A visual saliency-driven extraction framework of smoothly embedded entities in 3D point clouds of open terrain

Abstract

Three-dimensional documentation of natural and cultural geosites is gaining increasing attention as an indicative tool for environmental change. However, the entities therein pose a challenge to current extraction schemes due to their varying dimensions, complex shape, and most importantly, their seamless embedding in the surrounding topography. It is common to approach the extraction of these entities by developing landform-specific methods which are applied in a localized manner. Nonetheless, these methods hardly generalize, and different entities are dealt with independently, even when located at the same site. We propose in this paper a general, content-driven framework for the detection of smoothly embedded entities that, unlike prevalent approaches, seeks no specific form. We focus on salient entities, which attract visual attention within the point cloud and develop a new detection scheme, driven by homogeneity in the entities’ saliency. We show how such formulation requires no approximate location or starting points and does not suffer from weak responses. Therefore, our framework can be readily applied to a multitude entities in various scenes, regardless of type or acquisition technique. We demonstrate our solution on airborne and terrestrial laser scans and detect entities of different types and shapes that feature in both natural and culturally-important sites. As we show, the proposed framework yields improved results compared to state-of-the-art counterparts.