Automated Correction and Updating of Road Databases from High-Resolution Imagery

Transcription

1 Automated Correction and Updating of Road Databases from High-Resolution Imagery M.-F. Auclair Fortier 1, D. Ziou 1, C. Armenakis 2 and S. Wang 1 1 Département de mathématiques 2 Center for et informatique Topographic Information Université de Sherbrooke Geomatics Canada 2500 Blv Université, Sherbrooke 615 Booth Str, Ottawa Québec, Canada, J1K 2R1 Ontario, Canada, K1A 0E9 - Technical Report No Abstract Our work addresses the correction and update of road map data from georeferenced aerial images. This task requires the solution of two underlying problems: a) the weak positional accuracy of the existing road location, and b) the detection of new roads. To correct the position of the existing road network location from the imagery, we use an active contour ("snakes") optimization approach, with a line enhancement function. The initialization of the snakes is based on a priori knowledge derived from the existing vector road data coming from the National Topographic Database of Geomatics Canada, and from line junctions computed from the image by a new detector developed for this application. To generate hypothesis for new roads, a road following algorithm is applied, starting from the line intersections, which are already in the existing road network. Experimental results are presented to validate the approach and to demonstrate the interest of using line junctions in this kind of applications.

2 1 Introduction Road detection is a time consuming operation when performed manually. Fortunately, aerial imagery tends to facilitate this process, but complete automatization is still far off with the methods used to date. Like edge detection [16, 1] road detection is a type of feature extraction which is almost always context-dependent. In our case, we use high-resolution georeferenced images, that is, the geographic position of each pixel is known, and the existing road topographic network from the National Topographic Database of Geomatics Canada. There are two problems with this database : 1) it is not accurate enough (Figure 1) and 2) new roads are not included. The aim of our work is to automatically correct and update this road network. Figure 1: Example of imprecision in the road database. White lines represent road location from database. Much works has been done on automating road detection [2, 3, 6, 8, 9, 10, 11, 14]. Among the methods developed, we can mention particularly the work of Klang [13]. He developed a method for detecting changes between an existing road database and a satellite image. First he used the road database to initialize an optimization process, using a snake approach to correct road location. Then, he ran a line-following process using a statistical approach to detect new roads, starting from the existing network. We have noticed that none of the reviewed methods make use of junctions in the image. These junctions are generally reliable information and since roads form networks, they are very relevant in this context. There are significant differences between our approach and Klang s. First, we add line junctions to his correction scheme to improve the matching between the road database and the image. Second, we generate hypothesis for new roads by following lines starting from line junctions near the known road network. We describe our approach in Section 2. In Section 3, we present line and line-junction detection. In Section 5, we describe the correction of road location. In Section 6, we present a hypothesis generation scheme for new roads. In Section 7, we present results obtained from a georeferenced image and a road database, both supplied by Geomatics Canada. Finally, the conclusions are discussed in Section 8. 2

3 2 Approach Geometric correction of the database is the first step in updating road maps. Because the road database has been obtained from various sources including scanning road maps, the location of roads is not precise. However, this information allows a good distinction between image features which are roads and those which are not. In aerial images, roads are generally long, thin features. Thus we want to find lines which are near the initialization location provided by the road database. This prompts us to use active contour models (snakes), which are optimization curves requiring an initialization close to the solution. To bring this initialization closer to the solution, we match each intersection in the road database with at most one junction found in the image. Thus, we re-localize road segments according to the difference between database intersections and image junctions. After re-localization, we apply a line optimization process, using the snake approach. The second step in updating road maps is the addition of new roads. We know that roads are built in networks and that new roads are linked with old ones. We thus generate hypothesis for new roads by following lines from line junctions near the known network. The algorithm can be summarized by these following six steps: 1. Reduction of the image resolution, when it is not low enough to have roads as lines, by resampling the image. 2. Extraction of lines using Steger s algorithm [15], which is presented in Section Using the results of the line detection step, we find location of line junctions (Section 4). 4. Matching of extremities of road segments with line junctions and displacement of road segments according to the difference between their extremities and image junctions (Section 5.1). 5. Re-localization of each road segment individually, using the snake approach. 6. Line following starting from junctions placed near the existing road network to generate hypothesis for new roads. 3 Detection of Lines from the Image In Aerial and Satellite images, roads are long, thin elements. Usually, they are bright lines against a dark background. To detect these, we use Steger s line detector which can be summarized as follows [15] : A bright line is a maximum in the gray-level image in a particular direction. Many line detection algorithms have been proposed [16]. The basic idea behind the one we are using is that the second directional derivative of a bright line profile on a dark background has a negative minimum and its first directional derivative is 0 (Figure 2). In 2D, the direction that yields maximum curvature of the function has to be found. Let f (x; y) be a discrete image. This direction can be determined by computing eigenvalues and eigenvectors of the Hessian matrix : H(x; y) =» fxx f xy ; f xy f yy 3

