Widgetized Section

Go to Admin » Appearance » Widgets » and move Gabfire Widget: Social into that MastheadOverlay zone

Sensor Urban Environments can prevent and fight robberies, terrorism attacks, natural catastrophes and other emergencies.

Keynote Speaker Charalampos Doulaverakis Eng MSc after his speech in Athens, met with our Publisher Victor Bjorgan.

This document was published by the criminology scientists Charalampos Doulaverakis, Nikolaos Konstantinou, Thomas Knape, Ioannis Kompatsiaris and John Soldatos.

Doulaverakis was the keynote speaker at the EISIC 2011 SecuAll scientists are related to three institutions: a)  Informatics and Telematics Institute, Centre for Research and Technology Hellas  in Thessaloniki, Greece  b) the Autonomic and Grid Computing Laboratory, Athens Information Technology, Athens, Greece

c) Data Fusion International, Dublin, Ireland.
Abstract—This paper introduces a novel sensor information
fusion system enabling security and surveillance in large scale
sensor saturated urban environments. The system is built over
state-of-the art sensor networks middleware and provides information
fusion at multiple layers. A distinguishing characteristic
of the system is that it support seamless integration with semantic
web middleware (including ontologies and inference mechanisms),
which enable intelligent high-level accurate reasoning.
This is a key functionality for efficient surveillance in large scale
environment, where manual inspection of individual tracking
systems becomes extremely resourceful and overall impractical.
A proof-of-concept implementation of the system manifests its
benefits and technical challenges, while also outlining lessons
The proliferation of ubiquitous computing and the parallel
decrease of the price of sensors has recently allowed sensor
applications to make their appearance in a variety of domains.
Especially, sensor applications with emphasis on large scale
deployments in urban environments and the emerging wave of
participatory sensing applications such as WikiCity [1], City-
Sense [2], and Google Latitude [3] has offered fundamental
changes and advancements both in the academia as well as in
the heart of the society itself. Nowadays, more than ever, we
are witnessing the materialization of visions and concepts such
as the Internet-of-Things (IoT) [4] and M2M communications
Recent incidents have manifested that modern cities are
very susceptible to terrorist attacks. For instance, the collapse
of New York’s Twin Towers on 11th September 2001, the
bombing of packed commuter trains in Madrid on 11th March
2004, the London bombings in July 2005, or the Moscow
metro in March 2010 demonstrate that prominent applications
need to be devised in order to serve critical surveillance and
security needs in urban environments.
However, numerous challenges are associated with such
efforts. The large scale nature of both the geographically
dispersed environment as well as of the volume of the data,
the multiple distributed heterogeneous components that need
to be assembled, spanning sensor, sensor processing, signal
processing (including A/V) components, often from multiple
vendors are some of the issues that need to be addressed. Additionally,
the need for automation, since manual observation
of multiple camera feeds is not possible, and the inclusion of
high-level intelligent reasoning for event inference are features
without which the added value of the system is limited.
In order to tackle these issues in global sensor networks, various
frameworks have been developed, offering programmable
and configurable solutions, e.g. [6]–[8]. Among the most
important in these frameworks is the limited support for perceptual
processing components – with the exception of some
more heavyweight frameworks [9] – and limited support for
semantics embodiment, inference, and high-level reasoning.
Employment of Semantic Web technologies in sensor networks
research focuses mainly on sensor modelling in order
to enable higher level processing for event/situation analysis.
Initiatives such as [10] use SUMO as their core ontology under
which they define a sensor ontology for annotating the sensors
and it can be then queried for sensor discovery. Other frameworks,
such as the one proposed in [11], define ontologies for
sensor measurement and sensor description. The framework
emerges from the W3C Semantic Sensor Network XG1 which
aims at providing ontologies and semantic annotations that
define capabilities of sensors and sensor networks.
Ontologies have been proposed for situation awareness
(SAW) in sensor fusion applications where their ability to
model a domain or a “part of the world” is utilized. Several
approaches have been proposed as in [12] where the authors
define a core ontology for SAW which can be used as a basis
from which to build separate ontologies for arbitrary situations
that are able to express objects, relations and their evolution
over time.
In [13], we  the authors used an ontology to create a unified
expression of the Situation Theory [14], [15]. The Situation
Theory Ontology (STO) is expressed in OWL which enables
situations to be described using a formal language, thus
allowing inference through a reasoning engine or by using
appropriate rules.
A framework for the designing of ontology for SAW is
presented in [16] where a six step guide to design ontologies

