A cost-effective approach to uncovering and mapping buildings’ exterior arcades using street view imagery

Abstract

The exterior arcade of a building is a distinct architectural feature that provides a constructed space offering a covered and protective area. This structure not only benefits people but also serves as valuable data for various research fields, such as construction safety management and urban planning. However, research to date has not tackled the detection and mapping of arcades. This study proposes an efficient deep learning-based approach to uncover arcades and achieve arcade mapping using readily accessible street view images. The novelty of this study lies in exploring buildings’ arcades based on a YOLOv5 framework with proposed annotation approaches for arcade detection in which arcades are treated as untouchable “abstract objects” rather than the common approach which treats them as touchable “material objects.” Moreover, a simple approach to arcade mapping is proposed by combining the detected arcades with imagery parameters and location information. The experiments conducted in downtown Taipei, Taiwan have demonstrated excellent performance, which yield impressive results with an F1-score of 0.87 for arcade detection among models, as well as an RMSE of 0.52 m for on-road types and 0.64 m for at-intersection types in arcade mapping. Our method contributes a new and innovative approach to exploring arcades and establishing an arcade dataset using deep learning, which enhances geospatial investigation and development in real-world environments.