Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
Fusion Techniques for Reliable Information: A Survey Hyun Lee*1, Byoungyong Lee*2, Kyungseo Park*1 and Ramez Elmasri*1 *1 CSE Dept., University of Texas at Arlington, Arlington, TX, 76019, USA *2 U-ECO City Test-Bed Center, Korea Land & Housing Corporation, South Korea
[email protected],
[email protected],
[email protected], and
[email protected] doi: 10.4156/jdcta.vol4.issue2.9 multiple classifications may suffer from the wrongly selected data set [60], [87].
Abstract Information fused by multi-sensors is an important factor for obtaining reliable contextual information in smart spaces which use the pervasive and ubiquitous computing techniques. Adaptive fusion improves robust operational system performances then makes a reliable decision by reducing uncertain information. However, these fusion techniques suffer from problems regarding the accuracy of estimation or inference. No commonly accepted approaches exist currently. In this survey, we introduce the advantages and disadvantages of fusion techniques which can be used in specific applications. Second, we categorize well-known models, algorithms, systems, and applications depending on the proposed approaches. Finally, we discuss the related issues for fusion techniques within the smart spaces then suggest research directions for improving the decision-making in uncertain situation.
Figure 1. A Fusion System consists of Three Layers Therefore, in this survey, first we review the related fusion concepts: 1) terminologies, 2) objectives, and 3) advantages and limitations. Second, we compare wellknown models, algorithms, systems, and applications based on different approaches to discuss uncertainty in estimation. Finally, we introduce related current issues and research directions that support to obtain reliable contextual information by adapting different methods. For instance, heterogeneous sensor management [71], a multiple classifier [81], an ensemble method [72], and an adaptive multi-level fusion [11], [39] are adapted to pervasive and ubiquitous computing areas. The remainder of the paper is organized as follows. In section 2, concepts about some fusion techniques are introduced. Fusion models are discussed in section 3. In section 4, we compare famous fusion algorithms and theories that are applied to the estimation or inference methods. We discuss related systems for specific areas in section 5. Section 6 deals with issues and suggests research directions for fusion techniques. In section 7, we then conclude the survey.
Keywords Survey, Reliability, Data, Multisensor, Information, Fusion, Pervasive and Ubiquitous Computing, System Performance.
1. Introduction In pervasive and ubiquitous computing applications such as smart homes, offices, hospitals, and spaces [36], [42], [59], [61], [88], [107], some fusion techniques have been proposed to reduce uncertainty and achieve reliable data processing and analysis in a fusion system as shown in Figure 1. However, it is still open problem to obtain the reliable contextual information in specific applications using the proposed fusion technique. First, no commonly accepted approaches, which can estimate uncertainty in a fusion system, exist. Second, raw-level data or input-output characteristics of each device may cause the total uncertainty of estimation or inference. Third, a higher inference can not correct possible errors that may occur in a lower level of data processing. And last, identification tasks and decision makings based on
2. Concept 2.1 Terminology Fusion means the combination of sensory data that merges heterogeneous sources into a union. Integration
74
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
with the range of a temporal change. They then control and manage the fusion processes actively. In order to formalize the fusion process, three procedures operate as below:
means the aggregation of sensory data to make whole thing by simply bringing the parts of data together [14].
Table 1. Performance Factors for Information Fusion No Advantages Limitations 1 Robustness High-Complexity Extended Spatial 2 Fault Diagnosis Coverage Extended Temporal 3 Fault Detection Coverage 4 High-Confidence Fault Analysis 5 Low-Ambiguity Unbelief 6 Reliability & Validity Uncertainty 7 Low-Vulnerability Misapplication Figure 2. Fusion Types based on the Definition
1) Related approaches will be defined when no predefined rule, which satisfies a set of contexts, exists. We have knowledge about which contexts a particular fusion operator satisfies and which fusion operators satisfy a set of contexts. For characterizing the fusion process, functional equations of the fusion process are used as a fundamental tool. 2) For deciding the most appropriate fusion process in the given situations, we select methods. The selected method determines the modeling capability of a fusion process. In this case, the selection of the fusion process corresponds to the trade-off between expressivity and simplicity. We then consider the fusion operator which will be utilized to build universal approximators. 3) We determine algorithms or mechanisms to find the best parameterization of the given fusion operators. In addition, most fusion operators are parametric and their behavior strongly depends on the parameters. We should consider the relationship between operators and parameters.
