|Ph.D Student||Shtain Zachi|
|Subject||Clustering of Laser-Scanning-Based Point Clouds for|
Geographic Feature Extraction
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Sagi Filin|
|Full Thesis text|
Application of laser scanning technology for mapping purposes has gained increased popularity because of the direct, accurate, and detailed surface characterization that these systems provide. Their broad area coverage, even by a single scan, and their independence of exterior energy sources allows them to operate with fewer limitations and at any hour of the day. While laser scanners provide dense and accurate information, the sheer data volume they acquire and the complexity of the irregularly 3-D distributed point clouds impede efficient information extraction and analysis. As the high-resolution translates also to great redundancy, it becomes desirable to develop simplification strategies that can reduce the volume and contribute to greater accessibility and to faster subsequent processing.
Point cloud simplification is empowered by the definition of similarity metrics which are aimed to identify homogeneous regions in the point-cloud. However, the variety of shapes and clutter in natural scenes, along with the significant resolution variations, occlusions, and noise, contribute to inconsistencies in the geometric properties, thereby making the homogeneity measurement challenging. Thus, the objectives of this thesis are to develop a point-cloud simplification model by means of data segmentation and to extract information in a better-suited way. The literature review shows that most approaches either apply volumetric data strategies and/or resort to simplified planar geometries, which relate to only part of the entities found within a natural scene. To provide a more general segmentation strategy, we propose a proximity-based approach that allows an efficient and reliable surface characterization with no limitation on the number or shape of the primitives which in turn, enables detecting free-form objects. To achieve this, a local, computationally efficient and scalable metric is developed, which captures resolution variation and allows for short processing time.
Our proposed segmentation scheme is demonstrated on several datasets, featuring a variety of surface types and characteristics. Experiments show high precision rate while exhibiting robustness to the varying resolution, texture, and the occlusions that exist within the sets.