Edge Detection of Road Boundaries
The Hough Transform (HT) is used to extract straight lines from an edge segmented image (or edge map). The edge map of a residential area is still very complex, so it is difficult to recognise a road network without knowing the location and number of roads. The advantage of using HT is that no prior knowledge of line position is needed. It works even for imperfect edge maps [Hough 1962]. HT pre-processing is carried out using a Canny edge detector.[Canny 1986] It has been shown by Chao and Trinder [1997] that low resolution images simplified the information content, both physically and geometrically, but fail to represent some details. Therefore, high resolution images were chosen for road detection.
Grey level edges reflect the steepness of the intensity changes. To enhance the linear elements on the extracted edge map, an appropriate threshold was selected by the human operator, to ensure the best contrast and quality of linear features. The input image to the HT is the edge map derived from the edge detection process. The output image presents all the detected linear elements in binary mode. The actual straight lines identified by this method which is based on low levelknowledge are presented in Figure 3 [Chao 1999].
Roof Group Extraction
Rooftop extraction is based on image intensities. House density studies by Forster and Jones [1988] and Ng [1990] have been applied to determine a suitable threshold for the extraction of roof blobs and to assess the number of houses in an area. According to Forster and Jones [1988], rooftops reflect approximately 15% of the visible light, but are not necessarily with high intensity. This means that approximately 15% of pixels in a black and white image of a residential area will be related to buildings. Three rates in close arithmetic sequence5%, 8%, 12% , were used to test the extraction of an appropriate number of roof topbright blobs, .
Output images from these 3 P-tile thresholds based on extraction rates of 5%, 8%, 12% , are then combined by adding the new images together. This shows that the number of roof blobs in the new image is more than is obtained in any of the three independent images. Hence the total number of roof blobs can be increased and the rooftop extraction rate improved. This is because some blobs that appeared in the 5% extraction rate image did not necessarily appear in the other extracted images. Table 1 shows that the accuracy of the rooftop extraction rate can be increased from 27 percent to 42 percent, up to 60 percent, when the extraction percentage is increased. However, this also increases the number of erroneous blobs, as indicated in Table 1. Therefore, shape representation techniques, (e.g. elongatedness and compactness,) are needed for accurate rooftop extraction and recognition. Results for accepted regions are demonstrated in Figure 3. Better accuracy of extraction is expected to be achieved when higher percentages are used.
Building shape representation is a major issue in 3D building reconstruction and visualisation processes. For precise single building reconstruction, a scale larger than 1/8,000 is required. However, a scale of 1/8,000 might provide a better display of the surrounds and associated information, such as the house density, and the housing patterns in a typical residential area.
| House density percentage |
Extracted blobs |
Extracted correct blobs |
Extracted but error |
Accumulated correct blobs (accuracy) |
| 5% |
19 |
19 |
0 |
19 (27%) |
| 8% |
30 |
28 |
2 |
30 (42%) |
| 12% |
47 |
38 |
9 |
42 (60%) |
Table 1 Results of rooftop extraction based on 3 different extraction levels.The total house number of ground truth is counted to 70.
Knowledge Modelling For Feature Extraction
Semantic Information of Major Residential Objects
Knowledge representation and modelling have been important topics in object recognition. For the contextual study of residential areas, some related factors are considered as follows:
1. Co-existence relates possible objects in a specific class. In this study objects such as houses, roads, trees, grasslands , all appear in a typical residential area.
2. Adjacency, derived from the region adjacency graph in Sonka et al. [1993], is expressed by the distance between individual objects, but it can also be explained by geometric relationships such as, neighbour of, next to, part of, contains, above, below, inside, outside, vertical, horizontal, parallel to, etc. as suggested in Guelch et al. [1990].
3. A Relational structure is based on semantic knowledge in terms of is_a, a_kind_of. It therefore, provides a meaning or a physical object for matching a symbol to an object.
