollero, “An unmanned aircraft system for automatic forest fire monitoring and measurement,” Journal of Intelligence and Robotic Systems, Vol. 65, pp. 533-548, 2011.

[3: Lindermuth et al. 2011]