Extracting Road Centerlines from High Resolution Remote Sensing Images by Active Window

EXTRACTING ROAD CENTERLINES FROM HIGH RESOLUTION REMOTE SENSING IMAGES BY ACTIVE WINDOW LINE SEGMENT MATCHING

Y. Yang1, C.Q. Zhu2, D. Zhang1

1 - Xian Research Institute of Surveying and Mapping, Department of Cartography and GIS Engineering, Xian, China

2 - Zhengzhou Institute of Surveying and Mapping, Zhengzhou, China

Abstract: A semi-automatic method based on active window line segment matching is proposed for extracting road centerlines from high spatial resolution remote sensing images. Road centerlines can be tracked automatically by defining a series of active template windows lying on a road centerline which start from a user-given input point, and then by using thresholding, line segment matching and improved SSDA. Meanwhile, a little manual processing is allowed in case there is something wrong with automatically road extraction. Experiments on 0.61m-resolution QuickBird and 1m-resolution IKONOS satellite images demonstrate that main road networks can be extracted rapidly and accurately by using the proposed approach. Moreover this extraction is robust to noise interference.

Key words: high resolution remote sensing images; road extraction; active window; line segment matching

1 Introduction

With the development of sensors technology, more and more high-resolution satellite images such as QuickBird and IKONOS images are used in a wide range of applications, thus facilitating information extraction from the remote sensing images. Road extraction is very important for data update in GIS, image matching, target detection and digital automatic mapping [1]. However, satisfactory results have not been achieved in terms of accuracy and integrity by using existing automatic road extraction techniques [2-8]. Post-processing by human being is usually inevitable. Therefore, it is more practical to research semi-automatic methods with a little human being interaction for quick and accurate road extraction [9-12].

In this paper, a new semi-automatic method based on active window line segment matching is proposed for extracting road centerlines from high-resolution remote sensing images. It can be adaptive to the variations in brightness, and is robust to noise interference on the two sides of road centerlines. In addition, the searching strategy of sequential similarity detection algorithm (SSDA) is improved. As a result, faster matching can be obtained. Due to the combination of automatic tracking with little human interaction, this method is more practical.

2 Model and Methodology

2.1 Road Feature Analysis

Fig. 1 shows some examples of main roads in 0.61m-resolution QuickBird imagery. We can notice that road centerlines are clearly visible. They can be regarded as linear features which have distinctive brightness patterns compared to their surroundings. Moreover, most roads vary slowly without sharp changes in their directions. Therefore, the road centerlines can be approximately considered as curves which consist of many short straight lines connected. The extraction of road centerlines can naturally be transformed into the extraction of short straight lines. Based on the this idea, a new method, which is referred to as active window line segment matching, is developed for extracting road centerlines from high-resolution remote sensing images.

(a) Example 1 (b) Example 2

(c) Example 3 (d) Example 4

Fig.1 Examples of roads in 0.61m-resolution QuickBird imagery

2.2 Active Window Line Segment Matching Method

The basic idea of active window line segment matching is as follows:

2.2.1 Defining a template window

At first, select a point in the road centerline as a reference central point to define a template window, as shown in figure 2(a). Then, a thresholding operation is performed to transform the image into a binary one within the window.

The thresholding operation is represented as:

if a centerline is darker than its surroundings；

if a centerline is brighter than its surroundings.

where is the threshold which is calculated by using the optimum threshold algorithm presented in [13].

After the above operation, the road centerline within the window will become a black line whose intensity is zero, as shown in figure 2(b). Then the black line obtained can be used for the following line segment matching.

(a) Original image (b) Thresholding result

Fig.2 Original image and thresholding result

2.2.2 Line segment matching

By translating and rotating the template window in a prescribed scope, we can get some different target windows. The optimum matching is achieved when obtaining maximum similarity between line segment features from target window and template window. The principle of line segment matching is illustrated in figure 3.

The similarity transform from template window to target window can be expressed as follows.

where is the coordinates of a point within a template window, a point within a target window, the rotation angle between the target and template windows, and the respective shift in and directions.

Fig.3 Line segment matching

The similarity measure in line segment matching is defined as

where and are intensity values at point within the template and target windows, respectively. is a weighting factor to determine relative importance of matching points, which reaches its largest value 1 along the axis of the line segment and decreases gradually with the increase of distance from the axis. The size of each window is and line width for matching is . In a matching, the smaller the value of , the more similar in line segment feature between the template and target windows.

2.2.3 Searching optimum target window based on SSDA

In order to find the target window whose line segment feature is most similar to the template window, the template window is first moved in the road direction to create an initial target window. Then the initial target window is rotated and translated in a prescribed scope, and their line segment features are matched to find out the most similar window. In this operation, SSDA searching strategy is used for the sake of its fast matching speed.

The basic idea of SSDA is that if the accumulated error in the computation of similarity is greater than a prescribed threshold, the computation is terminated, thereby significantly reducing the amount of computation for non-matching points and enhancing the matching speed. In this algorithm, it is difficult to find out a suitable threshold. SSDA can be improved by ignoring the threshold, and matching is performed in the way of self-study. At first, the accumulated error of all pixels within the initial target window is calculated. From the second target window, the calculation of the accumulated error stops when it reaches or surpasses the minimum one occurred before. In this way, the optimum target window corresponding to the smallest accumulated error can be obtained once a thorough searching is completed. If the target window corresponding to the smallest accumulated error is found in the earlier stage, the subsequent amount of calculation will be very small. For this reason, we designed a priority order for the translation and rotation of the initial target window, as shown in figure 4, to find the optimum target window as soon as possible.

