A Generic Bayesian Network for Identification andAssessment of Objects in Maritime Surveillance

Max KrugerHochschule Furtwangen University (HFU)

Robert-Gerwig-Platz 1, D-78120 Furtwangen, Germany

e-mail: max.krueger@hs-furtwangen.de

Jurgen Ziegler and Kathrin HellerIndustrieanlagen-Betriebsgesellschaft mbH (IABG)

Einsteinstraße 20, D-85521 Ottobrunn, Germany

e-mail: jziegler@iabg.de, kathrin-heller@mytum.de

Abstract – Identification and assessment of objects arekey capabilities of surveillance and information systems formaritime environments. Bayesian methodology providesestablished instruments for fusion of uncertain identifica-tion and assessment indications from various informationsources. Particularly, Bayesian Networks are well qual-ified for modeling and inference in this application con-text. In order to standardize and simplify generation ofBayesian Identification Networks for new operational sce-narios, a generic Bayesian Network for identification is pro-posed. This generic model defines adequate and neces-sary node types, well suited for modeling of identificationand assessment tasks. In addition, the model provides adependency structure by subgraphs, which simplifies pro-vision of more complex cause-effect relations in BayesianIdentification Networks. Nevertheless, the generic model isflexible enough to cover various application scenarios withmanifold operational demands. Closing, generation of aBayesian smuggler detection Network for maritime surveil-lance in an exemplary littoral application scenario is pre-sented.

Keywords: Identification, Affiliation&Threat Assessment,Bayesian Network, Generic Model, Uncertainty Modeling,Maritime Surveillance Systems

1 IntroductionIdentification of a tracked object and assessment of its af-filiation and threat potential are essential capabilities in civiland military surveillance systems, e.g., in maritime or airen-vironment.Assignment of an identityandaffiliation&threatassessmentof an object are related tasks, see [1], [2], [3].This holds in problem formulation and solution techniques,so we subsequently subsume both by the termidentification(ID). In real-life scenarios, there is a large number of objectsto be judged simultaneously, and consequently a strong needfor automated assistance.

Bayesian techniques, and in particular Bayesian Net-works provide a capable framework for an ID process ([1,ch. 5-7], [2, ch. 8,12], [3, ch. 7-9]). Existing Bayesian ID ap-proaches tend toward a modeling that is driven by individual

technical sensor capabilities and application specific struc-ture, e.g., see [4], [5]. This makes approaches application-dependent and inflexible with respect to technical changesand enhancements, for instance new sensor technologies.More generically designed approaches, such as the standard-ized Identification Data Combining Process(IDCP, see [6],[7]) are constricted by certain independence assumptions ofsource measurements and preassigned source types, result-ing from its initial context of air defense.

This paper proposes a generic Bayesian IdentificationNetwork (Bayesian ID Network) with extended applicationspectra in various environments, fulfilling extended opera-tional demands. These extensions include readiness for inte-grating a variety of new technical sources as well as applica-tion in changing and new scenarios in asymmetric warfare.

Outline of this work: In section 2 we describe the use ofBayesian techniques and Networks for identification of ob-jects. A generic Bayesian ID Network is introduced and itsdetails are discussed in section 3. The following section 4considers selected aspects of generation and configurationofBayesian ID networks, and benefits of a source type taxon-omy of contributing information sources. Section 5 presentsa detailed example of maritime surveillance. Conclusionsand future work are outlined in section 6.

2 Bayesian IdentificationClassical identification by Bayesian inference uses theThe-orem of Bayes

p(idi|d1, . . . , dN ) =p(d1, . . . , dN |idi) · p(idi)

N∑j=1

p(d1, . . . , dN |idj) · p(idj)

(1)

to assess the posterior probabilities of possible identitiesid1, . . . , idK , given discrete measurementsd1, . . . , dN ofN ID sources ([2, pp. 496-497]). As a more sophisticatedprocess, the IDCP combines several of theseNaıve-Bayesassessment steps into one process ([6], [7]).

