Graph Neural Network based geometric modeling of terrain topology; its application to GPS denied navigation
Abstract
U.S. Army s mission of land dominance requires novel approaches to modeling the terrain for use in next generation applications, such as autonomous vehicles and modern intelligence gathering in asymmetric warfare. The existing terrain models are old, only encode the 3D structure and does not scale. In a sense, they represent a shrink-wrap base level representation of the surface, lacking topological relationships of surface features, textural information as well as semantic labels in the form of land-cover and land-use. The main objective of this proposal is to introduce a new geometric terrain model that extracts, learns, and compresses topological features for natural terrain that scales well and can be retrieved fast for applications requiring real-time retrieval. Without the loss of generality, the proposal showcases the applicability and usability of the developed model for positioning and navigation of autonomous ground systems on natural terrains. Compared to an urban terrain, natural terrains exhibit unorganized, complex, and unfamiliar content that makes it harder to develop terrain models, in particular, that could efficiently support real-time applications, such as autonomous vehicle navigation. In this proposal, in addition to terrain elevation, other related sensory information, including but not limited to semantic land- cover type and texture, is used to provide the developed model extended capabilities. The developed geometric terrain represents the encoded information in a subspace generated by graph neural networks. The graph neural networks provide topological modeling and operability using message passing mechanisms between the nodes. The message passing mechanisms encodes the terrain elevation and other relevant sensory information in the vicinity of each node and projects them onto a manifold. The nodes in the graph neural network have a six-neighbored hexagonal grid structure for increased spatial resolution. The graph neural network projects the hexagonal grids with similar terrain elevation and content to proximal locations on the manifold that allows compression of the geometric model. The compression model used in the proposal is based on clustering of the embedded feature space to find representative graph nodes as spatial models. Considering parts of natural terrain contain repeating variation in the elevation and semantic information, high compression rates can be achieved. The compressed geometric model is demonstrated for positioning and navigation of an autonomous ground system that will use the compressed model in context of location-based services in real-time. Traditionally terrain information is encoded as elevation, land-cover map and texture all geo registered but are separate from each other. The elevation is stored in the form of a digital terrain model generated from aerial lidar scans, texture is stored as satellite or aerial imagery and the landcover map is in the form of vector data. For three separate data sources, despite they are co- registered, accessing information requires significant time and is not suitable for real-time processing. The proposed geometric model advances these dated-models and provides a new approach to learning, analyzing, storing, and accessing geospatial data. The positioning and navigation application using the proposed geometric model emphasizes the importance of terrain modeling which was overlooked in most former studies.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 07, 2021
- Source ID
- W911NF2110356
Entities
People
- Alper Yılmaz
Organizations
- Army Contracting Command
- Ohio State University
- United States Army