4 t t t DMI, Université de Sherbrooke Technical Report N. 241 (a) Line profile (b) First derivative (c) Second derivative Figure 2: where f xx, f xy and f yy are the second order derivatives of f (x; y). The direction giving maximum curvature is the direction perpendicular to the line. This is defined as the direction of the eigenvector corresponding to the eigenvalue with maximum absolute value. If this eigenvalue is negative, we have a bright line on a dark background and if it is positive, we have a dark line on a light background. The eigenvalues of H(x; y) can easily be calculated by : (f xx + f yy ) ± = The normalized direction perpendicular to the line is ~n =(n x ;n y )= 8 >< >: q ( max f yy ;f xy ) q ( max f yy ) 2 + fxy 2 (f xx f yy ) 2 +4fxy 2 : 2 (f xy ; max f xx ) qf 2 xy +( max f xx ) 2 elsewhere if f xy = max f xx =0 (1) where max is the eigenvalue with maximum absolute value at point (x; y). Using this eigenvalue, we can define a measure of line plausibility. That is, when we search for dark lines, the line plausibility is max and when we search for bright lines, the line plausibility is max. Working now in 1D, to find line location along the profile direction, we prefer to use a polynomial approach than just a non-maxima suppression. Such approach allows to find sub-pixel position of the line. Thus we find the positions where the first derivative of the Taylor polynomial in direction ~n is 0 and its second derivative has a negative minimum. Let ~p @ 2 2! ~n = f x n x + f y n y f xx n 2 x +2f xyn x n y + f yy n 2 y where f x and f y are the first order partial derivatives of f (x; y). A pixel (x; y) is a line point if ~p 2 [ 1=2; 1=2] [ 1=2; 1=2] and max is used to select salient bright lines. When the line s width is greater than one pixel, we can use a suppression of non-minima (we are looking for bright lines) on max in the direction ~n. For the computation of derivatives, we use convolutions with derivatives of a Gaussian of parameter ff. Figure 3 presents an image, its line plausibility and the lines obtained at ff =1. ~n; 4

5 (a) (b) (c) Figure 3: a) Original image. b) Bright-line plausibility ( max )(ff = 1). c) Detected lines. 4 Detection of Line Junctions In the context of road detection in aerial images, we define junctions as line intersections. Image junctions are very useful features, particularly in solving matching problems, because they are generally reliable, stable elements. Unfortunately, we have noticed that they are not used for road detection. One reason is that, to the best of our knowledge, before the beginning of this project, no line-junction detector was existing. We use a line-junction detector developed in our Computer Vision Laboratory, based on a measure of curvature between the direction vectors of the line pixels within a given neighborhood [5]. This detector also finds line terminations. This measure is estimated by computing the mean of dissimilarity between the orientation vector of the line pixels : MX MX 1 fi( x; y)[1 j~n(x; y) ~n(x + x; y + y)j] N x= M y= M where ~n(x; y) is the normalized line direction at point (x; y) from Equation 1, is the dot product, fi( x; y) =ρ e ( x 2 + y 2 ) if a line exists between (x; y) and (x + x; y + y) 0 otherwise and N the number of points (x + x; y + y) such that fi 6= 0. The junction is localized in a high curvature region using a non-maxima suppression. For more details, see the report on this detector [5]. Figure 4 presents the results of the line-junction detector. 5 Correction of Road Location As we will see in more details in section 5.2, the correction of road location involves two steps. The first is re-localization of the known roads to bring them closer to the final snake s solution. The second step concerns the snake process itself. These two steps are presented in the next two sections. 5