The two concepts are similar to each other. However, the definition of “fusion” has a different meaning by depending on whether a machine or a user processes the events. As described in Figure. 2, "Machine fusion" is the integration by computers for obtaining data from multiple sources. This fusion is utilized to get a unified, a completed, and an accurate representation for objects [26]. For instance, the fusion of a signal processing of sensor measurements, which is used for identification, gets information from objects. Among machine fusion techniques, "Data fusion" considers a dynamic process such as the raw data association or correlation for fused product [101]. This fusion improves the accuracy than any product which utilizes separate raw data elements. "Sensor fusion" aggregates data from multi-sensors to provide an improved accuracy when it compared with only a single sensor value [42]. This fusion achieves a reliable inference by predicting the entity states and by reducing uncertainty. "Information fusion" is a multilevel process. This fusion supports data mining and the integration of information from selected multi-sources to produce the fused information [88]. It has a greater value than any of its parts. "Cognitive fusion" is one of user processes for integrating and combining contexts to generate knowledge about the situation. This fusion helps to make a decision for the situation awareness in pervasive computing applications [13].
2.3 Advantages and Limitations It is difficult to achieve reliable information based on the fusion process for some reasons even though we understand the fusion process and the existing methods correctly. There exist reduction, detection, association, correlation, estimation, data combining and controlling, dimensional reduction of uncertain data, and so on [88]. Therefore, we categorize some important performance factors for information fusion as shown in Table 1 to overview merits and demerits of information fusion. In this case, the advantages and limitations are defined by the utilized multi-data resources instead of the utilized single data resource. We explain the definition of each performance factor as below.
2.2 Goals of Fusion in Pervasive Computing The goal of fusion is that applied fusion techniques increase reliability of obtained information by reducing uncertainty. Some books [90], [100] formalize fusion processes then apply them into appropriate applications. For instance, they model uncertainties of sensory data
75
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
3.1 Information Fusion (IF) Models 2.3.1 Advantages
IF models are expected to play a crucial role in the support systems for aiding decision making and public security operation in home-based care or intelligent home security, since these models provide an extended spatial or temporal coverage. To present a coherent representation of an evolving situation, they allow for the management of large volumes of information or the correlation of unrelated deceptive information. Three models: JDL fusion model (JDL), visual fusion model (VDF), and united fusion model (UDF) are involved in IF models.
1) Robustness – One of sensors can do much to create information while others are unavailable 2) Extended Spatial Coverage – One of sensors can cover the geometrical space where others can not. Multisensor can cover whole space 3) Extended Temporal Coverage – One of sensors can detect an event at special time when others can not 4) High Confidence – Effective integration of the multiple sensors’ independent measurements on the same event increases the confidence and improves detection performance 5) Low Ambiguity – Multi-sensors’ measurement reduces the set of hypothesis and then achieved information has clarity 6) Reliability and Validity – Inherit redundancy of the multisensor improves the reliability of a system operation and increases understanding of data into the realized set 7) Low Vulnerability – Reduced dimensionality of the measurement space decreases the errors of a single portion of the measurement space
3.1.1 Joint Directors of Laboratories (JDL) model JDL model is one of famous models [104] in fusion system. The purpose of this model is unifying research terminology to promote technology and co-operation among six levels processing: the source processing (0), the object refinement (1), the situation refinement (2), the threat refinement (3), the resource management (4), and the cognitive refinement (5). The goal of the model is to serve as a functional model by diverse elements of the data fusion community. Hence, this model should clarify the elements of the problems and the solutions to facilitate the recognition of commonalities [16]. The extensions of JDL model refine any generalized data fusion model description by adapting important issues [57]. This model is used for target tracking or location tracking. The level 0, 1, and 2 are adapted to process data among objects and users in pervasive computing areas.