A Bayesian Network is a graphical probabilistic model.Following [8, pp. 33-34], it is defined asDirected Acyclic

Figure 1: Bayesian ID Network without tree structure

Graph (DAG) with each node representing a discrete ran-dom variableA1, . . . , AN , and directed edges representingdependencies between these variables. While modeling withBayesian Networks, careful attention must be paid to direc-tion of dependencies: Causes are modeled as parent nodes,effects as child nodes. Each variableAi is in one out of afinite number of mutually exclusive, node-dependent statesai,1, . . . , ai,Ni

, i.e.,Ai = ai,j . Additionally, each nodeAi

is associated with a Conditional Probability Table (CPT),which describes the conditional probabilities

p(Ai = ai,j |Ap1= b1, . . . , ApK

= bK) (2)

for all (given) state combinationsb1, . . . , bK of all parentnodesAp1

, . . . , ApK∈ pa(Ai) of Ai in the DAG.pa(Ai)

denotes the parent node set ofAi. Bayesian Networks rep-resent the joint probability distributionsp(A1, . . . , AN ) oftheir variables. By theChain Rule for Bayesian Networks

p(A1, . . . , AN ) =

N∏

i=1

p(Ai| pa(Ai)) (3)

the joint probabilities can be calculated ([8, pp. 36-37]).Generally, the state of a node variable is not known. Sourcemeasurements (evidences) in Bayesian ID Networks unveilstates of some node variables. Due to dependencies be-tween nodes, state probabilities of other nodes must be up-dated. This provides an inference mechanism in BayesianNetworks between source nodes and ID nodes.

Naıve-Bayes ID approaches and its IDCP extension canbe expressed as Bayesian Network ([1, pp. 283-285]). Be-cause of their modeling-inherent (conditional) independenceassumptions, both DAGs have tree-structure. But modelingwith Bayesian Networks allows more complex structures,capturing application-specific dependencies within an oper-ational scenario. Figure 1 shows a Bayesian ID Networkwithout tree-structure, taken from [4]. Finally, ID resultsand other higher-level information on tracked objects can beused to further improve tracking ([9]).

3 Generic Bayesian ID Network3.1 Generic ModelBayesian Networks for identification with a more complexstructure (see e.g. [4], [5]) have additional nodes next to thenecessarycauseandeffectnodes, reflecting source measure-ments and ID results. Abstraction reveals five basic nodetypes that occur in Bayesian ID Networks:

Identities&Assessments(IDA-node type):These nodes describe ID-characteristics such asFriendly/Hostile or Civil/Military in classical settings, orSmuggler,Pirate, Terrorist and Renegadein more recent scenarios.Generally, IDA-nodes are used to specify the possible iden-tification results from an operational perspective. Selectionof relevant possible ID results is based upon the operationalapplication demands. For instance, allegiance is a typicalcharacteristic to be described as IDA-node.

Affiliations&Intentions&Purposes (AIP-node type):Affiliations, intentions, and purposes form the major intrin-sic personality of an object, which impacts almost all facetsof an object directly or indirectly. Their aggregation can beinterpreted as operational personality of the object. There-fore, AIP-nodes describe the initial operational motivationand characterization, which determine appearance, behaviorand activities of the object. For instance, considerhostileattitudeof a terrorist towards a nation, orfriendly affiliationof an own air force fighter plane.

Behavior&Attributes&Capabilities (BAC-node type):Behavior, attributes, and capabilities are consequences ofAIP-characteristics. Their aggregation can be interpreted asoperational and technical behavior and properties of an ob-ject. BAC-nodes manifest intrinsic facets of AIP-nodes inbehavior and properties, which are (by adequate measuresunder certain circumstances directly) observable from theobjects outside. For example, BAC-nodes describe avail-ability or use of technical properties and features, e.g., pres-ence or utilization of anIFF mode IV deviceor thedraft ofa vessel, determining its ability to approach coastal waters.

