- 學位論文
運用影片中行走人物之虛擬投影進行影像中消失點及攝影機焦距自動校正之研究
The Vanishing Point and Camera Focal Length Auto-Calibration Technology Using the Virtual Shadows of a Walking People From Video
摘要
影像中消失點是估計攝影機焦距所須之必要參數，而攝影機焦距是還原影像絕對深度(absolute depth information)之重要資訊，故估計出一個準確的消失點，可有助於還原真實世界之絕對深度關係。有鑑於此，本研究提出了一個新穎的影像中消失點(vanishing point)及攝影機焦距(focal length)之自動估測演算法。過往的研究主要著重在找尋影像中的平行線段，或者是以不同時刻同一人物的頭腳位置拉設線段找尋影像中消失點，進而計算攝影機焦距。本研究的方法不同處在於，本研究先估測一個人物在不同時刻的虛擬投影在座標轉換後的長度變化，再以迭代演算之方式來反推影像中消失點所在之最佳位置。 首先，藉由前景偵測，追蹤同一人物在不同時刻的位置，並在追蹤結束後過濾出此人物的直立姿態，針對此組姿態進行虛擬投影。接著再將此組虛擬投影集合的兩個端點以座標轉換的方式轉換至世界影像座標系統(World Image Coordinate System)，並估測出此組虛擬投影集在世界影像座標系統的長度變化，本研究發現同一人物的虛擬投影在世界影像座標系統的長度變化會與消失點的位置有一個特殊關係，即越接近消失點的虛擬投影長度集合之標準差值，會隨著越靠近正確的消失點而逐漸遞減。故本研究針對此一特殊關係，發展一演算法逐步逼近此最佳消失點，本演算法首先會先根據一般場景攝影機之焦距範圍，定義一個特殊之迭代區間(iterative region)，並在此一區間觀察虛擬投影的長度遞增性，結合二分迭代法及循序迭代法快速的逼近一個最佳消失點，進而計算得最佳之攝影機焦距。 本研究的優勢在於只運用了一個攝影機即可得到攝影機之焦距資訊，並且無須額外的校正物體(calibration object)，也不須限制影片中人物的行走路徑，此外，本研究之運算複雜度亦較低，可更為適用於一般需要即時運算的影像監控系統。
並列摘要
The vanishing point in image is an important parameter for estimating the focal length. Moreover, the camera focal length is important information to revert the absolute depth information in image. Hence, if the optimized vanishing point can be confirmed, the absolute depth information in image can be generated easily. In this thesis, a new method to estimate an optimized focal length and vanishing point in image is proposed. In previous approach, the vanishing point estimation focuses on finding the parallel lines in image or the head-feet position of a people in different frames. In proposed approach, the set of virtual shadows of a walking people has been created. Moreover, the length increment of the set of virtual shadows will be computed after the coordinate transformation, an optimized vanishing point can be obtained quickly by using a special iteration method. First, the walking people will be tracked and the leg-crossing posture of this people will be filtered in each frame, by using background subtraction process. Moreover, the projection vector is proposed to create a set of virtual shadows, in order to transfer the head-feet coordinates of these virtual shadows to World Image Coordinate System (WICS). A special relationship between the set of virtual shadows and the vanishing point in image has been found, which is the standard deviation of the set of virtual shadows will decrease, when the transformation point close to the optimized vanishing point. By using this special relationship, a special algorithm can be developed to approach the optimized vanishing point. The algorithm base on a range of focal length in general camera, which can defined a special iterative area. In addition, the increment of length of virtual shadows can be observed in this iterative region, than the corresponding vanishing point can be obtained. To sum up the conclusion, the Bisection iteration method and Sequential iteration method can be combined together to approach an optimized vanishing point. The feature of proposed approach is that the focal length information can be obtained use only one camera; additional calibration target won’t be needed. Moreover, the action trajectory of human is not limited. Besides, the low computational complexity features can be applying proposed approach to real-time surveillance systems.