복셀 기반 3D U-Net을 활용한 대용량 포인트 클라우드의 시맨틱 세그멘테이션 기법 개발

Smart cities are emerging as a key driver of future urban development, and digital twins can maximize the efficiency of smart city design, operation, and maintenance. LiDAR (Light Detection and Ranging) is gaining attention as an effective device for acquiring essential 3D (Three Dimensional) point cloud data for digital twin creation. Semantic classification of raw data acquired from LiDAR is a crucial preprocessing step prior to digital twin generation. Previous studies have mainly focused on improving object-level classification accuracy by enhancing model architectures through fusion networks. While effective for accurately classifying small-scale data, these approaches have limitations in comprehensively learning and classifying data acquired from large-scale spaces. To overcome these challenges, this study proposes a custom semantic segmentation technique based on a voxelized 3D U-Net model. Unlike existing methods, the proposed approach does not require prior object-based segmentation of input data, and incorporates a lazy loading technique for efficient memory management when handling large-scale point cloud data. Experiments were conducted using various hyperparameter combinations on both the open S3DIS dataset and a simulated Plant dataset as training data. As a result, the proposed model achieved Overall F1-scores of 0.750 and 0.910 for 13 classes of the S3DIS dataset and 9 classes of the Plant dataset, respectively.