based on the Basic Formal Ontology (BFO) is presented. The
authors test their framework in a situation assessment in a postdisaster
environment context and claim that their BFO-based
ontology was able to capture the complexities of the testing
context and provide adequate inferential capabilities for higher
level fusion.
All the above methods do not propose an architecture which
will integrate sensors and low level processing modules with
higher level fusion process for situation awareness. In order
to address this limitation, a solution is presented in this paper
∙ Comprises a multi-level fusion system, at all JDL (see
Section II-B) levels
∙ Seamlessly blends ontologies with low-level information
∙ Combines semantic web middleware with sensor networks
The structure of the paper is as follows: Section II analyses
the architectural approach that is followed, Section III demonstrates
the Low Level Fusion (LLF) capabilities, Section IV
the semantic and reasoning capabilities, Section V presents
implementation details and example use cases, while Section
VI concludes the paper by presenting out comments and
In order to test our approach, a suitable environment for
setting up and deploying the proposed sensor analysis architecture
had to be selected. Security surveillance environments
offer an ideal set up in which the main characteristics that
make it distinguishing are:
∙ Usually security surveillance areas are sensor saturated
environments with electro-optical (visual and IR cameras)
and acoustic sensors being more common while
others like temperature or RFID, etc sensors can also
be found. The abundance of available sensors makes it
more possible to capture the various events that take
place in a surveillance session. However, the higher the
number of sensors, the more difficult it is to manage and
observe them. Additionally, it makes it harder to filter out
information that is irrelevant or discover information that
could potentially be useful.
∙ Surveillance environments are deployed in areas that are
densely populated in terms of people but also in assets,
e.g. buildings or vehicles, hence many events that could
be of interest are taking place. This fact, coupled with
the higher chance of these events to be captured by the
sensors (the previous characteristic), makes discovery of
important events difficult.
∙ Multiple processing algorithms and context-acquisition
components, which are used for extracting information
and detecting events, require a method for managing the
data that they produce.
∙ Due to the variety of sensor modalities that are deployed
there is a large degree of sensor and data heterogeneity

Fig. 1. Architecture of the proposed system
that need to be tackled by the sensor management and
analysis system.
Public security in urban areas can be a challenging task to
security personnel. Threats posed to citizens include crimes
such as robberies, terrorism attacks, natural catastrophes and
other emergencies. Being aware of threatening situations either
in the moment or being able to forecast events from security
monitoring applications enables security personnel to take the
right actions in context. Fusion of data from appropriate sensor
types supports inferring threatening situations in the context of
situation specific parameters. An example of such a deployed
sensor network is the case of urban surveillance in an urban
war environment where data feeds from multiple sensors, e.g.
visible spectrum and IR cameras, acoustic sensors, seismic
sensors, etc, are directed to a command center and critical
decisions have to be made in short time regarding the threat
level in a situation.
Taking into account the above facts, a method is required
that will enable efficient information processing and management.
Semantic web and ontologies can efficiently handle
heterogeneous information through semantic description of
knowledge. Additionally, ontologies are used to model domain
knowledge through class definitions and relations between
classes. The knowledge model and the underlying data can
then be used for supporting reasoning services in order to infer
new knowledge that is not explicitly stated. These features
can be utilized in a sensor network in order to provide the
backbone for intelligent sensor fusion.
A. System architecture
Figure 1 illustrates the approach followed in the hereby
presented work. At the bottom layer are the sensors, that are
translating their perception of the world into raw sensor data.
This data is processed by the leaf nodes, that are operating
at the LLF layer which adds structure to the data. The nodes
are named “Leaf” nodes because, if the architecture can be
considered as tree-like, then these nodes can be considered
as leaves, in the sense that they do not have children nodes
of inferior capabilities or at a lower level. When deemed
necessary, signal processing components can be hosted in
dedicated hardware since, specifically in A/V processing, algorithms
may be extremely resource-hungry. At the high-level
fusion (HLF) layer, intelligence is added to the system, by

