Research On Extraction And Contour Vectorization Of Buildings From High-resolution Remote Sensed Images

Artificial objects including buildings, roads and large engineering facilities are the principal part in urban high-resolution images, among which buildings and roads account for about 80% of images. As the primary object and important mapping element, recognition and extraction of buildings from images occupies a large proportion of the whole work and directly affects the automation level of mapping.Although at present some progress has been made on researches of building detection and extraction, most work are completed manually, which takes up more than half or even over 80% of the entire workload. Existing theories and methods are still not mature and robust enough to meet the requirements of practical applications.Therefore, It's of great significance to extract buildings quickly and automatically from a huge mass of image data, which is also the major orientation in this paper. According to the characteristic of high-resolution remote sensed images in Tokyo, an effective building extraction and vectorization method is proposed in order to provide a more accurate object extraction method with multi-source data and reference to automatic building modeling.In this paper, the high-resolution aerial imagery and digital surface model of Tokyo are chosen as the research data. Firstly, four segmentation algorithms are proposed and compared, the conclusions are as follows: (1) The improved marker-controlled watershed algorithm based on aerial imagery can avoid over-segmentation efficiently, but the building contour could not be well described.(2) Building areas of interest are precisely obtained by use of the building extraction method based on DSM.(3) Based on the segmentation of DSM, a refined building contour can be extracted as level-set method is adopted.(4) Compared with the level-set method, the multi-scale one based on the segmentation of DSM occupies higher precision and the building contour can be better described.Then, based on the regional segmentation, building contour vectorization method by use of corner detection is put forward. Experiments show that compared with the traditional vectorization methods, the proposed one offers a suitable and robust solution to contour vectorization of buildings in high-resolution images with high accuracy and efficiency. It provides a new sollution for modeling and 3D reconstruction of buildings with theoretic and realistic meaning.