Sources(S-node type):Technical and non-technical sources provide measurementsand information of BAC-aspects. Sources are modeled as S-nodes, accounting in particular for uncertainties and inaccu-racies related with measurements. Possible results of sourcemeasurements are described as states of S-nodes. Therefore,assignment of conditional probabilities for these states takesaccount of related uncertainties and inaccuracies of theirmeasurement process. Considerfrequency types of radardevicesor adherence to sealanesas example.

Environment&Conditions&Settings (ECS-node type):Environment, conditions, and settings cover all object-external aspects that influence objects appearance, behav-ior and activities as well as corresponding measurementsby sources. Being not object-inherent, ECS-nodes are in-terpreted as causal factors, that are located outside the con-sidered object. Typical examples for ECS-aspects are envi-ronmental conditions, such as weather or terrain.

The DAG of the proposed generic Bayesian ID Networkconsists of five subgraphs, one for each node type, as shownin figure 2. Note that each subgraph contains only nodes of

Figure 2: Generic Bayesian ID Network

the dedicated type. The large arrows in figure 2 depict allpossible directed dependencies between nodes of differentsubgraphs. Therefore, inter-subgraph dependencies are onlyallowed as follows:

• IDA-nodes depend only on AIP- and BAC-nodes.

• AIP- and ECS-nodes have only subgraph-internal par-ent nodes.

• BAC-nodes depend only on AIP- and ECS-nodes.

• S-nodes depend only on AIP-, BAC-, and ECS-nodes.

Beside these depicted ones, there are no other dependenciesbetween nodes from different node types. Within subgraphs,arbitrary dependencies are admitted, as long as the DAG-property is preserved within each subgraph. Obviously, theDAG-property then holds for the entire directed graph, ifit satisfies the structure given by the generic Bayesian IDNetwork in figure 2. An example of a Bayesian ID Networkwhich complies with the generic Bayesian ID Network isgiven in figure 3.

3.2 Structural PropertiesSelected structural and content properties of the proposedgeneric Bayesian ID Network are discussed in this subsec-tion. They make the network appear appropriate for a va-riety of ID tasks in different environments under extendedoperational demands:

The five basic node types, IDA-, AIP-, BAC-, S-, andECS-nodes describe and distinguish the relevant aspects ofID-modeling on a sufficient level of abstraction. Besides theDAG-property, there are no other restrictions in modelingpossible dependencies of nodes within each node-type spe-cific subgraph. In Naıve-Bayes modeling, these inner-typedependencies are ignored, because their description is notpossible. In Bayesian ID Networks, carful and frugal use of

Figure 3: Model-compliant Bayesian ID Network

setting dependencies is recommended in order to keep themodels manageable and applicable.

Identity and assessment results, modeled as IDA-nodes,are always consequences of deliberate intentions and activ-ities: ’No source measures identity directly’. Consequently,in the generic Bayesian ID Network, IDA-nodes only de-pend on AIP- and BAC-aspects and not directly on techni-cal measurements, given by S-nodes. Differing from mostclassical approaches, identity aspects are expected to bemodeled in several IDP-nodes. This is due to their mul-tifaceted meaning and structure, and facilitates the under-standing and configuration of the underlying model. Forexample, an object can be identified with respect to itsCivil/Military , Friendly/Neutral/Hostile, Combatant/Non-combatant, andPrivate/Commercial/Governmentalcharac-teristics simultaneously. Classical approaches sometimesaddress this issue by using parallel assessment processes fordifferent ID aspects based on different models.

Operational personality is modeled by AIP-nodes. Thesecharacteristics depend only on inherent intentions and ex-istence purposes of an object. Therefore, AIP-nodes canonly have AIP-subgraph internal dependencies. With someexceptions, AIP-nodes characteristics are not measured di-rectly, but induce technical properties that can be measured.Operational expectations on the frequency of different ob-ject types determine the a priori parameters in the CPT ofAIP-nodes.