6 Figure 4: Line junctions. 5.1 Matching the Road Database with Image The active contour model is an iterative process which needs an initialization close to the final solution. This initialization must fall within the attraction zone of the snake. We propose to use line junctions to re-localize road location from the road database, because they indicate the position of road intersections in the image. For this purpose, we use a correlation measure between certain points in the road database and line junctions. The road database provided by the Center for Topographic Information of Geomatics Canada is made up of lists of georeferenced points defining line segments. We call these line segments vectors. These points are in georeferenced coordinates, thus we have to convert them in coordinates relative to the beginning of the image. Then, we group the vectors into road segments in order to reduce the number of snakes. These road segments must be as straight as possible because this allows us to add a constraint of high rigidity to the snake. This high rigidity will help avoid noise and shadowing problems, because the optimized contour does not follow small intensity variations caused by these phenomenon s. We link two vectors if they are adjacent, if the cosine of the angle between them is below a threshold (in our experiments, we always set this threshold to cos 138 ffi ) and if there is no other vector starting at the extremity common to the two vectors. Thus, we find road segment extremities first, where there are more than two vectors ending at a single point; second, where the angle between two vectors is lower than a given value; and third, at terminations. A road segment is described by all vectors between its two extremities. Figure 5(a) shows an example of road segments from the database. There are five road segments in this image. Having the list of all road segments, we will match each extremity with a junction, if possible. For this purpose, we compute a correlation measure between the extremity and each junction localized in a N m N m neighborhood around this extremity. Let (x; y) be the extremity and (u; v) be the junction; then, the correlation is: corr(x; y; u; v) = MX m=2 MX m=2 RoadSeg(x + i; y + j)line(u + i; v + j); (2) i= M m=2 j= M m=2 where RoadSeg(x; y) = ρ 1 if (x; y) belongs to a road segment 0 if not: 6

7 (a) (b) (c) Figure 5: a) Road segments. Black squares are extremities. b) Matching between image junctions and road-segment extremities. White squares are junctions matched with extremities, while black squares are either extremities or other junctions. Neighborhood is c) Road segments after re-localizing the two extremities matched with junctions. and line(x; y) is the line plausibility ( max ) as defined in Section 3. We match (x; y) with the junction (u; v) having the highest correlation. Figure 5(b) presents the results of matching between extremities of Figure 5(a) and junctions of Figure 4. We used a M m of 25 in all of our trials. After matching, we re-localize each segment according to the difference between its extremities and the junctions matched. As the differences at the two extremities may not be the same, we move each vector extremity in the road segment according to the difference between the nearest road segment extremity and its matched junction. That means that every vector keeps the same orientation except the one containing the midpoint of the road segment, because its two extremities will move according to the two different extremities of the road segment (Figure 6). Figure 5(c) presents road segments after re-localization. 5.2 Active Contour Models - Snakes As we mentioned earlier, we want to find lines near the initialization provided by the road database. Active contour models or snakes is an optimization process which needs an initialization near the solution [4, 7, 12]. Aerial images can be affected by noise and roads can be partially hidden by shadows of trees, houses or clouds. Snakes manage these problems with constraints on rigidity and elasticity. In the continuous domain, snakes (or active contours) are parametric curves (v(s; t) =(x(s; t); y(s; t)); s2 [0; 1]) that minimize an energy function based on internal and external constraints [4] at time t: E tot = E int + E ext = Z 1 0 ff1 jv 0 (s; t)j 2 + fi 1 jv 00 (s; t)j 2 ds + Z 1 0 max (v(s; t))ds; (3) where E ext is the external energy, E int the internal energy and a regularization parameter. The internal energy represents the snake s capacity to distort. ff 1 ;fi 1 2 R +, v 0 and v 00 are the first and second partial derivatives in s. The term jv 0 (s; t)j 2 influences the snake s length (or tension) and jv 00 (s; t)j 2 influences the curvature (or elasticity) of the snake. The parameter can be absorbed by ff and fi (ff = ff 1 and fi = fi 1 ). 7

