CITYWIDE ESTIMATION OF PARKING SPACE USING AERIAL IMAGERY AND OSM DATA FUSION WITH DEEP LEARNING AND FINE-GRAINED ANNOTATION
Keywords: Deep Learning, Aerial Imagery, Image Segmentation, Parking Space Management, OpenStreetMap
Abstract. Emerging traffic management technologies, smart parking applications, together with transport researchers and urban planners are interested in fine-grained data on parking space in cities. However, there are no standardized, complete and up-to-date databases for many urban areas. Moreover, manual data collection is expensive and time-consuming. Aerial imagery of entire cities can be used to inventory not only publicly accessible and dedicated parking lots, but also roadside parking areas and those on private property. For a realistic estimation of the total parking space, the observed use of multi-functional traffic areas is taken into account by segmenting not only parking areas but also roads according to their purpose. In this paper, different U-Net based architectures are tested for detecting all these types of visible traffic areas. A new large-scale, high-quality dataset of manual annotations is used in combination with selected contextual information from OpenStreetMap (OSM) to depict the actual use as parking space. Our models achieve a good performance on parking area segmentation, and we show the significant impact of OSM data fusion in deep neural networks on the simultaneous extraction of multiple traffic areas compared to using aerial imagery alone.