都會區建築物偵測

Abstract

建築物結構均不免有水平與垂直之構形，不論在遠觀都會區建築群或個別建築進景影像中均可看到其存在。 本論文以前述之觀察為基礎，試作一自動化之都會區建築偵測系統，嘗試切割出都會影像中可能有建築物的區塊。系統運作略分為五個階段：

(1) 以特定之水平與垂直特徵偵測方法，初步定出水平與垂直分佈圖各兩個；同一特徵的兩個分佈圖具有交錯分佈特性。

(2) 透過特定之梯度方位投票機制，改善(1)階段的四個分佈圖結果。

(3) 透過特定之交錯分佈整合機制，分別整合水平與垂直兩個分佈圖經(2)階段強化過的兩組分佈圖。

(4)就水平分佈圖作區塊面積排序並篩選刪除小區塊；垂直分佈圖亦作類似處理。

(5)透過特定之幾何篩選條件，進一步整頓(4)階段之結果；最後透過特定之條件整合水平與垂直分佈圖，完成建築物區塊之標定。

經由測試觀察，都會建築群中部分之建築可由本系統標定出，驗證水平與垂直特徵確為一可用之偵測指標。

Horizontal and vertical beams are essential components to modern building structures, as can be seen in many images of vast metropolitan scenes where groups of modern architectures exist. Based on such a straight observation, a pioneering system was developed for segmenting images of vast metropolitan scenes and retrieving regions where building structures with horizontal and vertical beams may potentially exist. The process of the automated system comprises of 5 stages.

1.Two horizontal-duo maps hz-duo hereafter are generated via the use of two specifically designed horizon-bean extractors, where the pixels distribution in the hz-duo show complimentary nature in regions of interests. Likewise, two vertical-duo maps (vt-duo hereafter ) are developed similarity by two vertical-bean extractors.

2.The fair maps from (1) are further refined by a specifically conceived voting process utilizing gradient orientation in the regions under concern.

3.The hz-duo are fused as on map, by a fusing mechanism devised for the system, where regions of horizontal feature are initially located. The vt-duo are fused as one vt-fuse map in much the same way.

4.Ranking among regions in hz-fuse map is conducted according to area sizes, from which those with top rankings are kept as candidates. The same process is applied to vt-fuse map too.

5.The remaining regions in hz-fuse and vt-fuse maps are put through some geometrical screening test such that only qualified regions survive. Surviving regions then go through a marking process by which major horizontally and vertically featured regions in the metropolitan scene are identified.

Initial experiment on a set of images shows that the use of horizontal or vertical beans as the features for detecting buildings in vast metropolitan view is plausible.