- 學位論文
基於景象的平面限制運用基因演算法實現3D立體攝影機自我校正
Using Genetic Algorithm to Achieve 3D Stereo Camera Self-Calibration Based on Plane Constrain of Scene
摘要
在兩張2D平面影像轉換成3D立體視覺中，攝影機參數校正是一件很重要的基本工作，這個工作通常需要事先手動量得校正用的特殊圖案的3D資訊及其對應影像，利用此3D資訊及影像，可經由離線計算的步驟取得攝影機的校正參數。此步驟較為麻煩，且不易運用在即時系統中，例如目前先進研發的自主行動視覺機器人、自動導航車等。 在本文中，我們提出了一個方法，讓攝影機可以即時自我校正並獲得攝影機校正參數。此研究方法是利用空間的自然景物中，很多物體常存在著平面，基於平面的條件限制，利用基因演算法的演化可得到攝影機的校正參數。這個演算法，先從3D立體攝影機中取得左右兩平面影像，再從兩平面影像上取得數個對應點，利用這些對應點共一平面的特性，計算最接近的平面方程式及這些對應點與平面方程式的最小距離，平均的最小距離就是基因演算法中的適應值，最後搭配基因演算法演化得到攝影機的校正參數。 使用此基因演算法取得攝影機的校正參數，只要隨機產生第一代的染色體基因，基因演算法會演化收斂到最接近的攝影機校正參數。使用這種方式不需要特殊的校正圖形及預先的量測做攝影機參數的校正，因此可以容易的運用在即時的系統中。
並列摘要
It is a very important basic task to apply camera calibration to the conversion of 2D planar image to 3D stereo vision. For this task, the 3D information of the special pattern and the corresponding image for calibration usually has to be measured manually in advance. With the 3D information and image, the camera calibration parameters may be acquired through an offline calculation procedure. Such procedure is more complex and is not easy to be applied in a real time system, such as currently advanced developed autonomous mobile visual robot, Autonomous Guided Vehicle (AVG) etc. In this paper, we propose a method for the camera to perform instant self-calibration and acquire camera calibration parameters. The research method is based on conditional constraints of planes, in which there are many objects, in a natural scene of a space to acquire camera calibration parameters by evolution of the Genetic Algorithm (GA). For this method, two plane images, left and right, are acquired from a 3D stereo camera, several corresponding points are acquired from the two plane images, the co-plane feature of these corresponding points is utilized to calculate the proximal plane equation and the least Root Mean Square (RMS) distance between these points and the plane equation as a fitness value of the GA and the GA is adapted to evolve and acquire camera calibration parameters. To use the GA to acquire camera calibration parameters, only the first generation chromosome genes are generated randomly for the GA to evolve and converge on the proximal camera calibration parameters. This method does not need special calibration patterns and pre-measure for the camera parameter calibration, and can easily be applied in a real-time system.