2.3.2 Limitations 1) High-Complexity – Too much data increase the complexity of the observation, if the data is collected coincidently 2) Fault Diagnosis – Misaligned data caused by non-overlapping region can confuse the system then it can make improper results 3) Fault Detection – Increased time for collecting data from dynamic object delaying association can increase the miscorrelation 4) Fault Analysis – Combining multi-modal data, which is not uniquely associated to the context of the situation can make wrong results 5) Unbelief – Combining multi-modal data into a measurement sometimes can result in a fused estimate that means nothing 6) Uncertainty – No-independent fusing data are just double counting measurements and fusing bad data with good data based on estimation can increase uncertainty 7) Misapplication – No understanding an adapted assumptions of the algorithms sometimes can make misapplied applications
3.1.2 Visual Data Fusion (VDF) model VDF model is one of models which use the integral part of human visual system for computing data and for extracting knowledge from complex data [98]. This is critical to the integration of multiple resources, without the beneficiaries of such data would be overwhelmed by volume or complexity. We will understand complex auditory and visual information that is often using one to disambiguate the others. For instance, the automated analysis based on VDF model faces severe challenges such as the lack of accurate statistical models for the signals and the high dimensionality for sampling rates at a low level [77]. However, VDF model is useful in a secured smart space where the volume and complexity of data is bigger than others, since it can disambiguate one from the others.
3. Models
76
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
and numerical calculations [65]. CDM model includes information about the possibility of their applications. This model uses a continuous integrated process for the approximation of comprehension and prediction in the interaction with the environment [97].
3.1.3 Unified Data Fusion (UDF) model UDF model utilizes a Bayesian method (Bm) which maintains the probability density on joint target statespaces [4]. This model handles nonlinear motions, nonGaussian state distributions, measurement error, or nonlinear relationships between state and measurement in [94] by computing true Bayesian posterior on joint state spaces. This model is used for pattern recognition of some users in smart spaces by learning the recursive patterns.
3.2 Decision Making (DM) Models Depending on the selected criteria, DM models can be separated into different types of categories. In this survey, we introduce three types of models: Rational Decision Making (RDM) model, Naturalistic Decision Making (NDM) model, and related Cognitive Decision Making (CDM) model. The DM activity performed in a command and control environment is represented in these models.
Figure 3. Logical Processes with Seven Steps of Rational Decision Making (RDM) model
3.2.1 Rational Decision Making (RDM) model
3.3 Situation Analysis (SA) Models
RDM model comes from the organizer behavior and the logical process follows the path from problems through solutions using seven steps orderly as shown in Figure 3. However, many assumptions exist in steps. Thus, the assumptions should be considered in terms of the quality, quantity, or accuracy [38]. Prospect theory [8], regret theory [85], expected utility theory [56], and the heuristic multi-attribute decision making [48] are used to support the RDM model related areas.
The lack of knowledge in cognitive engineering has created new problems regarding trust in designed tools and human-in-the-loop concerns. Thus, it is necessary for balancing human perspective factors with that of the system designer using the SA models. SA models allow for the design of systems taking into account the human role in a dynamic decision making process like the command and the control in pervasive computing. Endsley's model [35], [79] and Prescriptive Situation Awareness (PSA) model [3], [68] are involved in this category.
3.2.2 Naturalistic Decision Making (NDM) model The framework of the NDM model has emerged to perform complex functions cognitively in demanding situation [99]. This model studies information, which addresses initial stages of observing events and those of developing descriptive explanations. The situation marked by time pressure, uncertainty, organizational constraints, unclear objectives, changing conditions, or different experiences are involved in this NDM model. Recognition-primed model [20] and Image theory [6] are utilized to support the NDM model.
