Semantic Localization for IoT
371
Types, Toponyms/Placenames, (Geo) Spatial Relationships, Coordinate Reference
Systems, Geospatial Metadata, and (Geo) Web Services. The relationships between
these, however, are highly unstructured and lacking in formal properties that can
exploited algorithmically. More recently, geospatial ontologies like GeoDataOnt [15]
have been developed to provide a unified ontology for this domain.
A popular (non-RDF) spatial ontology today is codified in a JSON schema called
GeoJSON [16]. This is used by many location based services. In contrast to the
semantic web, GeoJSON is good at representing geometries, but not higher level
ontological concepts and relationships. It supports points, lines, polygons, and collections of polygons in 2D or 3D. Given the extensive support for GeoJSON in existing
apps and software, it is a useful standard to leverage for geometric concepts. But
restricting spatial ontologies to exclusively geometric concepts is a mistake. Spatial
relationships are more complex.
On the opposite end of the complexity spectrum, the Open GIS Geography Markup
Language (GML) Encoding Standard [17] is a 437 page specification document
describing an XML schema for spatio-temporal ontologies. It follows the ISO 19101
definition of a feature as an “abstraction for real world phenomena” and represents the
world as a collection of features defined as name, type, value triples. The increased
complexity allows for the description of more sophisticated data such as spatial
geometries, spatial topologies, time, coverages, and observations. The format can be
extended to application schema such as IndoorGML [18] which is targeted for indoor
navigation. IndoorGML focuses on layered graph representations of relationships
such as adjacency and paths between semantic objects in indoor space. It models the
world as a collection of cells representing geometry and topology via the Poincaré
duality to achieve a “Multi-Layered Representation” of a given space in different
contexts.
A variety of geometric data structures and algorithms are employed in the field
of computational geometry when high performance is desired for computationally
difficult spatial analysis [19]. For example, a doubly-connected edge list is used for
the thematic map overlay problem, in which the overlay of spatial subdivisions is
computed.
1 A trapezoidal map is another geometric data structure employed to solve
point location queries: given the coordinates of a point and a map subdividing the
plane into regions, determine which region contains the point.
Point clouds are another computationally useful format for spatial information
in the domain of computer vision. Visually oriented sensors such as stereo cameras
or time of flight cameras (e.g. Light Detection and Ranging, LiDAR) measure the
location of individual 3-dimensional points in the world. These points represent
sampled measurements of real-world objects. Once collected, software such as the
open source point cloud library [20] can use a point cloud data set to reconstruct a
sampled surface or perform segmentation to semantically identify objects.
The diversity of these standards for spatial representation is daunting. Yet it is easy
to see how applications in the IoT with different purposes for spatial information and
different sensors for collecting that data benefit from different data representations.
1 Imagine overlaying two circles to form a venn-diagram, but with polygons instead of circles.
371
Types, Toponyms/Placenames, (Geo) Spatial Relationships, Coordinate Reference
Systems, Geospatial Metadata, and (Geo) Web Services. The relationships between
these, however, are highly unstructured and lacking in formal properties that can
exploited algorithmically. More recently, geospatial ontologies like GeoDataOnt [15]
have been developed to provide a unified ontology for this domain.
A popular (non-RDF) spatial ontology today is codified in a JSON schema called
GeoJSON [16]. This is used by many location based services. In contrast to the
semantic web, GeoJSON is good at representing geometries, but not higher level
ontological concepts and relationships. It supports points, lines, polygons, and collections of polygons in 2D or 3D. Given the extensive support for GeoJSON in existing
apps and software, it is a useful standard to leverage for geometric concepts. But
restricting spatial ontologies to exclusively geometric concepts is a mistake. Spatial
relationships are more complex.
On the opposite end of the complexity spectrum, the Open GIS Geography Markup
Language (GML) Encoding Standard [17] is a 437 page specification document
describing an XML schema for spatio-temporal ontologies. It follows the ISO 19101
definition of a feature as an “abstraction for real world phenomena” and represents the
world as a collection of features defined as name, type, value triples. The increased
complexity allows for the description of more sophisticated data such as spatial
geometries, spatial topologies, time, coverages, and observations. The format can be
extended to application schema such as IndoorGML [18] which is targeted for indoor
navigation. IndoorGML focuses on layered graph representations of relationships
such as adjacency and paths between semantic objects in indoor space. It models the
world as a collection of cells representing geometry and topology via the Poincaré
duality to achieve a “Multi-Layered Representation” of a given space in different
contexts.
A variety of geometric data structures and algorithms are employed in the field
of computational geometry when high performance is desired for computationally
difficult spatial analysis [19]. For example, a doubly-connected edge list is used for
the thematic map overlay problem, in which the overlay of spatial subdivisions is
computed.
1 A trapezoidal map is another geometric data structure employed to solve
point location queries: given the coordinates of a point and a map subdividing the
plane into regions, determine which region contains the point.
Point clouds are another computationally useful format for spatial information
in the domain of computer vision. Visually oriented sensors such as stereo cameras
or time of flight cameras (e.g. Light Detection and Ranging, LiDAR) measure the
location of individual 3-dimensional points in the world. These points represent
sampled measurements of real-world objects. Once collected, software such as the
open source point cloud library [20] can use a point cloud data set to reconstruct a
sampled surface or perform segmentation to semantically identify objects.
The diversity of these standards for spatial representation is daunting. Yet it is easy
to see how applications in the IoT with different purposes for spatial information and
different sensors for collecting that data benefit from different data representations.
1 Imagine overlaying two circles to form a venn-diagram, but with polygons instead of circles.