Fine-grained Feature Extraction for Semantic Segmentation of Point Clouds for Civil Engineering Structures

Abstract: Semantic point cloud segmentation plays a role in creating geometric-semantic as-is models (GSMs), e.g., for planning and construction in existing contexts or creating digital twins. Deep learning architectures have proven useful in learning complex geometric patterns but focus on large context windows, neglecting fine-grained texture features by subsampling the input. In this contribution, we describe a feature extraction based on local point cloud neighborhoods and full-waveform terrestrial laser scanning data for material classification. Through sensitivity analysis, we show that this provides a useful basis for subsequent use in large-scale machine learning models.