用立體機器視覺擔任行動機器人的運動導向時,須由不同視角的影像求取視差值來獲得立體空間的資訊。我們選擇了利用角度差值的方法去算視差值。這方法不僅快速而且能夠包含每一點的資料。但是這個方法卻只能在其可偵測的範圍內才能成功的求取視差值。也因為如此,一種多層級的方法被使用在求取視差值。在這個層級的方法中,我們架構了一個影像金字塔並且隨著影像的解像度越來越小時,視差值也會隨著縮小。如此一來,所要求取的視差值也許會落在濾波器的工作範圍內,這樣就可以成功的被計算出來了。前人的研究是不管解像度大小就把所有的濾波器結合起來去求取視差值。但是相同濾波器的長度在小解像度的影像資料上會涵蓋比大解像度更長的範圍。也有可能錯誤的資訊會被包含在裡面。在本研究中則發展出一套去偵測吻合度的層級方法,我們只選取那些具有高吻合度的資料來求取視差值。這套方法的實驗結果證實了它是相當精確的。
Using machine stereovision for the guidance of mobile robot needs to extract the 3D information from images captured at different viewpoints. We choose a phase based approach to estimate the disparity values. It is fast and dense. But the disparity value only can be detected within a working range. Thus a hierarchical method is applied. An image pyramid is constructed in the hierarchical method and with the reduction of the resolution will let the disparity value smaller than original images. Thus the disparity value will within the working range of the filters in each pyramid level. But the length of the filters will cover more range in lower resolution and the missing data may be covered. Conventional approach is combined all the filters to estimate disparity value. This research develops a hierarchical coherence algorithm. It only combines the units which are coherent to estimate disparity. This approach is shown to be accurate.