Automatic Line Network Extraction from Aerial Imagery of Urban Areas through Knowledge-Based Image Analysis

Abstract

Abstract : Investigations of three different methods for the automatic extraction of the traffic network from aerial imagery of rural as well as urban areas have shown the strengths and weaknesses of these methods. A method for the automatic extraction of line objects from large aerial images of prevailing rural areas works successfully in this domain. However, when applied to images of urban areas, success and reliability of the method decrease remarkably due to the complex mixture of natural and man made objects at those locations. A next method we have examined is best described as knowledge-based approach for the analysis of aerial images. In this approach image analysis is performed ast three levels of abstraction, namely iconic or low-level image analysis, symbolic or medium-level image analysis, and semantic or high-level image analysis. Domain dependent knowledge about prototypical urban areas is incorporated via a semantic network. However, depart from using enhanced segmentation, we decided not to follow this approach, because a third method we have investigated will promise reasonable results at lower cost. This method uses a blackboard-oriented production system for image understanding. Under the premise that the crossings of the traffic network can be detected, this method is able to extract the traffic network from aerial images of a complexity like the Bietigheim image (South-West Germany). Our approach for the automatic extraction of the traffic network from aerial images consists of a combination of the three methods described above.