UAV영상과 딥러닝 세그멘테이션을 활용한 설계도면의 효율적인 벡터화 방법

Topographic surveying involves mapping critical national infrastructure such as roads and ports, and visualizing them using computer-aided design (CAD). However, manual drafting in this process poses concerns regarding potential errors and extended working hours. This study aimed to explore the potential for improving the vectorization process of design drawings by combining unmanned aerial vehicle (UAV) high-resolution orthophotos with deep learning algorithms. For training data, high-resolution UAV orthophotos (ground sample distance, GSD: 5 cm) covering a total area of 5.55 km² from five municipalities in South Korea were used, and ground truth (GT) data was derived from drawings approved by the Spatial Information Quality Control Agency. The training and inference results demonstrated a segmentation accuracy of 90% in Overall pixel accuracy and 84% in full intersection over union (FIoU). Polygon generation accuracy using the Hierarchical supervision (HiSup) algorithm was also confirmed to be 94%. This study highlights the potential of AI solutions in the field of topographic surveying, where the rapid and accurate generation of spatial information for large areas and numerous objects is essential.