8 (x2 + dx1, y2 + dy1) (x2, y2) (x3 + dx2, y3 + dy2) (x4 + dx2, y4 + dy2) (x3, y3) (x4, y4) (x5 + dx2, y5 + dy2) (x5, y5) (dx2, dy2) (x1 + dx1, y1 + dy1) (dx1, dy1) (x1, y1) Midpoint of the road segment Original road segment Modified road segment Figure 6: Example of displacement of a road segment after matching. The external energy represents the image force (f (x; y)) attracting the snake. In our case, this force is the bright-line plausibility (f (x; y) = max (x; y)). Equation 3 is an evolution problem (snake position evolves over time). The solution to this minimization problem is the simplified Euler evolution t) ffv 00 (s; t) +fiv 0000 (s; t) ext(v(s; t)) ; where v 0000 is the fourth derivative in s. The parameter fl represents the viscosity of the curve. The higher the viscosity, the greater the inertia and the slower the evolution of the curve. The discrete version of Equation 4 is obtained by setting a discrete version of the parametric curve at moment t: V t = fv i;t = (x i;t ;y i;t );i = 1::ng, where n is the number on points of the curve. We can approximate the first derivative by: and the second derivative by: The temporal derivative is approximated by: v 0 i;t = v i;t v i 1;t v 00 i;t = v i+1;t 2v i;t + v i 1;t = v i;t v i;t 1 : The solution can be written in the following vectorial form: ffvt 00 + fiv0000 t ext(v t 1 ) : 8

9 At point v i;t,wehave: fl(v i;t v i;t 1) ff(v i+1;t 2v i;t + v i 1;t) +fi(v i+2;t 4v i+1;t +6v i;t 4v i 1;t + v i 2;t) ext(v i;t 1) We can rewrite Equation 5 in matrix (6) (K + fli)v t = flv t ext(v t 1 ) ; where I is the n n identity matrix and K is a n n matrix. The elements of K are determined by Equation 6 and the initial conditions; the snake has free extremities, i. e. v 00 (0;t) = v 000 (0;t) = v 00 (1;t)=v 000 (1;t)=0. Thus K is defined by: fi 2fi fi 0 ::: ::: ::: ::: ::: 0 ff 2fi 2ff +5fi ff 4fi fi 0 ::: ::: ::: ::: 0 fi ff 4fi 2ff +6fi ff 4fi fi 0 ::: ::: ::: 0 0 fi ff 4fi 2ff +6fi ff 4fi fi 0 ::: ::: fl is given by [4]: ::: ::: 0 fi ff 4fi 2ff +6fi ff 4fi fi 0 0 ::: ::: ::: 0 fi ff 4fi 2ff +6fi ff 4fi fi 0 ::: ::: ::: ::: 0 fi ff 4fi 2ff +6fi ff 4fi 0 ::: ::: ::: ::: ::: 0 fi 2fi fi fl p j ; n where is the mean progression of the curve. The initial must be determined by the user. We keep the same until the total energy of the curve increases and then we decrease by a factor of 0:75. We can approximate the total energy of the curve at time t by: E tot;t = nx i=1 Eext (v i;t )+ffjv 0 i;t j2 + fijv 00 i;t j2 To summarize the correction step, for each road segment we do the following: 1. Initialize a snake curve (V 0 ) with the re-localized road segment found in the matching step. Each pixel belonging to the curve is a v i;0. Set t to Compute V t from Equation If the total energy of the curve decreases then increment t and go to step If is not too low, decrease, increment t and go to step 2. Figure 7 shows an example of a snake used to optimize a line plausibility function (ff = 100 and fi =0). For all trials, we used an initial =1. Clearly, the initialization is beside the line. After 18 iterations, the snake is at the center of the line. 9

10 (a) Figure 7: Snakes a) Initialization. Every pixel is a point in V 0. b) After 18 iterations 6 Generation of Hypothesis for New Roads As mentioned earlier, the second problem in updating road database is the detection of new roads. These new roads are also bright lines on a dark background, so the result of the line detector can be used. Our work is based on the supposition that new roads are connected with the existing road network. This hypothesis is realistic because roads are built to allow people to travel from one place to another. For this reason, we consider only lines which are connected to the road network as given in the existing database. The line-following algorithm starts from junctions in a N N neighborhood of the roads, corrected using the snake approach (Figure 8). We follow all lines originating from these (b) Figure 8: Image junctions near the corrected roads (N =7). junctions. We consider each line pixel in a 5 5 window around a junction as a starting pixel for the line-following process (Figure 9). For each line, we proceed as follows : Initially, we take one of the starting pixel (anyone). Let p be the last chosen pixel (we call it the parent). First, we consider three neighbors of this pixel in 10