mapping the collected data into ontology concepts, achieving
thus uniform information representation throughout the system
and enabling reasoning.
The Central Control layer that is logically on top of the
infrastructure offers monitoring and control capabilities. In
essence, it comprises:
∙ A database (DB) where high-level information such as
events and threats is stored
∙ An Environmental Service (ES) that enables geospatial
∙ The Semantic Fusion component that performs analysis
on the system-wide high-level collected information in
order to infer events and potential risks and threats
∙ The Common Operational Picture (COP) that offers a
visualization of the system state with regard to deployed
sensors, detected objects, sensed events and inferred
In the following sections, we analyse how the aforementioned
layers and components function and interoperate in
order to achieve intelligent information fusion in an urban
B. JDL model for sensor fusion
In order to improve communications among military researchers
and system developers, the Joint Directors of Laboratories
(JDL) Data Fusion Working Group began an effort
to define the terminology related to data fusion. The result
of that effort was the creation, in 1986, of a process model
for data fusion and a data fusion lexicon [17]. The JDL
process model is a paper model of data fusion and is intended
to be very general and useful across multiple application
areas. The JDL data fusion process model is a conceptual
model which identifies the processes, functions, categories of
techniques, and specific techniques applicable to data fusion
(Figure 2). According to the model data fusion process is
conceptualized by sensor inputs, human-computer interaction,
database management, source preprocessing, and four key
1) Level 1 (Object Refinement): is aimed at combining
sensor data together to obtain a reliable estimation of
an entity position, velocity, attributes, and identity;
2) Level 2 (Situation Refinement): dynamically attempts
to develop a description of current relationships among
entities and events in the context of their environment;
3) Level 3 (Threat Refinement): projects the current situation
into the future to draw inferences about enemy
threats, friendly and enemy vulnerabilities, and opportunities
for operations;
4) Level 4 (Process Refinement): is a meta-process which
monitors the overall data fusion process to assess and
improve the real-time system performance.
Revisions of the model suggest the addition of 2 more
levels: A lower Level 0 (Source preprocessing/Sub-object
refinement) which aims to estimate the state of sub-object
entities such as signals and features, and a higher Level 5

(Cognitive or User Refinement) which introduces man in the

fusion loop where the aim is to generate fusion information
according to the needs of the system user.

Fig. 2. The JDL model levels and how they are related to the proposed
system’s layers and nodes
The current paper focuses in the definition of a sensor fusion
architecture which addresses matters that correspond to the
higher JDL levels 2, 3 and 4, i.e. situation, threat assessment
and process refinement respectively, as is depicted in Figure
2. Ontologies, with their inherent ability to model relations,
can be employed to provide enough specificity in describing
the higher level concepts of JDL but also to describe the
relations between these concepts [18]. The ontology provides
the standardized form in which situations are defined so
that appropriate algorithms for situation assessment can be
In order to perform Low-Level Fusion, our work relies on
the GSN2 middleware. The middleware constitutes an opensource,
Java-based implementation. It was designed in order
to allow processing from a large number of sensors and as
such, it covers the functionality requirements of low level
fusion in sensor data streams. In order to acquire information,
GSN introduces the concept of “virtual sensor”. Any data
provider, not only sensors can provide data to a GSN instance,
as long as a virtual sensor configuration file (in the form of
an XML) defines the the processing class, the sliding window
size, the datasource and the output fields. GSN Servers can
communicate between them, thus forming a network where
information is collected, communicated, fused and integrated
in order to produce the desired results.
As far as it concerns data acquisition, each GSN node can
support input from more than one data stream. In order to
combine the information, an SQL-like procedural language is
offered. This allows to the user to define LLF functions in the
following manner, while GSN takes care behind the scenes
about crucial issues such as thread safety, synchronization,
SELECT source1.S1 AS NumberOfPersons, source2.S2
AS ExistenceOfSmoke
FROM source1, source2
WHERE NumberOfPersons>0 AND ExistenceOfSmoke=”true”
In the example above, there are two data providers for one
GSN Server.