Operational and technical behavior and properties aremodeled by BAC-nodes. They describe operational andtechnical characteristics, for instance presence or use oftechnical devices. Besides dependencies on other opera-tional and technical behavior and properties, BAC-nodesdepend only on operational personality (AIP-nodes) of theobject, and additionally on environmental factors (ECS-nodes). For example, the BAC-aspectPerforming Smug-gling Activities depends on the AIP-aspectIntention toSmuggleand the ECS-aspectPoor sight conditions.

Characteristics of sources, modeled by S-nodes describe

the technical and non-technical sources that provide mea-surements and information on objects. This description in-cludes the technical properties of sources, in particular un-certainties and inaccuracies of measurements and its inter-pretation. With some exceptions, source properties dependonly on operational and technical behavior and properties(BAC-nodes), which are measured by the considered source.For instance, a sourcesudden maneuvermay indicate a mea-surementsudden change of course. While the BAC-nodestate represents an actual activity, the S-node state representsthe corresponding measurement. In some cases not an op-erational or technical behavior or property but operationalpersonality (AIP-node) is measured directly. Consider in-telligence information onSmuggling Intentionof the objectas example. In such cases measurements can be interpretedas update of the AIP-node a priori distribution. Concluding,S-nodes can depend on BAC-, ECS-, and AIP-nodes, and asusual, on other S-nodes .

Environmental aspects model external factors by ECS-nodes. Therefore, they do not depend on any object aspect,but are causal to S- and BAC-nodes.

To our observation, other properties and dependenciesseem to be subsidiary for sufficient identification. There-fore, they should be omitted in support of reduced com-plexity of modeling and algorithms. The generic BayesianID Network provides a modeling structure that preservesthe causal structure of a given ID problem, compare [10].Due to its generic characteristics, it is flexible enough tocover various identification tasks for manifold operationaldemands.

4 Aspects of Practical Operation4.1 Generation and ConfigurationThe generic Bayesian ID Network is basis for generationof a Bayesian ID Network for a particular application sce-nario. A proceeding of generation is described in [10]: Op-erational experts perform a tool-based definition of the ap-plication network by defining relevant nodes, their states andtheir dependency structure in compliance with an underly-ing generic model. The proceeding in [10] takes accountof the fact, that knowledge of Bayesian modeling cannot beexpected from operational experts. Nevertheless the outputof this first step is a qualitative model for the applicationscenario, covering relevant entities and dependencies.

Quantitative parameters of the application model, i.e.,configuration data of the Bayesian ID Network, can begained byUser-oriented Configuration([11]): This ap-proach provides a process and techniques of Bayesian con-figuration-data acquisition. Problems related to appropriatehandling of statistical information in Bayesian Networks areavoided. Particularly, this includes the known problems ofmisinterpreting conditional probabilities [11].

4.2 Source Data Exchange and TaxonomyIn order to improve overall identification quality of an ob-ject, all available source information should be used by

Figure 4: Contents of Source Data Exchange

an ID process. Particularly, this includes information gen-erated by other dislocated, cooperating identification in-stances, compare [6]. Exchange of source information thatis independent of the underlying ID assessment methodol-ogy provides the advantage of being able to use many ad-ditional sources. Standardized coding of (source) informa-tion is necessary precondition ofSource Data Exchange.In figure 4 minimal content of Source Data Exchange issketched. Generic fields can be adapted to formats usedby data transmission systems, e.g., Tactical Data Link. Thefields ID Source NumberandSource Declarationare non-generic fields and to be standardized.