11 l s l l s j s l l s l l Figure 9: A marked line junction (j), its starting pixels (s) and line pixels (l). the line direction. For example, if the line direction at point p = (p x ;p y ) is in ]ß=8; 3ß=8], the three neighbors to consider are (p x +1;p y +1), (p x ;p y +1)and (p x +1;p y ) (see Table 1). The line direction in p is computed from ~n (Equation 1). Each of these three neighbors which is marked as a line pixel is considered as a child. The next pixel selected is the child that minimizes: j c p j (8) where c is the line direction of the child and p is the mean line directions of the parent respectively. The mean line direction is computed on the last L pixels visited, as follows: a = P L 1 i=0 a i ; L where a i is the line direction of the ith pixel visited before pixel a. In the case where none of the three neighbors are marked as line pixels, we look in a neighborhood further away in the line direction; that is, when the direction of the vector between the parent and the pixel falls in the interval [ p ß=8; p + ß=8], we consider this pixel as a child. In this case, the next pixel is the one minimizing: (1!)jjpt c pt p jj 2 +!j c p j (9) where pt c and pt p are the locations of child and parent, respectively and! 2 [0; 1] is a weighting parameter. We do it iteratively from the beginning with the retained child. If a parent has no child, the line-following process is terminated. If the new road is shorter than a given threshold (S), it is dropped. Let a line junction near the existing network, for each line pixel in a 5 5 neighborhood of this junction, we can summarize the line following algorithm as follows : 1. Set the pixel as parent. 2. Compute the three neighbors in line direction of the parent (see Table 1). If they are line pixels, set them as children. 3. If there is at least one child: (a) Set the child that minimizes Equation 8 as parent. (b) Return to step 2. 11

12 ] ß=8;ß=8] p ]7ß=8; 7ß=8] p ]ß=8; 3ß=8] p ] 7ß=8; 5ß=8] p ]3ß=8; 5ß=8] p ] 5ß=8; 3ß=8] p ]5ß=8; 7ß=8] p ] 3ß=8; ß=8] p Table 1: Range of line directions and neighbors considered. p is the parent pixel and indicates neighbors to verify. 4. If there is no child: (a) Compute further children in the line direction. (b) If there is now at least one child : i. Set the child that minimizes Equation 9 as parent. ii. Return to step If the length of the new line is less than S, then drop it. Figure 10 shows the results of the line-following process with L =10;S =7;! =0:9. Figure 10: New roads found by line-following process with L =10;S =7and! =0:9. 12