3.3.1 Endsley model According to Endsley [35], "Situation awareness is the perception of the elements in the environments within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future." Endsley model consists of three levels. The key events are identified in the environment to define the situation in the level 1. Key events of the situation semantically for higher level of abstraction of current situation are tagged. In level 2, events of the level 1 are combined to make a comprehensive holistic pattern. The level 2 serves to define current status and to support rapid decision making and action. In level 3,
3.2.3 Cognitive Decision Making (CDM) model CDM model has a human-like decision making that does not include some complex optimization methods
77
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
current situation of the level 2 is projected into the future to make a prediction of the evolution of situation. The level 3 can support a short-term planning and an option evaluation when time permits. This model is a descriptive model that can identify some basic issues in dynamic environments with uncertainty. However, this model is difficulty to support a quantitative simulation process, which makes cues for actual emulation of the human-like decision making behaviors with embedded simulation studies [79]. 3.3.2 Prescriptive Situation Awareness (PSA) model A situation is assessed by the rule. If a set of event “E” occurs, then the situation is “S”. A hypothesized situation “S” is used to generate the expected event set “E*”, which is compared with the observed event set “E”. Fundamentally, the situation assessment is an inferential diagnostic reasoning process in which the situations are represented as hypothesized reasons, the events are considered as effects, and the sensory data are considered as symptoms [68]. PSA model starts with the detection of the event occurrences. After the events are detected, the belief impacts on the situations are evaluated by backward tracing the relation between situation and event by using Bayesian belief networks. And then the updated situation assessments drive the projection of future event occurrences by forwarding an inferential reasoning to the next step of the event detection. Based on this model, the situation awareness simply means that one knows what is going on around oneself. Maintaining coherent awareness of situation is essential to successful task completion. A formal basis for situation awareness draws on sources is proposed. This model utilizes techniques from the logic, humancomputer interaction and data fusion [3]. Even though Endsley model is useful to predict the future directions of users in smart spaces, PSA model is more adaptable to activate actuators in smart spaces.
Figure 4. Currently used Theoretical Frameworks for fusion techniques
4.1 Qualitative and Symbolic (QS) approach The aggregation of information is a very important task for many tasks in high-level data fusion. Many tools are provided by QS approaches. Propositional Logic (PL), First Order Logic (FOL), Modal Logic (ML) & Knowledge Logic (KL), and Non-Monotonic Logic (NML) are involved in this approach. 4.1.1 Propositional Logic (PL) PL is a formal system that represents propositions using formulae F = F (A; Ω; Z; I). PL is formed based on the combination of atomic propositions. The system of the formal proof rules allows certain formulae to be established as theorems. The alpha set A is a finite set of elements called propositional variables or atomic formulae. The omega set Ω is a finite set of elements called logical connectives. The zeta set Z is a finite set of transformation rules whenever they acquire logical applications. The iota set I is a finite set of initial points when they receive logical interpretations. This theorem is used for data fusion in [93]. However, some inconsistencies may arise in this theorem. For instance, this theorem requires new additional propositions to existing formulas in order to restore consistency. It is one of open problems in pervasive computing areas, which need dynamic logical connectivity.
4. Algorithms and Theories Fusion processes deal with knowledge in pervasive computing applications. While a quantitative approach is candidate for uncertainty representation, a qualitative symbolic approach seems better suite to reasoning on knowledge. A hybrid approach combines merits of two approaches: the quantified evaluations of uncertainty and high reasoning capabilities. Figure 4 shows three categories of theoretical frameworks to distinguish the characteristics of each algorithm and theories based on applied applications.
4.1.2 First Order Logic (FOL)
78
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
FOL is a deductive system that extends the PL by allowing the quantification over individuals of a given domain of discourse. The PL formula can be translated into the FOL equivalently by replacing a propositional variable with a zero arity predicate. Variables in FOL represent not propositions that can be true or false but objects the formula is referring to. This theorem is used for context-aware in [1]. However, the representation of information within FOL is not modular in [1]. It is difficult to express all possible exceptions with FOL.
4.2.1 Probabilistic Theory (PrT) Random variables, stochastic processes, and related events are combined to analyze random phenomena. A mathematical abstraction of non-deterministic event or quantities may either be simply occurrences or evolves over time in an apparently random fashion [40]. The sequence of the random events shows certain statistical patterns after repeating many times. The system studies and predicts the context or the situation. To resolve uncertain reasoning in human activities, this model is used with "Possible Theory (PoT)" [34] or "Fuzzy Set Theory (FST)" [30]. In pervasive computing areas, it seems well adapted to problems involving independent random variables, which are essentially arisen at low levels of the fusion process.