Development of aSource Type Taxonomyis in progress,which allows coding of source information, for instanceIDSource Number, using an appropriate format. Each sourceinstance, e.g., technical devices or human observations, isassigned to a combination ofSource Type, Source Subtype,andSource Device Classindices according to a predefinedtaxonomy. Source Types have to be designed in a such way,that all information can be encoded. For air defense andsurveillance, a taxonomy of source types has been definedin [7]. For environments other than air, definition of an ad-equate taxonomy and corresponding assignment of sourcetype indices is work in progress. A selection of possiblesource types in maritime context includes:

• Classifying Non-Imaging Systems, e.g., ESM,

• Movement Plans&Procedural Routing,

• Track Behavior,

• Identification By Origin,

• Protected Network Location and Identification,

• Visual Sightings,

• Platform Performance,

• Events,

• Self-Identification, e.g., AIS, and

• Imaging Classification Systems.

Note that these source types are not complete and belongto the most abstract taxonomy level. Expandability of IDprocesses and adaptability to new application scenarios aremajor guidelines within the design process. The SourceType Taxonomy is intended as foundation of Source DataExchange and cooperation in networks of heterogeneous IDsystems.

Figure 5: Littoral area at Scott Islands

5 Maritime Application ExampleThe Evaluation of Techniques for Uncertainty Represen-tation Working Group(ETUR-WG) prepared aShip Locat-ing and Tracking scenario(see [12]) based upon illegal im-migration detection in a maritime environment. We use thisscenario as context for our subsequent Bayesian Networkapplication. This exemplary application can be transferredto other surveillance tasks and domains, e.g., air and ground,as long as stable tracking service is available. A similar sce-nario assessing arms smuggling with speed boats is analyzedin [13]. By use of the IDCP ([6], [7]) a Bayesian ID processwas implemented for demonstration.

5.1 ScenarioThe scenario ([12]) in located in littoral waters, exemplarilylocated at Scott Islands, Canadian West Coast near Vancou-ver, see figure 5. A sealane from/to Asia runs in West-Eastdirection and a major Tanker route passes Scott Islands inNorth-South direction. Both routes cross in large fishinggrounds. Besides cargo and oil tanker traffic, there is a lotof fishing and leisure activity in the area. Military and gov-ernmental vessels supervise the area.

From intelligence sources it is known, that people smug-glers intent to transport illegal immigrants on cargo shipsand offload them by several trips with zodiacs from ship tocoast. All vessels and boats involved in smuggling try tohide their activities and spoof identities by imitating fisheryor leisure behavior and use of other measures. Identifica-tion task is to find the objects involved in smuggling and todiscriminate them from regular commercial, fishery, privateand military/governmental traffic.

5.2 Modeling of the Bayesian ID NetworkRelevant components to be modeled in this scenario areidentitiesaccording to the operational demands, maritimeinformation sourcescontributing to ID processes,intentionsand behaviorof vessels, andenvironmental factors, that in-fluence the ID processing. A dependency modeling for thesecomponents should take uncertainties and inaccuracies intoaccount. Modeling an ID process for this scenario resulted

in the Bayesian ID Network for smuggler detection in lit-toral waters, given in figure 6. This network consists of 40nodes with a total of 97 states, 46 direct dependencies, and399 parameters. Behind each node name in figure 6 its nodetype is denoted in brackets.

In more details, the Bayesian Network in figure 6 re-sulted from application of the generic Bayesian ID Network.According to scenario description ([12]), theIdentities&-Assessmentssubgraph contains a nodeSmuggler (IDA),which provides major discrimination between people smug-glers and others.Civil/Military Discrimination (IDA) andvesselCategory (IDA)node cover secondary identificationdemands in this scenario. Together, these three ID resultsprovide an appropriate operational description of a vessel.

Within the Affiliations&Intentions&Purposessubgraph,the Type Affiliation (AIP)node depends onSmuggling In-tention (AIP)andMilitary Affiliation (AIP) node.Type Affil-iation (AIP) discriminates betweenTransport Ship, Passen-ger Vessel, Fishing Vessel, Leisure Boat, andPatrol Vessel.Additionally, each AIP-node is causal to its correspondingIdentities&Assessments-node.

