7.3 Automated procedures for absolute orientation developed in this
7.3.4 Practical applications of the map to image matching
7.3.4.1 Isle of Wight case study
The diapositive number 164 of Isle of Wight was scanned using the Sharp JX-600 scanner at 600dpi resulting in a -26Mb image. The flying height inscribed on the photograph is 1750m and the focal length of the camera used is 152.75mm, therefore the scale of the photography is 1:11 600 and the pixel size of the resulting digital image is 0.50m. The photograph was taken in 1985. The 1:10 000 Ordnance Survey map of the same area represented on the photograph was used to digitise the field boundaries. This map was surveyed at 1:2 500 scale between 1969-72 and revised for significant changes in 1973, the publication date is 1975. The vector file obtained was processed as described in section 7.3.1.2. The maximum resolution on a map is 0.2mm, therefore a precision greater than 2m can not be expected from a measurement taken from a 1:10 000 map. This means that even if a resolution smaller than 2m per pixel is used, this will not increase the accuracy of the measurements made on the digital map. Taking this into account and at the same time trying to minimise the amount of data to be processed, both the map and the image were scaled to a pixel size of 3.92m, thus resulting in a -660x660 pixel image. In this case the highest point on the photograph has a height of 100m, even if this point was at the extremity of the photograph where the relief distortion is maximum, according to formula 7.1 this distortion would be 0.6mm, which corresponds to 1.4 pixels on the digital image. Since the image was reduced 8 times, this distortion is irrelevant. The digital photograph was prepared using the procedure described in section 7.3.1.1: it was smoothed, thresholded, the edge detector was applied and finally the smallest polygons were removed from the image. The results of each step of this process are shown in Figure 7.6. The map was translated into the raster format previously described in section 7.3.1.2.
Chapter 7. Map to image matching
scaled MAP scaled IMAGE
Image processing to extract polygons Extraction of the attibutes
that characterise each map ______ polygon_______
Extraction of the attibutes that characterise each
image polygon
Map to Image patch matching based on shape
Extraction of the matched polygons from the original ^ map at a refined scale
Prepare map for
matching Prepare image for matching
Match map to image
Map/Image Conjugate points
MAP
(field boundaries coyerage)
IMAGE (digitised photograph)
Map and Image corresponding polygons Map and image conyersion
to a common reduced scale
Extraction of the matched polygons from the original
image at a refined scale
•<.\^> <• '4'-':/*• ;> V j I f & ÿ
'% & -A
# '
'■^3?;>:.> ,Figure 7.6 The Isle of Wight original image (top left), the smoothed image (top right), the thresholded image (bottom left) and the edges detected in the image (bottom
Chapter 7. Map to image matching
2 0
26
Figure 7.7 shows both the image and the map prepared for the matching process, the polygons are numbered according to the order used to read them. The programs
code_image and code_map were applied and a record of the frequencies of the directions between neighbouring pixels, the frequencies of the change in direction segments, the area of the polygon, the perimeter of its boundary, the width and height of the bounding rectangle and the coordinates of the centre of gravity of each polygon were extracted. Figure 7.8 shows an example of the information recorded for one of the polygons represented in the Isle of Wight image. Appendix C shows the whole set of polygons extracted with the respective descriptions of their attributes. It can be observed that the frequencies of the change in direction segments are always 0 for the values 3, 4, and 5. The frequencies of chain code differences are always 0 for the value 4, which represents a 180 degree turn at a point. The values 3 and 5 barely occur because the polygons are generally 1 pixel width, there are very few joined pixels. The values 2 and 6 occur very rarely showing that the edges are relatively smooth.
Polygon number = 1 perimeter =135 area = 227 Code = 61 3 1 13 39 13 3 1
Differences = 93 23 1 0 0 0 0 17
Width of bounding rectangle = 65 Height of bounding rectangle = 9 Centre of gravity (X) =377.8 Centre of gravity (Y) = 3.6
Figure 7.8 Information extracted from a polygon represented in the Isle of Wight image
The program match_code was written to compare the information relating to each of the polygons extracted from the map and from the image. For each polygon in the map differences were computed between each of its attribute values and the correspondent attribute values for all polygons in the image. After adding all the differences, the minimum value was found and the pair of polygons that correspond to that minimum is considered matched. As mentioned by Abbasi-Dezfouli and Freeman (1994b) the chain code approach followed in this study is very often a reliable way to distinguish between different shapes. However the method by itself is not infallible and the other parameters used to characterise a polygon should provide the information
Chapter 7. Map to image matching
to compensate for a possible uncertainty derived by the frequencies of the chain code direction segments and change in direction, if these frequencies were used alone.