13 7 Experimental Results Our algorithm has been implemented in C++ on a Sun platform and in Visual C++ on a Windows platform. It was experimented on various sections of a georeferenced ortho image ( pixels) that we re-sampled in order to reduce the initial spatial resolution by a factor of two. The digital georeferenced ortho image mosaic used has initial ground spatial resolution of 1:5m and is from the Edmonton, Alberta area. It was generated by scanning 1 : scale 1995 aerial photography at 1000ppi (0:025mm) with 8-bit quantization. The Helava/Leica digital photogrammetric workstation was used for the automatic DEM generation, while the exterior orientation parameters for each image were determined through automatic aerotriangulation using control points from the Aerial Survey Database (ASDB). The road vector data were obtained from the National Topographic Database (NTDB). These data were revised and enhanced in 1996 using GPS methods and have an estimated accuracy of about 10m. Figure 11 presents the results obtained at each step of our correction algorithm for an image section. Figure 11(a) is the original image, on which the vectors from the road database are superimposed. Visually, we can see that in many places, vectors do not correspond to roads. Figure 11(b) shows the line plausibility. Figure 11(c) shows the result of line detection, where the scale (ff) is 2 and the line threshold is 20. Figure 11(d) shows junctions (threshold is 0:1) and road-segment extremities, represented by black squares. The white squares are junctions matched with extremities (N m = 13). We note that except for dead ends, there is only one extremity which is not matched (upper part of central curve), because there is no junction in the neighborhood of the extremity. If the nearest junction had been matched, the central segment would have shifted to the right, away from the road. Figure 11(e) shows the initialization after re-localizing road segments. Figure 11(f) presents the final position of the snakes (ff = 100 and fi = 4). We note that the final snake positions are closer to the center of the lines than the original vectors. The mean displacement of road segment points is 2:76 pixels and the variance of this displacement is 4:09. We had a mean number of iterations of 31 per pixel in the snake step, for a total of 641 pixels treated. The results without junction detection (Klang s approach with the same parameters as in Figure 11) are presented in Figure 12. The mean displacement of road segment points is 3:63 and the variance of this displacement is 4:95. The mean number of iterations was 34 per pixel. Comparing Figure 11(f) and 12(b), we can see why the mean displacement in Klang s approach is higher than ours. Some snake sections are displaced further from the road. Roads are better localized with our approach. Figure 13 presents the results for a second image section. Figure 13(a) shows the original image, with road segments from the database and line junctions (the threshold for line detection is 20 and the threshold for line junction detection is 0:15) matched with road segment extremities (N m = 9). Figure 13(b) shows line plausibility (ff = 2:5), and Figure 13(c), the final result of the correction step (ff = 100 and fi = 0). As we can see, the lower right-hand section has been well corrected. In the middle of the image, we have a street in form of U. The right lower part of the U section is badly positioned comparing to the original vectors. We can find the explanation in the line plausibility image. As we can see in Figure 13(b) line plausibility is low and the snake is attracted outside the line. We also observe from the left-hand part of the horizontal segment (lower left, the one which cross the river), that the rigidity of the snake is very useful. Despite the fact that the line detector doesn t give a straight line everywhere in this segment, the result obtained is almost straight. Further along 13

14 this same road, we can see that it lies just beside a dark line on a lighter background. This may be a shadow or something else. The snake is thus attracted by the side of the road instead of the middle. This is a frequent problem in urban regions. The mean displacement of road segment points is 2:26 pixels and the variance of this displacement is 1:47. We had a mean number of iterations of 47 per pixel in the snake step, for a total of 1957 pixels treated. For purpose of comparison, we worked without line junctions as did Klang [13]. We took the point in a neighborhood (same as in Figure 13(a)) where Equation 2 gives maximum correlation. The other parameters are the same. Figure 14 shows the results. The mean displacement of road segment points is 2:35 pixels and the variance of this displacement is 2:27. The mean number of iterations was 39 per pixel. The main difference between the two approaches is the variance of displacement is higher when line junctions are not used. Figure 15 presents the results for a third section of the image. Figure 15(a) shows road segments from the road database and line junctions, superimposed on the original image. The scale parameter ff is 2.5 and the thresholds for line detection and junction detection are 20 and 0:09 respectively. The vertical road is shifted slightly to the left. Figure 15(b) shows the bright-line plausibility, and Figure 15(c), the final result of the correction step (ff = 100 and fi = 0). We can see that the vertical road is better localized. However one intersection is slightly moved toward the bottom and the other is slightly moved toward the top of the image (Figure 17c). From Figure 15(b), the line plausibility is lower in the intersection s location than the neighborhood in the line direction. The snake is then attracted by the highest plausibility. One possible solution to this problem could be to find an application condition for the snake. It could avoid displacing the road segment when the line plausibility is not good enough. We had a mean displacement of 1:68 pixels with a variance of 1:32. It took 22 iterations per pixel, for a total of 908 pixels. Now we will present the results concerning generation of hypothesis for new roads. Figure 15(d) shows the final result of the new-road detection step carried out on Figure 15(a). Figure 16 presents the new-road hypothesis generation step for the image in Figure 11(a). Figure 16(a) shows the line junctions near the known network (N = 7) and Figure 16(b) shows the final result (L = 10 and S =7). We will now briefly discuss the influence of the different parameters. We begin with the scale parameter of the line detector (Figure 17). The scale parameter must be chosen carefully, depending on the width of the road to be detected [15]. If it is too low, the line is seen as two separate step edges. In addition, noise will have a greater influence on the results. If the scale parameter is too high, smoothing may flatten the line and the plausibility will be too low. For a discussion of line-junction parameters see [5]. In the matching step, we have a parameter determining the minimal distance allowed between the line junction and a road extremity (N m ). This neighborhood must be large enough, in case of high imprecision, to allow re-localization of the road segment closer to the line. On the other hand, it should not be too large because the road segment extremity may be matched with the wrong junction and then we would move the road segment away from the line (Figure 18). The snake process has three parameters, ff; fi and. First, when the rigidity parameter (ff) is high with respect to the elasticity parameter (fi), the final road segment is straight (Figure 19(a)). Conversely, when ff is small compared with fi, the final segment tends to oscillate (Figures 19(b) and 19(c)). The main difference between Figure 19(b) and Figure 19(c) is the number of iterations, which is higher for 19(c) (53:6 iterations per pixel) than for 19(b) (35:2 iterations per pixel). In 14