Seven behavioral facets, i.e.,Military , Cargo&TankerTransport, Passenger Transport, Fishing, Costal Cruising,Transporting Illegals, and Offloading Illegals, are givenin the Behavior&Attributes&Capabilitiessubgraph. Thesefacets each depend on one AIP-node. The additional depen-dency ofFishing Behavior (BAC)on Transporting IllegalsBehavior (BAC)is due to the fact, that smugglers transportvessels imitate fishing behavior. For the same reason,CostalCruising (BAC), which manifests leisure affiliation, dependsonOffloading Illegals Behavior (BAC).

TheSourcessubgraph contains all sources, that can pro-vide evidences resulting from human observations and tech-nical devices. In order to keep the model simple, each sourcenode gives evidence for (or against) only one behavioralfacet. Therefore, each source node has few states, mostlyonly two, and depends only on one (behavioral) BAC par-ent node. Exceptions are the nodesCivil Rendezvous Report(S)andAIS (S). The former models the observation of a ren-dezvous between smuggler cargo vessel and zodiac. Thisnode gives evidence forTransporting Illegalsas well as forOffloading Illegalsbehavior, and consequently has two BACparents. Same holds forAIS (S). The source nodesElec-tromagnetic Emissions (S)andVisual&EO/IR Classification(S)provide measurements that rather give evidence on affil-iation then on a certain behavior of a vessel. Consequently,these two sources depend directly on theaffiliation (AIP)node and not on a (behavioral) BAC node.

Few environment modeling is provided. TheEnviron-ments&Conditions&Settingssubgraph contains only threenodes. NodesVisibility (ECS)andSea State (ECS)dependon the common cause-nodeWeather (ECS)and influencemeasurements by S-nodes.

Configuration of the Bayesian ID Network has been per-formed with data available in or in compliance with the sce-nario description ([12]). Application results of this quan-titative modeling part essentially depend on an appropriate

Figure 6: Bayesian ID Network for smuggler detection in littoral waters

qualitative modeling of the Bayesian Network graph. Foroptions of tool-based configuration support, we refer to [11].

5.3 ExperienceSound modeling with a graph structure not too complex,i.e., not too many dependencies, allows easier configura-tion of the Bayesian ID Network. Consequently, a majorpart of generating a Bayesian ID Network involves defini-tion of an adequate DAG. We judge the generic BayesianID Network as helpful in defining Bayesian Networks forapplication, because the approach reduces definition of theneeded DAG in large parts into defining appropriate nodesand corresponding states. Construction is further supportedby semantics of nodes, given by the subgraph-specific nodetypes.

Looking at dependencies between nodes from differentsubgraphs, the cause-effect direction is instantly provided bythe generic Bayesian ID Network. Considering the remain-ing ’free-style’ part of modeling dependencies, we found thefollowing hints helpful, while using the generic network:

• Normally, nodes do not need more than one parentnode from each other subgraph. Otherwise, configu-ration can be difficult, given conflicting states of par-ents. If possible, source nodes should be modeled asdepending only on its strongest BAC-type parent.

• Dependencies within each subgraphs have to be mod-eled carefully and economically, i.e., only direct andstrong cause-effect relationships are relevant.

• Many simple nodes with only few states are preferablecompared to fewer complex nodes with many states. Inparticular, this holds for S- and BAC-type nodes if dif-ferent evidences indicate more than two different typesof behavior or intentions.

• For nodes with more than two states, splitting should beconsidered, in particular if nodes have several parentsfrom same or other subgraphs. If not possible, try toidentify a common property of all same-type parentsand create a new property node.

These hints are (obviously) not obligatory. They are relatedto particular structure of Bayesian ID Networks complyingwith the generic Bayesian ID Network, and might not be ap-plicable in other Bayesian Networks. Recommendations formodeling of general-type Bayesian Networks can be foundin introductory literature, e.g., [8, ch. 3], [14, ch. 5].