This virtual sensor definition will produce events only in
the case when the WHERE condition is satisfied. The example
above demonstrates the concept of fusion: results are provided
by taking into account the inputs from both the sensors.
Using the abovementioned approach, a variety of signal
processing components can be integrated into one unifying
architecture. For instance, the example above can be supported
using a Body Tracker [19], [20] and a Smoke Detector, in order
to produce alerts when persons are detected near smoke. In the
same manner, results by processing components such as face
detectors [21], unusual event detectors [22] can be fused in
order to achieve the desired system functionality.
Communication with these processing components constitutes
a difficult to tackle problem, mostly in terms of
implementation since:
∙ Two worlds have to be brought together: image processing
and the distributed systems. The former typically processes
images or videos that are fundamentally different
in nature from streaming data. As such, the component’s
input has to be modified in order to take into account
streams that may lack synchronization, be erroneous, or
overwhelming the network with data.
∙ Integration is hardly a trivial task. Prototypes in Matlab,
algorithms implemented in C++ will need to communicate
with the java-based GSN servers, running either
locally or remotely. Web services/socket interfaces can
be developed, imposing though, additional processing
Next, in order to allow GSN-GSN node communication, the
information produced can be forwarded to other GSN nodes
either in push or in pull mode. Data streams can be forwarded
in order to be processed at remote GSN nodes. In order to
forward these data streams, GSN supports subscriptions, onceoff
queries or even simple data forwarders. Data is forwarded
using a RESTful approach (remote wrappers return XML over
The Higher Level Fusion (HLF) layer of the proposed
architecture consists of the underlying ontology which will
be used as the backbone for performing the reasoning that
is required for situation assessment. OpenLink Virtuoso3 has
strong support for Semantic Web technologies and provides
services like RDF triple storage, a SPARQL compiler, rulebased
reasoning and can expose relational data as RDF.
Virtuoso acts as a federation layer between the relational and
semantic data thus enabling their seamless integration. A block
diagram of the HLF approach is presented in Figure 3 (Virtuoso Universal Server: http://virtuoso.openlinksw.com/).

Fig. 3. The High Level Fusion process

A. Ontology for high level fusion
The ontology is the main information gathering point in the
proposed architecture. It is defined according to the domain
where the sensor network is placed and it defines the entities
that are taking place in a situation, the events and the relations

between them. All data that are produced by the processing
modules and by the LLF nodes are eventually stored here
where reasoning is performed in order to infer new knowledge
which corresponds to events and situation assessment that
cannot be performed at the LLF level.
In order to take advantage of research that has been conducted
in the area of semantic enabled situation awareness, the
current system uses as its core ontology the Situation Theory
Ontology (STO) [13] which has been developed specifically
for that purpose and is based on Situation Theory. In short
STO is written in OWL and models the events/objects and
their relationships in a way that can be extended using either
OWL axioms and properties or in combinations with rules
for supporting complex cause-effect relations that cannot be
expressed in OWL alone. The main classes of STO are
displayed in Figure 4.


Fig. 4. Situation Theory Ontology with its main classes and properties
Situation is the central class. Instances of this class are
specific situations. The second class is the Individual class,
which is a counterpart of the individuals in situation theory.
Similarly, Relation captures the n-ary relations. Attribute is
a generalization of locations and time instants in situation
theory. Instances of this class are attributes of individuals
and situations. An attribute may have a dimension associated
with it so the class Dimensionality represents this fact. The
Polarity class has only two instances that correspond to the
two possible values associated with a tuple, either that a
given tuple holds or that it does not hold.

Classes of STO are related through a number of OWL properties.

Situations are linked with four kinds of entities. First, the property
relevantIndividual captures the individuals that participate in a
situation. The property relevantRelation is used to assert that
a given kind of relation is relevant to a given situation. Since
situations are objects, they can have attributes of their own.
The STO is extended with classes and relations that correspond
to the actual application scenario. In order to be able
to use it in a real sensor fusion environment two additional
ontologies are integrated. These are the “Time ontology”