4.1.3 Modal Logic (ML) & Knowledge Logic (KL) ML is any system of formal logic that attempts to treat modalities (e.g., possibility, probability, and necessity). A formal ML represents modalities using a modal operator. The basic modal operators are usually written L for "Necessity" and M for "Possibility" [25]. In addition, KL is a logical definition in multi-modal logic systems. KL is a special kind of knowledge for group agents. KL represents the state of knowledge of the agent in given state using set theory [43]. However, it is difficult to define how many modalities are used in the same situation and new modalities may arise in the middle of a decision support task.
4.2.2 Fuzzy Set Theory (FST) FST has a set, whose elements have degrees of a membership. It is represented by a pair (A; m) where A is a set and m: A [0; 1]. It permits gradual assessments of the membership of elements in a set [54]. It defines set membership as a possibility distribution in real unit interval [0, 1]. This is used for modeling which needs more than the PrT with uncertainty. For instance, the FST is used to capture a clinical uncertainty in medical data of pervasive computing applications [2].
4.1.4 Non-Monotonic Logic (NML) NML is the logic which consequence relation is not monotonic. It is a formal devised framework to capture and to represent defeasible inference. The set of the conclusions based on the given knowledge can not increase with the size of the knowledge itself. This NML is in contrast to the FOL case, which inferences being deductively valid can never be undone by new information. It adds a formula to a theory that never produces reductions of its set of consequences [23]. This is a very important framework and an active field of research in Artificial Intelligence (AI).
4.2.3 Possibility Theory (PoT) PoT treats uncertainty in incomplete information. It is a complement for PrT. The possibility distribution is a mapping from a set of states of affairs S to the real unit interval [0, 1]. A function represents actual states of affairs. For example, S = 0 means that the state S is rejected as impossible. Also, S = 1 means that the state S is accepted as possible [33]. A general possibilitybased optimization method that handles the possible variables in reliable estimation is suggested in [67]. In practice, this theory is used to model and to combine expert opinion into database interrogation systems.
4.2 Quantitative approach
4.2.4 Dempster-Shafer Theory (DST)
Numerical frameworks are described for knowledge processing and uncertainty in the view of their use in the domain of contextual information fusion & analysis. To represent and manage uncertainty in fusion process, Probabilistic Theory (PrT), Fuzzy-Set Theory (FST), Possibility Theory (PoT), Dempster-Shafer Theory (DST), Rough-Set Theory (RoST), Conditional Event Theory (CET), and Random Set Theory (RaST) are introduced in this section.
DST is a generalization of Bayesian Theory (BT) of subjective probability. It allows us to base degrees of belief for one question on probabilities for the related question. How much they differ from the probabilities depend on how closely two questions are related [9], [21]. DST is applied as a sensor fusion technique for managing uncertainty and for inferring mechanisms, which analogous to human reasoning process in [105].
79
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
This theory deals with the most kinds of uncertainty. However, this rule must be manipulated, since this rule produces the exponential increase of the number of focal elements and this rule can lead to non-intuitive results whenever the belief functions to be combined are not independent.
4.3 Hybrid approach Hybrid approaches such as Incidence Calculus (IC), Bayesian Networks (BN), and Valuation-Based Theory (VBT) are used for a global modelization of a situation. This approach mixes above two approaches. In recent, this approach is frequently used to represent contextual information in pervasive computing applications, since this approach express the spatial-temporal dependency, which represents the correlation of different contexts.
4.2.5 Rough-Set Theory (RoST) RoST is a formal approximation space ordered pair A = (U; R), where U is nonempty set of objects called universe. R is equivalence relation on U called indiscernibility relation that is difficult or impossible to perceive. RoST gives lower and upper approximation of an original set [109]. It deals with vague information and obtains decision attribute's class from multi-class cases. However, this theory can deal only with discrete values as it uses the granularity structure of the given data. Medical data analysis, image processing, or voice recognition utilizes this theory for their data analysis. Thus, RoST seems to be appropriate for data analysis.
4.3.1 Incidence Calculus (IC) IC is a mechanism for automatic reasoning that uses a logical formula and manipulates these formulas with rules of inference for uncertain knowledge [19]. It is a probabilistic logic to manage uncertainty in numerical methods and in symbolic methods. Incidence functions not only represent the uncertain state of propositions but also represent absolute true states of propositions. 4.3.2 Bayesian Network (BN)
4.2.6 Conditional Event Theory (CET) CET deals with conditional probability of an event B in relationship to an event A. Event B occurs given that event A has already occurred. The notation is P (B\A) called as the probability of B given A. The formula for CET is P (B\A) = P (A and B) = P (A). The CET is used for treating uncertainty with the belief and inference methods in [55], [70]. However, these rules are modeled by conditional events implication, which is incompatible with conditional probabilities. The aim of CET is to find a suitable mathematical framework in which rules can be consistently modeled as conditional events.
