ROBUST SKYLINE EXTRACTION ALGORITHM FOR

MOUNTAINOUS IMAGES

Sung Woo Yang, Ihn Cheol Kim and Jin Soo Kim

3-1-2, Agency for Defense Development, Yuseong P.O. BOX 35, Daejeon, Republic of Korea

Keywords: Skyline extraction, mountainous images, canny edge images, skyline candidate pixel.

Abstract: Skyline extraction in mountainous images which has been used for navigation of vehicles or micro

unmanned air vehicles is very hard to implement because of the complexity of skyline shapes, occlusions by

environments, difficulties to detect precise edges and noises in an image. In spite of these difficulties,

skyline extraction is a very important theme that can be applied to the various fields of unmanned vehicles

applications. In this paper, we developed a robust skyline extraction algorithm using two-scale canny edge

images, topological information and location of the skyline in an image. Two-scale canny edge images are

composed of High Scale Canny edge image that satisfies good localization criterion and Low Scale Canny

edge image that satisfies good detection criterion. By applying each image to the proper steps of the

algorithm, we could obtain good performance to extract skyline in images under complex environments.

The performance of the proposed algorithm is proved by experimental results using various images and

compared with an existing method.

1 INTRODUCTION

Skyline extraction is similar to a segmentation

problem which partitions the image into the sky and

non-sky areas. The skyline extraction in mountain-

ous images is very useful in that we can obtain many

local spatial features from skyline that hardly change

even though time goes by. It is used for navigation

of vehicles or micro unmanned air vehicles (Ettinger

et al., 2002; Messi, 2003; Truchetel, 2006). And it

can also be used for rendering cartographic data,

rendering self-shadowing textures, accelerating

flight simulation, visualizing scientific data, path

planning to avoid detection etc (Stewart, 1998). But

it is very difficult to extract skyline from moun-

tainous images because of compexity and diversity

of the skyline and influence by the noise caused by

complex environments. Clouds, fog and backlight by

the sun in an outdoor environment make the skyline

ambiguous. And they also make it hard to extract

skyline. Because of these difficulties, skyline extra-

ction in mountainous images is one of the most

difficult problems to solve in computer vision fields.

There are two approaches to skyline extraction. One

is region-based approach which uses the

characteristics of images that the sky often occupies

the upper part of an image (Fang et al., 1993; Stein

el al., 1992; Cozman et al., 1997, 2000). And the

other is edge-based approach which uses the fact that

skyline can be regarded as a boundary between two

distinctive regions (Talluri and Aggarwal, 1992; Lie

et al.,2005; Woo et al., 2005).

The approach proposed in Fang et al.’s work (1993)

uses the threshold to find the skyline. They calculate

the threshold using ten small sub-windows and the

contrast of the image. After the threshold is

determined, a vertical-line search from top to bottom

is performed. Then, the pixels whose intensity is

below the threshold are determined as skyline. Their

approach has weak robustness when the image has

complex environments, such as clear clouds above

the skyline. Cozman et al.’s approach(2000) also

uses vertical-line search. However, they use the

smoothed intensity gradient image. This approach

also has the drawback of weak robustness for

complex environments. In the Stein et al.’s work

(1992), they find the skyline using the segmentation

method which segments the sky and the ground. But

they don’t mention the crucial segmentation step,

and general segmentation methods have a limitation

to find exact skyline. Talluri and Aggarwal(1992)

use gradient value to extract the skyline based on

edge-based approach. But their approach is too

simplified to be applicable to complex images. A

more practical approach is advocated by Woo et al.

(2005) and Lie et al.(2005). They also use edge-

based approach. In Woo et al.’s work, they use

Dynamic Programming using contrast cost and

253

Woo Yang S., Cheol Kim I. and Soo Kim J. (2007).

ROBUST SKYLINE EXTRACTION ALGORITHM FOR MOUNTAINOUS IMAGES.

In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IFP/IA, pages 253-257

 SciTePress

homogeneity cost. In Lie et al.’s work, they use DP

utilizing edge image, vertex cost and link cost.

Unfortunately, DP has its limitation to extract

skyline in that false skyline can be extracted due to

the dependence on cost equation. A skyline candi-

date point is essential for using DP but the papers

don’t provide exact method to find it.

In this paper, we combine region-based and edge-

based approach to extract skyline. The paper

contains two fundamental contributions. The first is

that this research presents the solution to the problem

of extracting skyline using two-scale canny edge

images. The second is the development of robust

algorithm to extract skyline using the advantages of

each canny edge images, the characteristics of the

images and a proper linking algorithm. Detailed

steps and experimental results are explained in the

following sections.