(Time Ontology in OWL: http://www.w3.org/TR/owl-time/

which holds the timestamp of any concept instance that
is stored during runtime and the “WGS84 Geo Positioning
ontology” (Basic Geo (WGS84 lat/long) Vocabulary: http://www.w3.org/2003/01/geo/5)

which holds the latitude and longitude values of
B. Mapping relational data to RDF
As explained in Section III, the data that are produced by
the processing modules and by the LLF nodes are stored in the
relational database that backs GSN. In order to forward these
data to the ontology they have to be translated into semantic
notations. This task is performed by the mapping layer which
maps the relational schema to the semantic schema. There
are two strategies for accomplishing this transformation, using
either a push or a pull method.
The push method forwards the data to the ontology using
semantic notation as soon as they are generated. This has to be
implemented in the lower level layer with the semantic layer
having a passive role in the process. The advantage of this
method is that the transformations are executed fast and the
ontology is always up-to-date. The disadvantages is that each
lower level node will have to implement its own push method
while there is the risk that the ontology will be populated with
data even when no query is sent to the semantic layer.
The pull method transforms the relational data to semantic
on request, i.e. during query time. Virtuoso supports this
functionality through “RDF Views” where mappings, simple
or complex, between relational database tables and ontology
concepts and properties are defined. During query time, the
mapping process is triggered and data are transformed on
the fly. The advantages of the pull method is that the actual
mapping is defined at the higher semantic level rather than the
lower levels and that data are transformed on request so that
the ontology will accumulate instances that are needed for the
actual query evaluation. The disadvantage of this method is
that it could lead to longer response times during queries.
In the proposed system, the pull method which utilizes
Virtuoso’s “RDF Views” is used as it is best to separate the
semantic layer from lower level processing. This will make
extension of the system by adding more sensors and sensor
processing modules easier without having to mix the lower
level processing level with the high level semantic fusion level.
Additional information can be integrated by the HLF layer
in order to derive situations.

Information coming from sources like environmental services which could provide details about
the geographical locations of specific structures or points can
be queried, as long as they expose I/O interfaces, and these
data can be used for inferencing.
C. Reasoning

Reasoning uses the ontology structure and the stored instances
in order to draw conclusions about the ongoing situations.
Under this approach relations between classes like
rdfs:subClassOf, properties like owl:sameAs and relations
that are defined by experts in the ontology are utilized
for inference. The knowledge base can further be extended by
using rules to describe situations/events that are too complex
to be defined using OWL notation only.
For utilizing the above, two different reasoners are integrated
in the system, Virtuoso’s internal reasoning engine
and the Jena Semantic Web Framework (The Jena Semantic Web Framework: http://jena.sourceforge.net/.)

Virtuoso provides inference capabilities using OWL only thus does not provide
support for an external rule set. However, Virtuoso implements
a powerful geospatial inference mechanism which is a crucial
feature when dealing with sensor networks and sensor fusion.
On the other hand, Jena supports external rules, advanced
reasoning capabilities and seamless integration with Virtuoso
where in this case, Virtuoso is treated only as a triple store.
Both of them can be employed simultaneously on the same
dataset and their activation is managed programmatically by
the application.
Another issue that had to be dealt with is that sensor analysis
modules and the LLF processes can generate a significant
amount of data over time, enough to make the reasoning
service not responsive. To overcome this problem, the solution
that is proposed is to use a time window where reasoning is
performed using only facts in a specific time interval. Jena
and Virtuoso can support this type of inference by assigning
a RDF triples to a specific “context” type, thus dealing with
RDF quads.
The reasoning service can also use information and data
from external services in order to use them in the inference
process. In the proposed architecture, an Environmental
Service provides the location information of sensors, events,
waypoints and critical or important landmarks which are
subsequently used for geospatial reasoning.
It should be noted here that in a real world application the
different information sources that provide data to the proposed
fusion architecture could make use of different ontologies
for their purposes. In order to enable a seamless integration
of these sources an ontology mapping process between the
information source ontology and the HLF ontology would have
to be defined.
Implementation is based mostly on two types of GSN-based
servers that communicate with the sensor world, with a central
node that monitors and controls the network. As illustrated

