利用搭載雙環場攝影機取像系統之自動車作自動建立房間立體空間圖之研究

Abstract

本研究提出了一個利用有視覺自動車導航作自動建立房間立體空間圖的系統，整個過程無須人為指引。在假設空房間擁有彼此垂直的牆面以及像是門和窗戶的扁平物件的條件之下，此系統可用來獲取空房間的平面空間圖。為了取得環境影像，我們設計了一種新型的環場攝影機取像系統，該系統包含兩個同光軸且背對背連接的環場攝影機，可分別取得環境中上半及下半圓球之影像。我們採用空間對應表與取像系統所取得之環場影像去計算空間特徵點的三維深度資料。我們所提出房間立體空間圖的建立過程包含三個階段，分別為(1)跟隨踢腳進行自動車導航，(2)平面空間圖之建立，以及(3)房間立體空間圖之展示。在第一階段，我們透過跟隨房屋中的踢腳進行自動車導航，直接在環場影像進行影像處理以獲取影像中的踢腳點，再透過空間對應表的查詢，計算出踢腳點的空間位置。基於最小平方法，我們進行圖形分類，將這些踢腳點分成兩組互相垂直的集合，並使用這些踢腳邊線來表示其所在的牆壁。在第二階段，我們提出一個總體最佳化的方法，用以建立平面空間圖，此方法調整所有踢腳邊線，使整體的直線鑲合(line fitting)誤差最小。在最後階段，我們以離線作業方式對自動車航行時蒐集的環場影像進行門與窗戶的偵測。我們將這些在門與窗戶偵測所得到的資料進行合併，且結合踢腳邊線作為房屋架構的三維空間資料，最後以圖學形式展示整個房屋架構。實驗結果顯示我們所建立的平面空間圖是相當精確的，其近似平均誤差為3%，而自動車可多次成功的進行航行並建立房屋立體空間圖。這些結果證明了本系統的可行性。

An automatic house-layout construction system via vision-based autonomous vehicle navigation without human involvement is proposed. The system can be used to acquire the floor layout of an empty room, which consists of mutually perpendicular wall lines as well as shapes of flat objects like doors and windows. To achieve acquisition of environment images, a new type of omni-directional camera is designed, which consists of two omni-cameras aligned coaxially and back to back, with the upper and lower cameras taking respectively images of the upper and lower semi-spherical spaces of the environment. The pano-mapping approach is adopted to compute 3-D depth data of space feature points using the upper and lower omni-images taken by the proposed imaging system. The proposed automatic house layout construction process consists of three stages, namely, vehicle navigation by mopboard following, floor layout construction, and 3-D graphic house layout display. In the first stage, vehicle navigation is conducted to follow the wall mopboards in a house. An image processing scheme is applied directly on the omni-image to extract mopboard points. Then, a pattern classification technique is proposed to classify the points in two perpendicular groups, which are then fitted by lines using an LSE criterion. Such mopboard edge lines are used to represent the walls including the mopboards. In the second stage, a global optimization method is proposed to construct a floor layout from all the mopboard edge lines in a sense of minimizing the total line fitting error. And in the last stage, doors and windows are detected in an offline fashion from the omni-images taken in the navigation session. The data of the detected door and window then are merged into the mopboard edge data to get a complete 3-D data set as the house structure, which finally is transformed into a graphic form for 3-D display from any viewpoint. The experimental results show that the constructed floor layout is precise enough with an approximate average error of 3%, and that automatic vehicle navigation may be repeated successfully to construct graphic house layouts. These facts show the feasibility of the proposed approach.