A Point-based Level-set Approach for the Extraction of 3D Entities from Point Clouds – Application in Geomorphological Context

Abstract

The commonly used level-set models have been proven efficient in delineating and detecting objects in a variety of image processing applications. Such models do not require training or labelling, need little prior information as to existing objects, and want little adaptation between scenes. Despite their great potential, their utility for point cloud processing has been limited to 2D space. In this paper, we introduce a novel level-set-based approach which extracts entities directly from the point cloud while retaining their three-dimensional point form. To do so, we adapt the level-set scheme to 2D smooth manifolds represented by unstructured points in 3D space. Surface derivatives are then computed using a local surface parametrization based on a weighted least squares approximation. This alleviates the need for a triangulated mesh and facilitates the level-set evolution within the point cloud. As a driving force, we utilise visual saliency to focus on pertinent regions. As the saliency estimation is performed pointwise, the proposed model is completely point-based, resulting in three-dimensional entities extracted by their original points. We apply the proposed method to extract geomorphological entities in two fundamentally different scenes. Such entities present a challenge to existing extraction schemes, as they are embedded within their surrounding and they do not conform to closed parametric forms. The proposed approach enables the detection of various entities simultaneously, without prior knowledge of the scene, and regardless of their position. This promotes flexibility of form and provides new ways to quantitatively describe morphological phenomena and characterise their shape.