15 general, we have observed that when ff =0and fi increases, the number of iterations also increases. For roads, it is suitable that rigidity should be more important than elasticity, but the snake must be able to fit a low-curvature road. The initial mean progression ( ) of the snake influences the number of iterations required to optimize the energy. If the initial increases then the number of iterations decreases. For example, to obtain the result showed in Figure 19(d), it took 60:3 iterations per pixel and to obtain the result showed in Figure 8, it took 54:6 iterations per pixel. However, beyond a certain value, the snake may fall outside the attraction zone and then evolve in a wrong direction (Figure 19(e)). In the new-road hypothesis generation step, there are two influencing parameters : L and S. L is the number of pixels on which the mean direction is computed. It mainly influences the search direction in the case where there is no direct neighbor and a remote searching is required (step 4 of the line-following algorithm). For example, in Figure 20(a), we used L =5, and we can see that some lines are missing in comparison with Figure 16(b) where L =10. S is the minimum length of line to be kept (Figures 20(b) and 20(c)). 8 Conclusion We have presented our approach for automatically updating road maps from high-resolution aerial images. The updating of road database involves the addressing of two problems, i.e., the correction of existing road location and the addition of new roads. Since roads are long and thin elements in high-resolution images, the method is based on the detection of line. Existing road location is not precise, but can provide a good indication of where roads are located as opposed to other elements such as rivers or railways. For correction, we use a snake approach, with the existing location as the line initialization. To bring the initialization closer to the road, we re-localize road segments on the basis of line junctions from the image. We include a line junction detector developed in our Vision Laboratory at Département de Mathématiques-Informatique at Université de Sherbrooke [5]. We have presented some results obtained with an image and a database provided by Geomatics Canada. For the addition of new roads, we apply a line-following algorithm starting from image junctions which are in the existing network. We have shown that there is an advantage for using line junctions by comparing with Klang s approach. References [1] M.-F. Auclair-Fortier, D. Ziou, C. Armenakis, and S. Wang. Nouvelles perspectives en détection de contour: Textures et images multi-spectrales. In Vision Interface, Trois-Rivières, Canada, May [2] M.-F. Auclair-Fortier, D. Ziou, C. Armenakis, and S. Wang. Survey of Work on Road Extraction in Aerial and Satellite Images. Technical Report 247, Département de mathématiques et d informatique, Université de Sherbrooke, [3] A. Baumgartner, C.T. Steger, H. Mayer, and W. Eckstein. Semantic Objects and context for Finding Roads. In David M. McKeown, Jr., J. Chris McGlone, and Olivier Jamet, editors, 15

