Virtual analysis of urban road visibility using mobile laser scanning data and deep learning

This study proposes a new computer-aided framework for virtually analyzing urban road visibility using mobile laser scanning (MLS) data. The proposed framework compromises three main parts: 1) based on the data reorganization procedure, the 3D U-net is successfully introduced to tackle the issue of non-stationary noises that significantly hinders accurate detections of stationary sight obstructions, 2) a multi-step procedure is developed to extract the road areas to be estimated and fill data gaps caused by occlusions, and 3) a virtual scanning method (VSM) is proposed to achieve a fast and accurate visibility assessment of the extracted road areas. The proposed VSM also facilitates the application of deep neural networks in the automated driving domain to classify sight obstacles. By enabling multiple outputs, the proposed virtual framework provides a comprehensive understanding of urban road visibility, which can help road administrators detect and understand poor-visibility locations on urban streets.