BN, a probabilistic graphical model, solves decision problems under uncertainty. It represents probabilistic relationships between variables set and independencies. The goal of BN is to make estimations and inferences in multisensor fusion based knowledge-based situation aware of pervasive computing [106], [112]. The local computation technique provides a solution to problem of computational complexity which is involved in joint probability distributions. Medical diagnosis is a main application of BN. However, the BN is difficult to deal with temporal variations that should be considered in decision-supporting systems. The dynamic BN, which considers temporal dimensions, is more useful theory in pervasive computing areas [69].
4.2.7 Random-Set Theory (RaST)
4.3.3 Valuation-Based Theory (VBT)
RaST deals with the object with values being sets. Random elements take values as a subset of spaces and serve as a model for set-valued observations and irregular patterns. The random set direct estimation is used for multi-user detection [10]. DST is combined to clarify independence of the confidence degrees from multi-sources based on RaST in [110]. RaST provides a general framework for representing and manipulating different uncertainty. It also appears to be a unifying framework where most theories of uncertain reasoning can be justified. However, the formalism is rarely used since it requires the large size of memory.
VBT is a map from the class of open sets of the topological space to the set of positive real numbers including infinity. VBT system, which is an axiomatic framework, may represent the uncertain knowledge in different domains including the BN, the DST of belief function, epistemic beliefs, and the PoT [45]. Based on VBT, the knowledge is represented by the valuations associated graph, with nodes and links corresponding to dependencies between variables.
5. Systems and Applications 5.1 Systems
80
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
used to manage the processing and the interpretation of the image data for context-aware [96]. 5.1.4 Multi-Agent System (MAS)
To process the fused data, Distributed system (DS), Knowledge based decision support system (KDSS), Blackboard system (BS), or Multi-agent system (MAS) is used. As shown in Figure 5, a pervasive healthcare monitoring system (PHMS) is one of MAS.
MAS is composed of multiple interacting intelligent agents to solve problems, which has a difficulty to be solved by an individual agent or a monolithic system.
Figure 6. The example of Pervasive Computing Applications with Fusion Techniques
Figure 5. Pervasive Healthcare Monitoring System 5.1.1 Distributed System (DS)
MAS deals with heterogeneous sensor data in different applications. For instance, MAS is adopted to support a reliable healthcare monitoring in [113]. Different types of a MAS have currently utilized to process contextual information of between the users and the environments.
Multiple sensors, platforms, or information sources, which are fused to generate information over networks, have led to develop DS. The key problem for fusion in DS is to determine the coefficient of the data [44]. In [108], sequential filtering and consensus algorithms are provided for scalable sensor fusions with dynamic DS networks. In [114], efficient algorithms for an optimal linear estimation fusion are used with DS. Underlying relationships among objects are assumed to unknown to the automated system or human.
5.2 Applied Applications Fusion techniques are used in pervasive application areas as shown in Figure 6. Multi-sensor fusion is used for detecting and localization of the objects. A unified framework for robust shape tracking [115] is a general application for data fusion. This framework optimally fuses uncertainties or noise from the system dynamics, measurements, and a sub-space model. The problem of jointly tracking and classifying several targets within sensor networks where false detections are presented is addressed in [103]. In addition, a data fusion is used in conditional monitoring for integrating detected data and medical applications. The on-line state monitoring and fault diagnosis system is developed to transform the power during its running [46]. A tool condition in CNC milling machine is used for online monitoring based on neural networks [24] and a home-used sleep condition inference and monitoring in multi-modality sensor systems is used in [76]. A pervasive healthcare is the famous data fusion applications. An ensemble of classifiers that contains complementary information
5.1.2 Knowledge-based Decision Support System KDSS has a potential to transform the knowledge of a human decision making into the intelligent expert systems [80]. ARTMAP information system [22] uses distributed code representations for producing the selforganizing expert systems. 5.1.3 Blackboard System (BS) BS are designed for data storing in sensor systems, which surrounded by software modules operated on the blackboard. BS is implemented as the communication medium among modules within the sensor nodes in [7]. Distributed algorithmic and rule-based system has been
81
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
are included in fusion issues. These concepts can use a fuzzy logic, an evidential reasoning, a neural network, or artificial emulating tools [28]. The question that is often ignored is "when not to fuse" question. Fusion system requires the definition of PM for designing the system. This characteristic has received relatively less attention relative to its importance [29]. This is very difficult to conceive a global PM of the fusion system admittedly. In terms of defining the Fusion Algorithm Measure of Effectiveness (FAME), a beginning has been made in this direction in [27].