16 Integrating Photogrammetric Techniques with Scene Analysis and Machine Vision III, Proc. SPIE 3072, pages , april [4] M.-O. Berger. Les contours actifs : modélisation, comportement et convergence. Thèse de Doctorat, Institut National Polytechnique de Lorraine, [5] F. Deschênes and D. Ziou. Detection of Line Junctions and Line Terminations Using Curvilinear Features. Technical Report 235, Département de mathématiques et d informatique, Université de Sherbrooke, [6] R. Fiset, F. Cavayas, B. Solaiman M.-C. Mouchot, and R. Desjardins. Map-Image Matching Using Multi-Layer Perceptron: the Case of Road Network. ISPRS Journal of Photogrammetry and Remote Sensing, 53(2):76 84, [7] P. Fua and Y. G. Leclerc. Model Driven Edge Detection. Machine Vision and Application, 3:45 56, [8] A. Gruen, P. Agouris, and H. Li. Linear Feature Extraction with Dynamic Programming and Globally Enforced Least Squares Matching. In A. Gruen, O. Kuebler, and P. Agouris, editors, Automatic Extraction of Man-Made Objects from Aerial and Space Images, pages 83 94, Basel, Switzerland, Birkhäuser Verlag. [9] B. Guindon. Application of Spatial Reasoning Methods to the Extraction of Roads from High Resolution Satellite Imagery. In IGARSS, [10] B. Guindon. The Extraction of Planimetric Features from High Resolution Satellite Imagery Using Image Segmentation and Spatial-Based Reasoning. In International Symposium, Geomatics in the Era of RADARSAT (GER 97), [11] C. Heipke, H. Mayer, C. Wiedemann, and O. Jamet. Evaluation of Automatic Road Extraction. Proceedings of International Archives of Photogrammetry and Remote Sensing, XXXII:47 56, [12] M. Kass, A. Witkin, and D. Terzopoulos. Snakes : Active Contour Models. The International Journal of Computer Vision, 1(4): , [13] Dan Klang. Automatic Detection of Changes in Road Databases Using Satellite Imagery. In Proceedings of International Archives of Photogrammetry and Remote Sensing, volume 32, pages , [14] D.M. McKeown and J.L. Denlinger. Cooperative Methods for Road Tracking in Aerial Imagery. In Proceedings of IEEE, International Conference on Computer Vision and Pattern Recognition, pages , [15] C. Steger. An Unbiased Detector of Curvilinear Structures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(2): , February [16] D. Ziou and S. Tabbone. Edge Detection Techniques - An Overview. Pattern Recognition and Image Analysis, 8(4): ,

17 (a) (b) (c) Figure 11: a) Original image ( ) with vectors from the road database. b) Line plausibility (ff = 2). c) Detected lines (threshold is 20). d) Junctions (threshold is 0:1) and extremities (black squares). White squares are matched junctions (N m =13). (d) 17

18 (e) Figure 11: e) Initialization of snakes after re-localization. f) Final result (ff =100and fi =4). (f) (a) Figure 12: Results without junction detection. a) Matching of road segments with image. White squares are points matched. b) Final result (ff =100and fi =4). (b) 18

19 (a) Figure 13: a) Super-imposition of vectors from the road database, junctions (white squares are matched junctions), and road-segment extremities on original image ( ). Thresholds for line detection and junction detection are 20 and 0:15 respectively and the neighborhood for matching is

20 (b) Figure 13: b) Line plausibility (ff =2:5). 20

21 (c) Figure 13: c) Final result (ff = 100 and fi =0). 21

22 (a) Figure 14: Klang s approach a) Super-imposition of vectors the from road database, maximum correlation points (white squares) and road-segment extremities ( ) on original image. Threshold for line detection is 20 and neighborhood (N m N m ) for matching is

23 (b) Figure 14: c) Final result (ff = 100 and fi =0). 23

24 (a) (b) (c) Figure 15: a) Original image ( ) with vectors from the road database (white lines) and line junctions (black squares). b) Line plausibility (ff = 2:5). c) Final result of the correction step (ff =100and fi =0). d) Final result of the new road detection step. (d) 24

25 (a) (b) Figure 16: Detection of new roads. a) Line junctions near the corrected network. b) Result (L = 10 and S =7) x x x (a) Line profile (b) First derivative Figure 17: Line detection for small ff. (c) Second derivative 25

26 (a) (b) Figure 18: a) Matching with N m = 15. b) Final result of the correction step with ff = 100;fi = 4; =1. (a) ff = 500;fi = 4; =1 (b) ff =0;fi =4; =1 (c) ff = 0;fi = 100; =1 (d) ff = 100;fi = 4; =0:1 (e) ff = 100;fi = 4; =3 Figure 19: Results of road correction with different parameter values 26

27 (a) L =5, S =7 (b) L =10, S =3 (c) L =10, S =15 Figure 20: New roads for different values of L and S. 27