雖然樓梯具有高度與深度未知的特性，但這是指其作為一般人使用的情況下。為得知其高度與深度等內容，在此採用了對其影像做處理，此內容流程包括了將彩色圖像轉灰階圖像、 Canny邊緣檢測、中值濾波器去除高頻雜訊、 利用Hough變換來獲取直線、選取感興趣的區域與擷取特徵點等。將獲取的這些特徵點輸入至不同的多層感知器類神經網路（MLNNs），依此結果來做估測一個樓梯的高度、深度與階數。這些多層感知器類神經網路皆為兩個輸入（即二維影像平面座標）和兩個輸出（即二維的大地座標）。因為不需要使用到3D的建模，故稱之為2.5D的類神經建模。 首先在實驗場地搜尋樓梯，將搜尋到的樓梯進行相關估測，根據估測的結果將人形機器人導引至樓梯前方附近，再根據類神經所估測之樓梯的高度及加上內插計算所得之深度，並估測其階數，依此執行人形機器人爬樓梯的動作以完成所設定的任務。最後，以相關的實驗來驗證所提之方法之有效性及可行性。 Although the stair possesses unknown height and depth, they are fixed as a general stair human used. The proposed image processing includes the transform to a grayscale, Canny edge detection, median filtering to remove high frequency noise, Hough transform for a straight line, the selection of regions of interesting, and the extraction of feature points. These feature points are fed into different learned multilayer neural networks (MLNNs) for the estimation of height and depth of a stair. These MLNNs are all two inputs (i.e., 2D image plane coordinate) and two outputs (i.e., different 2D world coordinates). No 3D modeling is required; it is the so-called 2.5D based neural network modeling. First, a humanoid robot (HR) scans the field to find the stair, which is randomly distributed before an HR. Based on the estimated posture of a stair with respect to an HR, it will be navigated to the vicinity of a planned posture. Then the stair climbing using the nominal height and depth with the interpolation of their lower and upper values is designed to finish the assigned task. Finally, a sequence of experiments are arranged to confirm the effectiveness and efficiency of the propose methodology.