based on information fusion approach is researched for early diagnosis of Alzheimer's disease [82]. A contextaware fusion system for enhancing healthcare privacy is presented in [73]. In [74], a robust framework, which integrates an ear-worn activity recognition sensor with ambient blob-based vision sensors, is proposed for enhanced activity recognition.
6.1.3 Multiple Classifiers (MC) Diversity and fusion of multiple classifiers (MC) are necessary for fusing information. Despite of the proved utility of multiple classifier systems, no general answer to questions about the possibility of exploiting the strengths while avoiding the weakness of different classifier designs has yet emerged [81]. Diversity is not related to the ensemble accuracy unequivocally [18]. 6.1.4 Multiple Metrics (MM) We can apply multiple metrics into image fusion as a fusion process method. MM is suited for cases where the feature values are continuous variables [17]. Fusion at higher levels designed for situation awareness can be expected to encounter a spectrum of the different types of features, which include integer valued and symbolic valued data, which require more complex metrics [66]. The main objective of image fusions is to minimize the uncertainty and the redundancy while maximizing the relevant information particular to applications [41].
Figure 7. Issues in Fusion Techniques
6. Issues and Research Directions 6.1 Issues As shown in Figure 7, we deal with fusion issues as sensor managements (SM), performance measurements (PM), multiple classifiers (MC), multi-metrics (MM), knowledge discovery (KD), ensemble methods (EM), and hybrid sensor networks (HSN).
6.1.5 Knowledge Discovery (KD) KD is a potential objective multisensor information fusion process [92]. Data mining can be viewed as a means of accomplishing the objective of the KC. It is generally associated with only a single data source. A real intelligent decision system should able to not only fuse the information coming from multiple sensors, but offer the conceptual information and knowledge [47].
6.1.1 Sensor Management (SM) SM will give a sense of the control over the sensors and brings out an efficient use of the sensors. It brings all sensors into one family. It combines sensor systems to accomplish specific and dynamic mission objectives [71]. The ultimate goals are to optimize the overall performance of the tracking or the classification system in situation aware environments as described in [12].
6.1.6 Ensemble Methods (EM) There is no perfect match whenever we try to match models to information. So, a reasonable match requires that models have to be modified to fit the data. This requires that models and data have to be considered in various combinations. Numbers of combinations grows
6.1.2 Performance Measurements (PM) Many questions about concepts such as probability, statistics, decision, estimation, approximate reasoning techniques, image processing and pattern recognition
82
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
very fast with the complexity of model and the amount of data. In [15], [78], powerful new mathematical ideas are researched. EM issues, which come from modern data mining problems, are necessary for more reliable data processing. The data consists of distributed, noisy, and sampled variations on non-stationary environments. EM should examine different applications that lead to these modern data mining problems and how novel ensemble methods aid in solving these problems [72], [89].
A fault diagnosis and detection has become one of critical aspects of the advanced fusion system design in pervasive computing areas. Failures normally produce the change in system dynamic and pose a significant risk [5]. Some multiple sensor fault detection methods, which include distributed online methods, neural fuzzy networks, or Bayesian methods are proposed currently to increase the reliability of information [50], [63], [86].