2 TWO-SCALE CANNY EDGE

IMAGES

Canny edge detector (Canny, 1986) is one of the

most useful edge detectors in today’s computer

vision community and it is optimal in a mathematical

sense. In spite of these good reputations, it is not

adequate for skyline extraction. Because it has

Localization-Detection Tradeoff (Trucco and Verri,

1998; Forsyth and Ponce, 2003). Good detection

means the detector must minimise the probability of

false positives. Good localization means that the

detected edges must be as close as possible to the

true edges. Optimal edge detector must satisfy both

criteria but canny edge detector cannot improve both

criteria simultaneously. It depends on the smoothing

filter’s scale:

and threshold values at the

Hysteresis Thresholding step in canny edge dete-

ctor:

. By adjusting these parameters, we can

improve only one criterion, either good localization

or good detection.

We cannot improve two criteria simultaneously.

However, if we use each criterion in proper steps, we

can obtain the same effect as improving both criteria

simultaneously. So we make two-scale canny edge

images. The first image satisfies good detection

criterion. And the second image satisfies good

localization criterion.

Figure 1 explains two sale canny edge images. We

can obtain the center images of the Fig. 1 by setting

the four parameters

)(

lhyx

as (4, 4, 20, 1)

and the right images by setting them as (3, 3, 5, 0).

We can find out that the right images have more

detailed edge segments than the left images. We will

define the left image as High Scale Canny edge

image and the right image as Low Scale Canny edge

Figure 1: Three sample images in our image database and the corresponding two-scale canny edge images.

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

254

image. We will use High Scale Canny edge image

for the processes which require good localization

and Low Scale Canny edge image for good detection.

3 SKYLINE EXTRACTION

In this section, we introduce skyline extraction

algorithm using two-scale canny edge images. There

are two steps for skyline extraction. The first step is

the seed selection and the second is skyline search.

The seed is a definite point that we can think as a

skyline candidate point. From this point, we can

search the whole skyline using feasible search

algorithm. In the subsequent sections, we explain

detailed procedures.

3.1 Seed Selection

We need to find the skyline from many edge

segments in the canny edge image. For this, we look

for a definite skyline candidate point at first. We use

the topological information that skyline in the

mountainous image has a breakpoint at top of a

mountain, and the location of the skyline in an image

which may occupy the upper part of an image.

To find the breakpoint, we use a maximum point and

two local minimum points of an edge segment. We

define a maximum point as a point that has the

highest y position in an edge segment. From the

maximum point, we search an edge segment. A first

point that has the minimum y position is the local

minimum point. We can get two local minimum

points from left and right side of the maximum point.

We can make two lines from maximum point to left

local minimum point, and from maximum point to

right local minimum point. Joint angle between these

two lines determines validity of the seed point.

Considering the fact that the skyline usually pose in

upper part of an image, we search the canny edge

image from top to bottom and select a first maxi-

mum point of an edge segment that satisfies the

specific joint angle as the seed. We process this step

in High Scale Canny edge image. Fig. 2 shows the

seed selection. Three white dots in the image mean

left local minimum point, the maximum point and

the right local minimum point.

Figure 2: The example of seed selection.

We can observe advantages of adopting High Scale

Canny edge image for finding the seed. High Scale

Canny edge image only has strong edges that have

high gradient values. So if there are not strong noises

in the sky, for example thick clouds, there are not

severe noises enough to disturb finding the seed in

High Scale Canny edge image. If we use Low Scale

Canny edge image, we may not find the appropriate

seed. Fig.3 shows this situation. The first row is the

original images. The second row is the seed found in

High Scale Canny edge image and the third row is

the seed found in Low Scale Canny edge image.

Figure 3: Seed selection using High Scale Canny edge

image: the second row, and Low Scale Canny edge image:

the third row.

The disturbance of the noises that exists even in

High Scale Canny edge image can be overcome by a

simple verification step. Using the found seed, we

search the skyline. If the length of the extracted

skyline is below specific threshold, we decide that

this seed is not adequate, and find other seed from

that position. In this way, we can find the pertinent

seed.

3.2 Skyline Search

We start to search whole skyline from the seed found

in previous stage. Skyline search is processed along

an edge segment of the seed exists, and at first we

search the right side of the seed. We search right 5-

neighbor of the seed. If there is an edge pixel in this

region, we set this point as a skyline candidate pixel.

If there is no edge pixel, we search additional 9-

neighbor of the seed. Fig. 4 explains skyline search

algorithm. Black pixels are the skyline candidate

pixels and dashed boxes show 5-neighbor and

ROBUST SKYLINE EXTRACTION ALGORITHM FOR MOUNTAINOUS IMAGES

255

additional 9-neighbor of the skyline candidate points.

Grey pixels in the dashed box are not selected as

skyline candidate pixels because additional 9-

neighbor search is followed by 5-neighbor search. If

we find an edge pixel in 5-neighbor, we do not

search additional 9-neighbor. This procedure makes

the skyline search algorithm more accurate. If there

is not a skyline candidate pixel, we search reason-

able area around previous skyline candidate pixel.

When there is an edge pixel, we continue the search

process from this pixel. If we cannot find a pixel in

this area, we think that the last skyline candidate

pixel is the end of the skyline, and stop the search

process. Search process is equally done on the left

side of the seed.

Figure 4: Skyline search using 5-neighbor and additional

9-neighbor of the skyline candidate points.