Fig.4 Priority order of translation and rotation for initial target window

3 Road Network Extraction Scenario

Figure 5 shows the procedure of our road centerline extraction method. Firstly, a starting point and a directional point on a road centerline are inputted by a user to create a template window. To ensure that line segment matching is performed along the road centerline, it is necessary to rectify these two points to become respective midpoints of the road centerline.

Secondly, an initial target window can be obtained by moving the template window along the road direction. Thresholding is applied to the template and the target window, and line segment matching is performed.

Then, a series of target windows can be obtained by rotating and translating the initial window, and line segment matching is performed based on SSDA. Once the matching is successfully completed, a new road center and a new road direction are available.

Figure 6 shows the process of continuous active window matching. If the minimum value of is greater than a prescribed value E, the matching fails. In this case, manual processing is necessary, which includes proceeding with tracking, moving backwards or crossing road sections. As long as manual processing is completed, the matching will continue as before until complete road information is obtained.

Fig.5 The procedure of road centerline extraction

Fig.6 Continuous active window matching

4 Experimental Results

In this section, we used the proposed method to extract road networks in remote sensing images of Shenzhen city. These images were obtained from the satellite QuickBird whose resolution is 0.61m. Some parameters are set as follows: the size of template and target windows is 2525 pixels, the distance between template and initial target window is 20 pixels, the translation limit of initial target window is 22 pixels, the range of rotation angle is from -45º to 45º, the line width in matching is 5 pixels. Once a starting point and a directional point are inputted on a road centerline, the line segment matching will be completed automatically, thus obtaining a series of central points along the road. Figure 7 shows the extraction results of the roads in figure 1. Figure 8 shows the extraction results of another long and curved road. Automatic matching fails when a road centerline is lost or shaded. In this case, manual processing is necessary. After that, automatic matching will be restored.

(a) Result 1 (b) Result 2

(c) Result 3 (d) Result 4

Fig.7 Results of extracted road centerlines in fig.1

Fig.8 Another extracted result of a curved road

Another road extraction experiment is applied to a remote sensing image of part of Shenzhen city. The image is obtained from the satellite IKONOS whose resolution is 1 meter. The area is 12 km9 km, and the parameters in the experiment remain unchanged. The final experiment results are shown in figure 9. The total length of extracted road networks is 310 kilometers, and time required in the experiment performed in a 3 GHz Pentium(R) 4 computer is 60 minutes, including manual processing.

Fig.9 Results of road network extraction in Shenzhen

5 Conclusions

From above experiments, main road networks can be extracted accurately and quickly by using the proposed method. The road extraction in 1m-resolution images can be dealt with nearly in real time, which is 3 times faster than that performed in 0.61m-resolution images. However, in both cases, they are much faster than manual operation. Meanwhile, this extraction is robust to noise interference. Matching is only performed between line segment features within windows of interest, excluding the outside of the centerline, which can overcome the effect of noise outside the centerline. The windows in the matching are continuously updated and the threshold is obtained from the current window. This can make the matching adapt the variations in the intensity along road centerlines. In addition, it makes this method more practical to combine the automatic tracking with little manual processing.

References

[1] Shi Wenzhong, Zhu Changqing, Wang Yu. The review and prospect of extraction road features from remote sensing images [J]. Acta Geodaetica et Cartographica Sinica, 2001, 30(3)：257－261 (in Chinese)

[2] Barzohar M, Cooper D B. Automatic finding of main roads in aerial images by using geometric-stochastic models and estimation [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, 18 (7): 707－721

[3] Trinder J C, Wang Y. Automatic road extraction from aerial images [J]. Digital Signal Processing, 1998, 8(4): 215－224

[4] Laptev I, Mayer H, Linderberg T, et al. Automatic extraction of roads from aerial images based on scale space and snakes [J]. Machine Vision and Applications, 2000, 12(1): 23－31

[5] Zhu C, Shi W, Pesaresi M, et al. The recognition of road network from high-resolution satellite remotely sensed data using image morphological characteristics [J]. International Journal of Remote Sensing, 2005, 26(24): 5493－5508

[6] Shackelford A K, Davis C H. Fully automated road network extraction from high-resolution satellite multispectral imagery [C]. IEEE International Geoscience and Remote Sensing Symposium, Toulouse, 2003: 461－463

[7] Jin X, Davis C H. Automatic road extraction from high-resolution multispectral Ikonos imagery [C]. IEEE International Geoscience and Remote Sensing Symposium, Toulouse, 2003: 1730－1732

[8] Hu X, Tao C V. Automatic highway extraction from high resolution imagery by an energy minimizing based perceptual grouping method [J]. Geomatica, 2004, 58(1): 41－50

[9] Kass M, Witkin A, Terzopoulos D. Snakes: active contour model[J]. International Journal of computer vision, 1988, 1(4): 321－331

[10] Gruen A, Agouris P, Li H. Linear feature extraction with dynamic programming and globally enforced least squares matching [C]. Automatic Extraction of Man-made Objects from Aerial and Space Images. Basel: Birkhaeuser Verlag, 1995. 83－94

[11] Gruen A, Li H. Semi-automatic linear feature extraction by dynamic programming and LSB-snakes [J]. Photogrammetric Engineering & Remote Sensing, 1997, 63 (8): 985－995

[12] Kim T, Park S-R, Kim M-G, et al. Tracking road centerlines from high resolution remote sensing images by least squares correlation matching [J]. Photogrammetric Engineering and Remote Sensing, 2004, 70(12): 1417－1422

[13] Gonzalez R C, Woods R E. Digital Image Processing [M]. 2nd ed. Beijing: Publishing House of Electronics Industry, 2003 (in Chinese)