Accurate urban road centerline extraction from VHR imagery via multiscale segmentation and tensor voting

• A novel hybrid feature (spectral, structural and contextual) is proposed.
• A TV-NMS based method is introduced to extract smooth road centerline network.
• A fitting algorithm is proposed to obtain complete centerline network.
• A new and challenging dataset is public available for further research.

Accurate road centerline extraction from very-high-resolution (VHR) remote sensing imagery has various applications, such as road map generation and updating etc. There are three shortcomings of existing methods: (a) due to noise and occlusions, most road extraction methods bring in heterogeneous classification results; (b) morphological thinning is a fast and widely used algorithm to extract road centerline, while it produces small spurs; (c) many methods are ineffective to extract centerline around the road intersections. To address the above three issues, we propose a novel road centerline extraction method via three techniques: fused multiscale collaborative representation (FMCR) & graph cuts (GC), tensor voting (TV) & non-maximum suppression (NMS), and fitting based centerline connection. Specifically, FMCR-GC is developed to segment the road region from the image by incorporating multiple features and multiscale fusion. In this way, homogenous road segmentation can be achieved. Then, TV-NMS is introduced to generate a road centerline network. It not only extracts smooth road centerline, but also connects the discontinuous ones together. Finally, a fitting based algorithm is proposed to overcome the ineffectiveness of existing methods in the road intersections. Extensive experiments on two datasets demonstrate that our method achieves higher quantitative results, as well as more satisfactory visual performances by comparing with state-of-the-art methods.