Estimation of building height using a single street view image via deep neural networks

Building smart cities requires three-dimensional (3D) modelling to facilitate the planning and management of built environments. This requirement leads to high demand for data on vertical dimensions, such as building height, that are critical for the construction of 3D city models. Despite increasing recognition of the importance of such data, their acquisition in a low-cost and efficient manner remains a daunting task. Big data, particularly street view images (SVIs), provide an opportunity to efficiently solve this problem. In this study, we aim to derive information on building height from openly available SVIs by using single view metrology. Unlike other methods using multisource inputs, our method capitalizes on deep neural networks to extract a set of features – such as vanishing points, line segments, and semantic segmentation maps – for single view measurement and then estimates the height from single SVIs. The minimal input required by the method increases its competitiveness in large-scale estimation of building heights, especially in areas with difficulty to obtain the conventional remote sensing data. In addition to experiments that demonstrate the effectiveness and efficiency of the proposed method, we also conduct a thorough analysis of uncertainties and errors brought by the method, thereby providing guidance for its future applications.