Region-based perceptual grouping for road extraction from high-resolution images

Abstract

A novel region-based perceptual grouping algorithm is proposed for road segment identifying in urban areas from aerial images. Unlike the conventional edge-based approaches, our grouping is regional basis. An imagery scene is modelled by the spectrally homogeneous regions restrained in shape and size. Given a set of low level image features, perceptual grouping is performed on the regions to generate a higher level of structure, from which road segments are extracted. In our unique approach, perceptual grouping is based on the similarity of spectral, orientation and proximity of the candidate regions. The road candidates are further merged into larger regions in terms of their orientation and adjacency. The road segments are extracted from these merged regions on geometric measurement. Experiments were carried out on imagery of real scenes for evaluation. The results showed the approach has the advantage of distinguishing roads from artefacts caused by parking lots and buildings.