5.4 Application to Example CasesThe Bayesian ID Network given in figure 6 is implementedusing GeNIe&SMILE ([15]) software package. In thepresent subsection we apply our generated Bayesian ID Net-work for smuggler detection to several exemplary cases:

Military Vessel:A frigate leaves its home base and moves to a gunnery range.Available source evidences on this object are:

• Apparent Origin Naval Base (S)→ Match.

• Mission Plan (S)→ Match.

Identification results in figure 7 clearly indicate a militaryship and exclude any smuggling activity.

Figure 7: Assessed ID military vessel

Cargo Ship:Coming from Asia, a freighter transports goods to Canada.It shows normal behavior without any anomaly. Availablesource evidences on this object are:

• Sealanes (S)→ Match.

• AIS (S)→ Match.

• Apparent Origin Asia (S)→ Match.

Identification results in figure 8 indicate a commercial civilship with no smuggling affinity.

Figure 8: Assessed ID cargo ship

Fishery Vessel:False evidence is given on this fishing vessel. By techni-cal misassignment, a match on unexpected sealane leavingis given. Besides that, the vessel shows typical fishing be-havior. Available source evidences on this object are:

• Unexpected Sealane Leaving (S)→ Match.

• Low Speed&Unsteady Course (S)→ Match.

• Fishing Grounds (S)→ Match.

Despite false evidence, identification results in figure 9 in-dicate a fishing vessel. The slightly increased smugglingaffinity is due to the false evidence given.

Figure 9: Assessed ID fishery vessel

Private Yacht:A private yacht moves at medium speed in coastal waters.Available source evidences on this object are:

• Littoral area (S)→ Match.

Identification results in figure 10 indicate a small civil ob-ject. Smuggling affinity is week, but in this case there is fewevidence overall.

Figure 10: Assessed ID private yacht

Freighter Transporting Illegals :Transporting illegals, a freighter coming from Asia showsnormal behavior, but starts heading towards its rendezvouspoint. Available source evidences on this object are:

• Sealanes (S)→ Match.

• AIS (S)→ Match.

• Apparent Origin Asia (S)→ Match.

• Unexpected Sealane Leaving (S)→ Match.

Identification results in figure 11 indicate a commercial ship.There is a suspicion towards smuggling, but so far no clearsmuggling indication by any source. Contact seems to beworth of further observations. Compare to case ’CargoShip’.

Figure 11: Assessed ID freighter transporting illegals

Smuggler Zodiac Offloading Illegals:A zodiac meets a cargo ship and afterwards moves towardscoastal waters offloading illegals. The rendezvous is ob-served and reported by a fishing vessel. Available sourceevidences on this object are:

• Civil Rendezvous Report (S)→ Match.

• Littoral Area (S)→ Match.

Identification results in figure 12 indicate a small civil boat,probably involved in smuggling activities.

Figure 12: Assessed ID smuggler zodiac offloading illegals

Summing up application results of these exemplary IDcases, the generated Bayesian ID Network for detection of

smuggling activities yields plausible and comprehensiblere-sults. By means of Bayesian Network learning techniques(see e.g. [16, ch. 16-17,19]) and manual fine tuning (see [17,ch. 2,5]) by operational subject matter experts, slight im-provements appear possible.

6 ConclusionsIn this paper a generic model for Bayesian ID Networks hasbeen proposed. Our intention is to standardize and simplifygeneration of Bayesian Networks for identification and as-sessment of affiliation and threat. Provision of appropri-ate node types and a dependency structure reflecting thecause-effect directions simplify the generation of applica-tion networks and lead towards adequate and uniform struc-tures of Bayesian Networks for identification. Exemplarily,a Bayesian Network for smuggler detection in a maritime,littoral context has been constructed and gives encouragingapplication results.

Future work will need to address the question, if furtherdetailing of node types and dependency relations can im-prove usability of the generic model. Additionally, othermeasures for improvement of generation and configurationof Bayesian ID models for surveillance applications are tobe considered. In context of the Bayesian ID Network forsmuggler detection, discrimination capabilities, the setofcontributing sources and their possible declarations willbedetailed and completed. This might lead into an implemen-tation for further evaluation.

