Semantic structure from motion for railroad bridges using deep learning

Abstract

Current maintenance practices consume significant time, cost, and manpower. Thus, a new technique for maintenance is required. Construction information technologies, including building information modeling (BIM), have recently been applied to the field to carry out systematic and productive planning, design, construction, and maintenance. Although BIM is increasingly being applied to new structures, its application to existing structures has been limited. To apply BIM to an existing structure, a three-dimensional (3D) model of the structure that accurately represents the as-is status should be constructed and each structural component should be specified manually. This study proposes a method that constructs a 3D model and specifies the structural component automatically using photographic data with a camera installed on an unmanned aerial vehicle. This procedure is referred to as semantic structure from motion because it constructs a 3D point cloud model together with semantic information. A validation test was carried out on a railroad bridge to validate the performance of the proposed system. The average precision, intersection over union, and BF scores were 80.87%, 66.66%, and 56.33%, respectively. The proposed method could improve the current scan-to-BIM procedure by generating the as-is 3D point cloud model by specifying the structural component automatically.