in Figure 1, the Leaf Nodes are the ones that consume the

sensory information. These ones are in fact GSN-based servers
configured to process incoming data from the sensor layer.
In order to integrate Signal Processing components as in
Figure 5, our implementation allows the use of standalone
hosts. This happens because the components can either be
resource-hungry in terms of processing capabilities to the
extend that they utilize the whole processing power for their
efficient operation or simply because the implementation is
not portable, as in linux/windows-specific libraries.
A. System Implementation
In the scope of a proof-of-concept implementation and
also as a testbed for our experiments, the snapshot of the
architecture above that was implemented employs 1 computer
node that hosts a Leaf Node and two processing components:
a Smoke Detector and a Body Tracker, a camera that streams
its feed using RTP, a node that hosts a Semantic Node that
offers HLF capabilities and, finally, the Central Node that has
the overall system supervision as already depicted in Figure
Bottom-up, the system can be described as follows:
First, the camera generates an RTP feed with its perception.
The feed is processed by both the signal processing
components (the Body tracker and the Smoke detector).
The first component generates a stream containing at
all times the NumberOfPersons detected by the component.
The second one, simple however, is not as straightforward:
it splits the image into particles and reports the
NumberOfSmokeParticles detected. Then, the LLF Virtual
sensor, according to the example presented in Section
III fuses the data. Figure 5, illustrates the behavior of the


Fig. 5. Body Tracker Sequence Diagram

The sequence is initiated by the Body Tracker wrapper,
which is polling for results the Body Tracker component
host at fixed time intervals. The host, when it receives
the POLL command – note that the component may
support a number of functions such as START, STOP,
REQUEST_VIDEO_DETAILS etc. – it returns a string containing
a description in XML of the number of persons tracked.
This description is of the following form:

<person n=”1″>
<person n=”2″>


Note, next, that sampling the video source takes place
asynchronously. This happens because the camera will be
streaming at 25 fps, while for the needs of the LLF, 500
ms may suffice between two consecutive polls. In addition,
tracking a person is a more demanding task than detecting
it, since the former implies comparing consecutive frames for
differences between them while for the latter processing single
frames is enough. Therefore, the asynchronous behaviour is
explained from the fact that the camera fps is not aligned
with the messages per second that the component produces.
Under any circumstances, the LLF Wrapper receives a
notification from the BodyTracker. The same information flow
applies to the Smoke detection component. The LLF Wrapper
generates results only in the case when the fusion conditions
are true. Of course, more processing components and more
complex fusion conditions can be added in the processing
scheme hereby described.
For the HLF, the major difference is in the introduction of
an ontology and reasoning procedures where more complex
conditions for fusion can be applied. HLF only deals with
higher level detections and events meaning that it communicates
directly with the LLF and individual perception modules,
pulling data in order to make inferences. An example situation
where an alarm would be triggered by a rule, expressed either
as a Jena rule or as a SPARQL construct query, which would
state ”If smoke is detected near an object then raise an alarm”
with objects being a Person or Vehicle. The classes that are
inserted are
Smoke rdfs:subClassOf Non-RigidObject
Person rdfs:subClassOf RigidObject
Vehicle rdfs:subClassOf RigidObject
Alarms can then be retrieved with a simple SPARQL query
issued to Virtuoso as
SELECT ?x WHERE {?x rdf:type STO:Alarm}
This example illustrates the use of the subclass inference in

order to define situations that hold for a class of objects
which is something that cannot be defined using the LLF
layer only. In a similar way, the owl:sameAs property can
be used for reasoning. The property can be evaluated during
runtime. For example if two distinct detection instances is
decided that are actually the same entity, e.g. if they have
the same geographical locations, then by inference through
owl:sameAs the rules and properties that apply to one also
apply to the other thus enhancing the reasoning capabilities of
the system.
B. Advanced inference
In order to demonstrate the inference capabilities of the
proposed architecture, a scenario was set-up which makes
full use of the data chain from low level sensor data to high
level complex event detection and situation awareness/threat
assessment. The example demonstrates how external services
can be utilized in order to detect “critical” situations. For
setting up a realistic usage scenario, the Environmental Service
which exposes an interface where all interest points, e.g.
buildings or sensors, positions in absolute lat/long coordinates
are registered, is employed. These locations can be queried
in real time by the HLF layer and are used for inferencing
through geospatial reasoning. These data along with the lower
level processing modules are used by the HLF layer in order
to derive hypothesis about the criticality of a situation.
To demonstrate the use of sensors typically deployed in
urban environment we selected video cameras which can be
used to detect persons and also incidents, such as smoke, in
an area of interest. Sensor locations are associated with a
priori knowledge during reasoning. Detected smoke at a petrol
station marks a significant threat to public safety based on
a priori knowledge about the location and can be associated
with a clear emergency action plan. A public square with a
highly varying number of persons and smoke allows inferring
various situations which can be controlled as necessary based
on a security agenda. A specific public area can be subject
to an event schedule which can provide significant semantic
input to inferring on the threat level associated with detected
situations. A gathering of locals celebrating with a bond fire
on a public square is different from an unplanned detected
smoke event during mid-day hours.
1) “Smoke detection in critical location” situation: Reasoning
over triples in Virtuoso requires that the underlying
triple store was populated with inferred event data from the
smoke detector and body tracker. RDF views have been
developed and used to load the underlying data from the
sensor inference database. Using the scheduler component in
Virtuoso we trigger a procedural logic which updates the triple
store via RDF views from the sensor inference database and
subsequently invoke on the reasoning.
In the example of smoke detection at the petrol station
the relevant smoke event together with location data of the
video camera is forwarded to emergency personnel with the
relevant information. The reasoning process associates the
smoke detection event with a criticality factor according to