References[1] D. L. Hall and S. A. McMullen,Mathematical Techniques in

Multisensor Data Fusion, 2nd ed. Boston, London: ArtechHouse Publishers, 2004.

[2] S. Blackman and R. Popoli,Design and Analysis of ModernTracking Systems. Boston, London: Artech House Publish-ers, 1999.

[3] E. Waltz and J. Llinas,Multisensor Data Fusion. Boston,London: Artech House Publishers, 1990.

[4] P. van Gosliga and H. Jansen, “A Bayesian Network for Com-bat Identification,” inRTO IST Symposium on ’Military Dataand Information Fusion’ (Prague, Czech Republic, 20-22October 2003), ser. RTO Meeting Proceedings MP-IST-040.NATO Research & Technology Organization, March 2004.

[5] G. Laskey and K. Laskey, “Combat Identification with Bayes-ian Networks,” in2002 Command and Control Research andTechnology Symposium. Naval Postgraduate School (Mon-terey, California), February 2002.

[6] C. Stroscher and F. Schneider, “Comprehensive ApproachtoImprove Identification Capabilities,” inRTO IST Symposiumon ’New Information Processing Techniques for Military Sys-tems’ (Istanbul, Turkey, 9-11 October 2000), ser. RTO Meet-ing Proceedings RTO-MP-049. NATO Research & Technol-ogy Organization, April 2001.

[7] STANAG 4162 (edition 2): Identification Data Combin-ing Process (IDCP), NATO Standardization Agency (NSA),Brussels, 2009, (NATO unclassified).

[8] F. V. Jensen and T. D. Nielsen,Bayesian Networks and Deci-sion Graphs. New York: Springer Science, 2007.

[9] H. K. Chan, H. B. Lee, X. Xiao, and M. Ulmke, “IntegratedGMTI radar and report tracking for ground surveillance,” inProceedings of the 13th International Conference on Infor-mation Fusion (Edinburgh, United Kingdom 26 - 29 July2010). ISIF, July 2010.

[10] J. Ziegler and B. Haarmann, “Automatic Generation of LargeCausal Bayesian Networks from User Oriented Models,” inProceedings of the 6th Workshop on Sensor Data Fusion (IN-FORMATIK 2011: Informatik schafft Communities). Berlin:Gesellschaft fur Informatik e.V. (GI), 2011.

[11] M. Kruger and J. Ziegler, “User-Oriented Bayesian Identi-fication and Its Configuration,” inProceedings of the 11thInternational Conference on Information Fusion (Cologne,Germany 30 June - 03 July 2008). ISIF, July 2008.

[12] Evaluation of Techniques for Uncertainty Representation(ETUR) Working Group, “Use case 1: Ship Locating andTracking,” http://eturwg.c4i.gmu.edu/?q=UseCase1, Interna-tional Society for Information Fusion (ISIF), accessed on3rd March 2011, (Webpage and accompanying document”Higher-level fusion ship locating and tracking scenario forETUR” by Pierre Valin (DRDC, Canada)).

[13] M. Kruger and J. Ziegler, “Analysis of IDCP - Applicationin MAJIIC,” IABG, Ottobrunn, Technical Report in projectMAJIIC, 2007.

[14] A. Darwiche,Modeling and Reasoning with Bayesian Net-works. Cambridge, MA(USA): Cambridge University Press,2009.

[15] Decision Systems Laboratory of the University of Pitts-burgh, “GeNIe software package for Bayesian Networks,”http://genie.sis.pitt.edu.

[16] D. Koller and N. Friedman,Probabilistic Graphical Models:Principles and Techniques. Cambridge, MA (USA): TheMIT Press, 2009.

[17] H. Chan, Sensitivity Analysis of Probabilistic GraphicalModels: Theoretical Results and Their Applications on Bay-esian Network Modeling and Inference. Saarbrucken: VDMVerlag Dr. Muller, 2009.