It was observed that different weights should be given to the differences between the parameters that characterise the polygon shape. More precisely, it had to be considered that the difference in area of the polygons to be matched had to be scaled to at least to half of its value since it was adversely affecting the matching results. This problem can be easily understood by considering that the area of the polygon includes all the pixels inside the boundary of the polygon, depending on the size of the polygon a slight difference in shape can contribute to a high difference in the area. If not reduced the difference in area between the polygons would be the main characteristic to be considered when adding all the differences between the different parameters used to characterise the polygons. Taking this into account the matching was performed with no problems. For each polygon in the map one polygon in the image was found whose characteristics were considered to be the most similar. The resulting value from adding all the differences between the parameters of each pair of polygons was then divided by the perimeter of the polygon to compensate for the fact that more differences are expected to occur on polygons with a longer perimeter. It was observed that to eliminate the bad map to image correspondences the matching value for each match did not provide a very robust way of identifying the bad matches. To solve this problem, it was also necessary to consider the relative geometry of the patches to detect the wrongly matched patches. Two approaches were considered to perform the selection of the good matches. The first method consists of computing geometric relations between the map polygons, such as distance and angles, and then checking that the correspondent image polygons maintain similar geometric characteristics among them. The second technique starts by considering that the best matching value should belong to a correct match. Based on this, a transformation is computed between the best matched polygons using the centre of gravity, and the furthest and the closest point from the centre of gravity of both map and image correspondent polygons. The parameters found by this transformation are then applied to transform each map polygon centre of gravity into the correspondent image polygon centre of gravity. The transformed centre of gravity is then compared with the centre of gravity of the image polygon that is considered to match the map polygon under consideration. If these two values coincide then the matching is accepted, if they are different then this means that the original correspondence is wrong and that map and image polygons do not match. Both techniques were implemented, the
programs are called match poll and match pol2 respectively. It was observed that the second approach can cause problems because the map to image transformation initially used focuses on a restricted area of the image where the polygon that starts the process is located, this may be quite different from the transformation to be used on polygons distributed in the different areas of image. Therefore the first procedure to filter the bad matches was adopted since it applies extra geometric constrains that assure a correct result. Table 7.1 shows the matching values, the values highlighted represent the good matches.
Once the common polygons are identified a refinement of the matching and a consequent identification of conjugate points can be performed using the map to image matching technique based on the dynamic programming algorithm described in section 7.3.2. Now that the conjugate polygons have been found using the map to image matching algorithm based on shape, the knowledge of the centre of gravity of these polygons together with the width and height of the bounding rectangle can be used to section the original image into areas only containing the polygon to be further processed. The resolution of the image and the map for this matching is 2m per pixel which corresponds to the minimum resolution to be used with a 1:10 000 map. Using the method outlined in section 7.3.3, the initial estimation of the transformation between the map and the image can be computed.
Table 7.1 Isle of Wight matching values and corresponding matched polygons
M atch in g V alue Map P o ly g o n Im age P o ly g o n 0 .6 2 9 1 6 1.641 2 6 1.476 3 17 1.826 4 13 0 .8 1 0 5 8 1.370 6 20 0.839 7 10 0 .6 6 0 8 10 1.491 9 9 1.644 10 18 1.780 11 4 0.623 12 18 0 .5 5 5 13 13 2.778 14 18 1.008 15 18 1.284 16 18 1.846 17 18 2.498 18 6 1.463 19 18
Chapter 7. Map to image matching
Figure 7.9 shows one of the map polygons with the respective image polygon used by the program map_image_match. These two polygons were independently processed using the techniques described in section 7.3.2.1 and 13.2.2 in order to prepare them to be matched. Figure 7.9 shows the points matched for one of the polygons considered.
The method followed to relate a map to an image performed well, and the results of its application to the Isle of Wight digitised photography were analysed and described in detail. The map to image matching system presented does not require any initial approximation to initiate it, providing the opportunity to achieve a higher level of automation than some of the methods previously attempted. The map to image matching techniques developed and applied in this study showed to be successful in automatically obtaining the required list of map and image corresponding points to perform absolute orientation. This list of coordinates will be used in the next chapters to assess how appropriate they are to carry out automatic absolute orientation.