An innovative framework combining a CNN-Transformer multiscale fusion network and spatial analysis for cycleway extraction using street view images

Cycling is a global activity promoting environmental sustainability and health benefits. Many cities have invested in developing dedicated bike paths. Data about these cycleways, often missing in road databases, form the cornerstone of cycling experience assessment and sustainable urban development planning. Street view images (SVIs) provide detailed street-level information, contributing significantly to extracting these cycleway infrastructure data. However, automatic methods for this task are currently sparse. In this study, we propose an innovative framework combining deep learning with post-processing to extract geo-tagged cycleways from SVIs. First, an innovative neural network, SWDE-Net, is proposed to segment various types of cycleways from SVIs. SWDE-Net integrates the strengths of Convolutional Neural Networks (CNNs) in handling local features and Transformers in managing global dependencies within the DeepLabv3+ encoder-decoder framework. A Cross-Contextual Atrous Spatial Pyramid (CC-ASP) module for enhanced multiscale features and a Spatial Global-local Fusion (SGLF) module for precise spatial information. Following cycleway extraction using SWDE-Net, a post-processing approach integrating roadside determination and majority voting is proposed to geotag the extracted cycleways to the road network. Experimental results show that SWDE-Net achieved a mean Intersection over Union (mIoU) of 85.57 % on the constructed cycleway dataset outperformed the state-of-the-art models. Applied to Ghent (Belgium), the framework achieved a geometric correctness of 86.25 %. Evaluations in two additional cities further illustrate the proposed framework's robustness. Therefore, this study provides a highly flexible and robust method for automatically constructing bicycle infrastructure datasets from SVIs, facilitating the development of active mobility solutions in urban environments.