Geometry-Aware Human Noise Removal from TLS Point Clouds via 2D Segmentation Projection

Area of Science:

• Geomatics Engineering
• Computer Vision
• 3D Data Processing

Background:

• Terrestrial Laser Scanning (TLS) generates large point clouds vital for digital twins and cultural heritage.
• Manual removal of human noise from TLS data is inefficient and time-consuming.
• Automated methods are needed to improve the processing of TLS point clouds.

Purpose of the Study:

• To develop a geometry-aware framework for automatic human noise removal in TLS point clouds.
• To validate the framework's effectiveness on real-world outdoor TLS datasets.
• To assess the method's performance against existing dynamic object removal techniques.

Main Methods:

• Utilized You Only Look Once (YOLO) v8 for 2D instance segmentation of humans.
• Projected 2D masks into 3D space and applied multi-stage geometric filtering (DBSCAN, PCA).
• Introduced 'geometric gating' to mitigate reprojection errors and background structures.

Main Results:

• Achieved high noise removal accuracy with Precision/Recall/IoU of 0.9502/0.9014/0.8607 (OMU) and 0.8912/0.9028/0.8132 (JM).
• Demonstrated stable performance on mobile mapping system (MMS) data without parameter recalibration.
• Showed competitive recall and reduced over-removal of static structures compared to DUFOMap and BeautyMap.

Conclusions:

• The proposed geometry-aware framework effectively automates human noise removal from TLS point clouds.
• The method offers practical efficiency for large-scale TLS data processing.
• The publicly released dataset supports further research in this domain.