EXTRACTING ROAD NETWORK FROM HIGH RESOLUTION REMOTE SENSING IMAGES BASED 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 Mappingy, Zhengzhou, China

yytoall@126.com

 

With the development of remote sensors and satellite technologies, more and more high-resolution satellite data such as QuickBird and IKONOS images are used in a wide range of applications. Extracting road is one of the potential applications of the high-resolution satellite images.

In this paper, we propose a newly developed method----- Active Window Line Segment Matching ---- to extract road centerlines from high-resolution remote sensing satellite images. The method is based on the road characteristics of high-resolution satellite remote sensing images. The basic idea of the method is to define a template window on a road centerline, and then to search a target window along the road whose line segment feature is most similar to the template window to determine next central point in the road. By repeating the above process, we can obtain a series of central points in the road and the corresponding final road centerline from them.

The basic steps of the proposed method are as follows.

(1)   Defining a template window

Select a point in the road centerline as a reference central point to define a template window. Then, a thresholding operation is performed directly to transform the image into a binary one within the window. After the above operation, the road centerline within the window will become a black line whose intensity is zero. Then the black line obtained can be used for the following line segment matching.

(2)  Line segment matching

By translation and rotation of 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 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. In a matching, the smaller the value of , the more similar in line segment feature between the template and target windows.

(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, we first move the template window in the road direction and create an initial target window. Then we rotate and translate the initial target window in a prescribed scope, match their line segment features and find out the most similar window. In this operation, the searching strategy of sequential similarity detection algorithm (SSDA) is used for the sake of its fast matching speed. And SSDA is also improved by ignoring the threshold and matching in a way of self-study.

(4) Determining the road centerlines

Once the matching is successfully completed, a new road center and direction are available. By repeating the above process, we can obtain a series of central points in the road and the corresponding final road centerline from them. If the matching fails, 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.

Based on the proposed method, some experiments to extract road centerlines on 0.61m-resolution QuickBird and 1m-resolution IKONOS images are carried out. It only needs one hour to extract centerlines of main road from a 1m-resolution IKONOS image with the length of 310 kilometers, including manual processing. Hence the proposed road extraction method is rapidly. Moreover this extraction is robust to noise interference.