Aprimoramento do processo SfM para fotogrametria terrestre pela detecção e eliminação de elementos móveis e céu usando YOLOv8

Enhancement of the SfM process for terrestrial photogrammetry through detection and removal of moving elements and background using YOLOv8

Abstract

This article proposes an automated method to enhance terrestrial photogrammetry processes by detecting and removing mobile (vehicles, people) and static (sky) elements using YOLOv8. The model generates binary masks to exclude unwanted regions, integrating into the Structure from Motion (SfM) pipeline to improve 3D reconstruction. Datasets such as Clouds-1000 (sky) and COCO (mobile objects) were used to train YOLOv8, validated in a case study of 3D documentation of a historical building. Results showed a 5.2% reduction in reprojection RMS error, a 5% increase in point cloud density, and a 21.7% decrease in outliers, along with a 6% reduction in processing time. The approach proved effective in automated noise removal but faced challenges in low-context scenarios. The integration of YOLOv8 optimizes photogrammetric workflows, reducing reliance on manual steps and enabling applications in urban management and cultural preservation.