Skyline search is processed in Low Scale Canny

edge image. It is not proper for the seed selection but

for skyline search. Because Low Scale Canny edge

image contains weak edge components that have

weak gradient values as well as strong edge compo-

nents. Fig. 5 explains the suitability of Low Scale

Canny edge image for skyline search process. The

second image is the result of the skyline search using

High Scale Canny edge image, and the third image is

the result using Low Scale Canny edge image. We

can see that the skyline of the third image is more

exquisite than that of the second image.

Figure 5: Skyline search using High Scale Canny edge

image: the second image, and Low Scale Canny edge

image: the third image.

4 EXPERIMENTAL RESULTS

In our experiments, 55 images acquired in different

environments were tested. And we could extract

skylines accurately in 50 images. We could extract a

part of the skylines in the other 5 images. We can see

that the proposed algorithm works accurately even

when there are complex environments in the images.

Some of the test images are shown here. The results

are compared with Woo et al.’s, 2005 gradient based

DP method. In Fig.6, the first column shows the

original test images. The second column shows the

seed using the proposed method, the third column

shows the extracted skyline using the proposed

method and the last column shows the extracted

skyline using Woo et al.’s method. We can see that

in Woo et al.’s method, it confuses the skyline and

other lines that have large gradient. And it can not

search the whole skyline accurately because of the

noises, for example clouds and fog. But the proposed

method still works well under the complex

environments.

5 CONCLUSIONS

In this paper, we presented a new robust skyline

extraction algorithm in mountainous images. The

major contribution is the use of two-scale canny

edge images and the development of new skyline

extraction algorithm. Using topological information,

location of the skyline in an image and a feasible

search algorithm, we developed a new skyline

extraction algorithm. By applying two-scale canny

edge images to proper skyline extraction steps, we

could overcome the defects of two images. We

applied our algorithm to 55 images that have various

complex environments. In about 90% of the images

we could find exact skylines. The experimental

results show the robustness of our algorithm under

complex environments.

REFERENCES

Canny, J., 1986. A computational theory for edge

detection. In IEEE Trans. On Pattern Recognition and

Machine Intelligence, vol.26, no.6, pp.679-698.

Cozman, F., Krotkov, E., 1997. Automatic mountain

detection and pose estimation for teleoperation of

lunar rovers. In Proc. Of the International Conference

on Robotics and Automation, pp.2452-2457.

Cozman, F. and Krotkov, E. and Guestrin, C., 2000.

Outdoor visual position estimation for planetary

rovers. Autonomous Robots, 9, pp.135-150.

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

256

Ettinger, Scott M., Nechyba, Michael C., Ifju, Peter G. and

Waszak, Martin, 2002. Vision-guide flight stability

and control for micro air vehicles. In IROS ’02, proc.

IEEE/RSJ International Conference on Intelligent

Robots and System, pp.2134-2140.

Fang, M.., Chiu, M.-Y., Liang, C.-C. and Singh, A., 1993.

Skyline for video-based virtual rail for vehicle

navigation. In Proc. IEEE International Sympos. On

Intelligent Vehicles, pp.207-212.

Forsyth, D. A. and Ponce, J., 2003. Computer Vision A

Modern Approach, Prentice Hall, Upper Saddle River,

NJ.

Lie, W.N., Lin, T.C.-I., Lin, T.-C. and Hung, K.-S., 2005.

SA robust dynamic programming algorithm to extract

skyline in images for navigation. In Pattern

Recognition Letters, 26, pp.221-230

Messi, Mark, et al., 2003. Vision chip flight stability and

control for micro air vehicles, In ISCAS ’03, proc.

IEEE International Conference on Circuit System,

pp.786-789.

Stein, F. and Medioni, G., 1995. Map-based localization

using the panoramic horizon. IEEE Trans. On

Robotics and Automation, 11(6), pp. 892-896.

Stewart, J.A., 1998. Fast horizon computation at all points

of a terrain with visibility and shading applications.

IEEE Trans. Visual. Comput. Graphics, 4(1), pp.82-93.

Trucco, E. and Verri, A., 1998. Introductory techniques

for 3-d computer vision, Prentice Hall, Upper Saddle

River, NJ.

Truchetel, F., et al., 2006. Attitude Measurement by

Artificial Vision. Measurement Science and

Technology, vol. 17, pp.101-110.

Talluri, R. and Aggarwal, J., 1992. Position estimation

for an autonomous mobile robot in an outdoor

environment. IEEE Trans. Robotics and Automation,

8(5), pp.573-584.

Woo, J.H., Kweon, I.S., 2005. Robust horizon and

peak extraction for vision-based navigation. In

MVA’06 Proc. IAPR workshop on Machine Vision

Applications.

Test image Seed selection Skyline search

Gradient based

DP search

Figure 6: Experimental results for six chosen test images and the comparison with the results of Woo et al.’s method.

ROBUST SKYLINE EXTRACTION ALGORITHM FOR MOUNTAINOUS IMAGES

257