events modelled in STO and information coming from the
ES. The process of modelling events in STO is described in
detail in [23]. Figure 6 illustrates the model with instance
data by example. We used the STO:FocalSituation class
to mark significant situations that prompt action by security
personnel. The class interlinks the data associated with events
which we query via SPARQL and forward to emergency via
a Web service call.

Fig. 6. Modelling of the “significant situations” in STO

The query of Figure 7a (Fig. 7. a) Query for retrieving focal situations and 7b) the results of the query
run against the modelled “significant situation ) retrieves focal situations with their
event related details such as time location and further textual
descriptions and was run against the model above with the
results shown in Figure 7b. Predefined SPARQL queries are
invoked on in Virtuoso/PL.
Integration with emergency departments can be achieved
by sending details on significant threats found in SPARQL
result sets to a Web service endpoint exposed in an emergency
SELECT ?focalsituation ?sentTxt ?timeTxt ?locTxt
?focalsituation STO:focalRelation ?event .
?event STO:hasAttribute ?time .
?event STO:hasAttribute ?location .
?time rdf:type STO:Time .
?location rdf:type STO:Location .
?sentence rdf:type STO:Sentence .
?location rdfs:label ?locTxt .
?time time:inXSDDateTime ?timeTxt .
?sentence rdfs:label ?sentTxt



In this paper we have presented a framework for implementing
intelligent information fusion in a sensor network
environment. The framework deals with all aspects above the
sensor layer i.e. it deals with perception modules integration,
communication of the perception modules with GSN, low
level fusion, high level fusion with integration of semantic
description of information, communication with external services,
situation assessment and alert generation. It is a generic
framework that can be applied to any sensor network and
it is not restricted to the security and surveillance area that
was demonstrated. Illustrative examples of how the framework
would be deployed in realistic scenarios have also been
demonstrated where inference capabilities are shown. The core
advantages of the proposed framework are it’s extensibility
with pluggable perception modules integration and the ease of
defining rules for LLF and HLF, using SQL-like syntax and
semantic rules/queries respectively.
Future work will be focused on implementing and integrating
probabilistic reasoning, probably through fuzzy-DL,
in order to drive inference as opposed to the deterministic
reasoning that is applied up to now, thus enabling resolution
of issues such as conflicts or missing detections in situation

Part of this work has been carried out in the scope of the
MEDUSA project (Multi sEensor Data fusion grid for Urban
Situational Awareness, co-funded by the European Defense
Agency (MDS-MoM-A026RTGC-C-0003-VCS). The authors
acknowledge help and contributions from all partners of the
[1] F. Calabrese, K. Kloeckl, and C. Ratti, “Wikicity: Real-time locationsensitive
tools for the city,” IEEE Pervasive Computing, July-September
[2] R. Murty, A. Gosain, M. Tierney, A. Brody, A. Fahad, J. Bers, and
M. Welsh, “Harvard university technical report,” Tech. Rep. TR-13-07,
September 2007. [Online]. Available: http://www.citysense.net/
[3] X. Page and A. Kobsa, “Navigating the social terrain with google
latitude,” in iConference, Urbana-Champaign, IL, 2010, pp. 174–178.
[4] H. Sundmaeker, P. Guillemin, P. Friess, and S. Woelffl, Vision and
Challenges for Realising the Internet of Things. European Union, March
[5] G. Lawton, “Machine-to-machine technology gears up for growth,” IEEE
Computer, vol. 37, no. 9, pp. 12–15, September 2004.
[6] K. Aberer, M. Hauswirth, and A. Salehi, “The global sensor networks
middleware for efficient and flexible deployment and interconnection of
sensor networks,” Tech. Rep. 006, 2006.

