An ontology defines common vocabulary for researchers who aim to share information in a specific domain. It includes both human and machine interpretable definitions of basic and complex concepts in the domain, as well as their hierarchy and the relations among them.
Some of the reasons behind the development of an ontology in the context of the BorderUAS project are:
- To decompose the domain knowledge and spell out its fundamentals as holistically as possible
- To expose hidden, high-level knowledge that is not explicitly provided by the deep learning algorithms
- To share common understanding of the structure of information among Law Enforcement Agency stakeholders or software agents
- To enable the reuse of the domain knowledge when extending the representation, in order to further capture more specific applications and tasks
- To make the domain assumptions distinct and eliminate ambiguity among the different users of the platform
- To separate the domain knowledge from the operational knowledge and bridge the gap between the technical terminology and the human comprehension
Figure 1 illustrates the four main representational primitives that constitute a semantic representation. As one can see, an ontology defines the concepts in a domain of discourse as Classes. A class represents a set of actual “things” that share common characteristics – not only physical things that people can touch. The members of each class are the most specific concepts encompassed in the knowledge base and are often called Instances or Individuals. The different types of relationships that exist between the individuals are expressed as Association Roles or Object Properties.
These roles are directed, meaning that they point from a subject to an object. The combination of a subject individual, an object individual and the association among them, results in a triple that expresses an assertation in the form of subject-predicate-object, as presented in figure 2. Finally, an ontology contains value slots assigned to each class, which govern its instances and describe their various features. These qualities are called Attributes or Data Properties.
In a domain such as border surveillance, concept definitions are often unclear, or they require special knowledge to interpret. In this context, BorderUAS ontology intents to structure the knowledge about the existence of things and possible events, aiming to align these concepts with the LEA’s terminology as much as possible, and cover all the scenarios introduced by the end-users.
The complexity of creating and using domain-specific ontologies can be reduced by dividing knowledge into multiple areas. Following the approach of modularization, that is dividing the ontology into individual modules for independent subject areas, we split the knowledge into 4 segments, that are:
- Core ontology: It contains basic abstract concepts such as physical objects, actions, context, enumeration values, events, groups of things, as well as the notions of location and time. These concepts are applicable to almost every knowledge-based representation.
- Domain ontology: It defines the central concepts of the border surveillance domain, by aligning the core concepts with the domain ones and extending the core ontology in terms of granularity. It includes both dynamically changing objects, such as vehicles, and static unchanging objects, such us buildings, activities and behaviours that involve the physical objects in active and passive roles and events that mark the changes in states of physical objects, critical points in time, and triggers or decisions for the start or end of activities.
- Analytics application ontology: It defines intermediate concepts which play a transitive role in the analysis of a scene or scenario from the surveillance media such as videos. It introduces concepts such as transitive actions, environmental and weather observations and low-level features such as bounding boxes.
- LEA application ontology: It contains the most specific concepts represented in the knowledge base. It involves the definition of specific physical objects and complex events from inference perspective.
Figure 3 visualises the architecture design of the BorderUAS ontology. The application ontologies use concepts from both domain ontology and core ontology, as indicated by the overlap of circles in the image.
Upon the conclusion of this post it is important to point out that many characteristics which one might want to declare between things, are too complex to be readily expressed using logical constructors exclusively. The combination of logical reasoning and rule-based reasoning enables a computer program to make complex inferences from a particular expression, thereby realising a high level of computer understanding. In this context, we aim to develop a complete set of reasoning rules for describing and semantically annotating complex events that might occur in the domain of border surveillance. This pre-defined set of semantic rules along with the appropriate ontology implementation tool, will provide the BorderUAS platform with the capability of notifying the end-users about the occurrence of specific events in real-time, through a data pipeline that ultimately ends with some sort of alert system, as portrayed in figure 4.