Improvement of 3D Modeling Efficiency and Accuracy of Earthwork Site by Noise Processing Using Deep Learning and Structure from Motion

Nowadays, we can relatively easily create 3D models of earthwork construction sites by taking many pictures from multiple viewpoints using Unmanned Aerial Vehicles (UAV) and by applying Structure from Motion (SfM) technique to the pictures. However, since many construction machines are moving at the construction site, they become noise or disturb the model when generating a 3D model. Thus, it is necessary to take measures such as removing the noise, i.e., construction machinery after surveying, stopping construction work during the survey, or surveying on weekends when all machines are parked at proper places, which can result in reduced efficiency and productivity. On the other hand, object detection technology using deep learning is making great progress. Therefore, this research proposes an efficient and accurate UAV photogrammetry system that generates high-precision 3D models by detecting and removing construction machinery from UAV aerial images in advance. First, the object detection method by deep learning is used to detect the construction machinery. Next, the parts of the construction machines are removed from the images, and the removed parts are completed by image interpolation or compensation. Finally, a 3D model is created using SfM. We developed a system based on the proposed methodology and applied it to an actual earthwork construction site. The result showed that accuracy and efficiency have been enhanced by using this system.