Paper Title
ENHANCEMENT OF UAS-SFM WORKFLOW FOR BUILDING EXTRACTION USING SEMANTIC IMAGE SEGMENTATION
Abstract
Unmanned Aerial System-Structure from Motion (UAS-SfM) photogrammetry mapping has become a popular technique for generating detailed 3D models of buildings and urban environments. However, the lack of inherent semantic information in generated dense point clouds limits their interpretability and usability. This paper introduces a solution by integrating semantic image segmentation within the Deep Learning (DL) framework.Theproposed approach seeks to enhance photogrammetry workflows by directly extracting objects from point clouds before 3D reconstruction, primarily focusing on building extraction. Despite challenges like insufficient ground truth data and UAS limitations, the proposed method outperforms in-built classification techniques, offering improved accuracy and efficiency. It provides SfM software with guidance for optimal point cloud generation, mitigating noisy reconstructions and intricate mathematical procedures through a blend of photogrammetry and DL. By addressing the interplay of influencing factors, the proposed method demonstrates its potential for advancing the precision and detail of generated point clouds, opening avenues for more accurate mapping results.
Keywords - Deep Learning, Semantic Segmentation, Photogrammetry, Building Segmentation