[7] C. Sharp, E. Brewer, and D. Culler, “Hood: A neighbourhood sensor
networks,” in Proc. of MobiSYS’04, 2004.
[8] I. Chatzigiannakis, G. Mylonas, and S. Nikoletseas, “50 ways to build
your application: A survey of middleware and systems for wireless sensor
networks,” in Proc. of IEEE Conference on Emerging Technologies
and Factory Automation, ETFA, 2007.
[9] N. Dimakis, J. Soldatos, L. Polymenakos, P. Fleury, J. Curin, and
J. Kleindienst, “Integrated development of context-aware applications
in smart spaces,” IEEE Pervasive Computing, vol. 7, no. 4, pp. 71–79,
[10] M. Eid, R. Liscano, and A. E. Saddik, “A universal ontology for sensor
networks data,” in IEEE International Conference on Computational
Intelligence for Measurement Systems and Applications, Ostuni, Italy,
[11] H. Neuhaus and M. Compton, “The semantic sensor network ontology:
A generic language to describe sensor assets,” in AGILE International
Conference on Geographic Information Science, Hannover, Germany,
[12] C. J. Matheus, M. M. Kokar, and K. Baclawski, “A core ontology
for situation awareness,” in Proceedings of the Sixth International
Conference of Information Fusion, 2003, pp. 545–552.
[13] M. M. Kokar, C. J. Matheus, and K. Baclawski, “Ontology-based
situation awareness,” Information Fusion, Special Issue on High-level
Information Fusion and Situation Awareness, vol. 10, no. 1, pp. 83–98,
January 2009.
[14] J. Barwise, “Scenes and other situations,” Journal of Philosophy, vol. 78,
no. 7, pp. 369–397, 1981.
[15] M. R. Endsley, “Theoretical underpinnings of situation awareness: a
critical review,” in Situation Awareness Analysis and Measurement.
Mahawah, NJ, USA: Lawrence Erlbaum Associates, 2000.
[16] E. G. Little and G. L. Rogova, “Designing ontologies for higher level
fusion,” Information Fusion, Special Issue on High-level Information
Fusion and Situation Awareness, vol. 10, no. 1, pp. 70–82, January
[17] D. L. Hall and J. Llinas, Handbook of Multisensor Data Fusion. CRC
Press, 2001.
[18] J. Llinas, C. Bowman, G. Rogova, A. Steinberg, E. Waltz, and F. White,
“Revisiting the JDL data fusion model II,” in Proceedings of the
Seventh International Conference on Information Fusion, P. Svensson
and J. Schubert, Eds., 2004, pp. 1218–1230.
[19] K. Bernardin, R. Stiefelhagen, A. Pnevmatikakis, O. Lanz, A. Brutti,
J. R. Casas, and G. Potamianos, Computers in the Human Interaction
Loop. Springer, 2009, ch. Person Tracking, pp. 11–23.
[20] A. Pnevmatikakis and F. Talantzis, “Person tracking in enhanced cognitive
care: A particle filtering approach,” in The 18th European Signal
Processing Conference (EUSIPCO 2010), Aalborg, Denmark, August
[21] N. Katsarakis and A. Pnevmatikakis, “Face validation using 3d information
from single calibrated camera,” in The 16th International
Conference on Digital Signal Processing (DSP 2009), Santorini, Greece,
July 2009.
[22] K. Avgerinakis, A. Briassouli, and I. Kompatsiaris, “Real time illumination
invariant motion change detection,” in Proceedings of the ACM
Multimedia 2010 Workshop – 1st ACM ARTEMIS2010 International
Workshop, October 2010.
[23] M. M. Kokar, J. J. Letkowski, R. Dionne, and C. J. Matheus, “Situation
tracking: The concept and a scenario.” in Situation Management
Workshop: SIMA’08. IEEE, MILCOM, 2008.