6.1.7 Hybrid Sensor Networks (HSN) Fusion assists in the seamless integration of smart sensors, actuators, devices and capabilities, softwarehardware agents, and sensor network applications for emerging hybrid sensor networks (HSN). It facilitates the information networks by integrating sensor nodes, distributed algorithms, different communication, and network architectures towards timely and intelligent decision making [11], [39]. Figure 8. Research directions of Fusion Techniques for reliable contextual information
6.2 Research Directions
6.2.3 Robust, Consensus and Fault Tolerant Fusion
There are many efforts of researchers to progress fusion methods. Among them, we will introduce four important directions of fusion techniques as shown in Figure 8. These categories are concurrently considered by researchers to obtain reliable contextual information in pervasive computing areas.
Robust, consensus, or fault-tolerant is an important factor for research on fusion. Precise information about the sensed environment is scarce in the real world then the sensors are not always perfectly functional even if a multi-sensor fusion requires exact information about the sensed environment. A robust, consensus, or faulttolerant fusion technique should apply to the pervasive system. Some research progresses are performed. For instance, a robust fusion algorithm in the presence of the various forms of uncertainty is necessary in fusion systems [31], [37], [84]. The practical activity maps information in [58] shows ambient intelligence-related context information gathered from both humans and their surrounding environments. Principal strategies of the incorporating reliability into fusion operators are over-viewed in [87]. Fault-tolerant information fusion algorithms are proposed based on the generic statement of the Sensor Managements in [75].
6.2.1 Multi-Level Sensor Fusion There is a clear need for multi-level sensor fusion for advanced systems, which have a higher robustness and flexibility [49], [64], since a single level fusion limits the capacity and the robustness of the systems due to the weaknesses caused by uncertainty, missing observation, or incompleteness of the sensors. To make this goal, a general architecture is designed for decision makings from fusion levels of the time-varying data, features and decisions. A low-level fusion fuses multisensor raw data. A medium level fusion fuses low level data and the feature data to obtain fused feature. And last, a high level fusion fuses a feature to obtain a final decision. As shown in Figure 1, the hierarchical layer composed of data capture layer, data processing layer, and data analysis layer is needed for fusion [83].
6.2.4 Conflict Managements Conflict managements are performed to reduce the error rates of information. Different measurements and methods have been proposed for the fusion. A general unified decision theory is still missing in the areas of decision making from reasoning under uncertainty for
6.2.2 Fault Diagnosis and Detection
83
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
all kinds of sources. The management and combination of high conflicting sources of information is difficult to solve because of uncertain, imprecise, and paradoxical sources of information [111]. So, Dezert-Smarandache Theory (DSmT) is proposed to solve these problems when conflict between sources becomes a higher and information becomes ambiguous and imprecise [32]. A generalization of the conflict repartition rules and the mixed combination rule (conjunctive and disjunctive rule) may be discussed in belief function theory. The unification of the evidence theoretic fusion algorithms, the combination of the particle filtering and the DSmT for conflict information handling, and a qualitative belief condition rule are studied from many researchers to solve complex, static or dynamic fusion problems beyond the limits of the DST [62], [91], [95], [102]. As one of researches, we also work on evidential networks for multisensor context reasoning in home-based care [51], [52], [53] to reduce the conflict information in uncertainty levels as shown in Figure 9.
8. Acknowledgements This research was supported by a grant (07High Tech A01) from High Tech Urban Development Program funded by Ministry of Land, Transportation and Maritime Affairs of Korean Government.
9. References [1] R. A. and R. H. Campbell, “An infrastructure for context awareness based on first order logic,” Personal and Ubiquitous Computing, Vol. 7, No. 6, 2003, pp. 353-364 [2] S. Agarwal, A. Joshi, T. Finin, Y. Yesha and T. Ganous, “A pervasive computing system for the operating room of the future,” Mobile Networks and Applications, Vol. 12, No. 2, 2007, pp. 215-228 [3] K. Baclawski, M. Kokar, C. J. Matheus, J. Letkowski and M. Malczewski, “Formalization of situation awareness,” the 11th OOPSLA workshop on Behaviors Semantics, 2002, pp. 1-15 [4] C. A. Barlow, L. D. Stone, and M. V. Finn, “Unified data fusion,” Proc. of the 9th National Symposium on Sensor Fusion, Vol. 1, 1996, pp. 321-330 [5] O. Basir and X. Yuan, “Engine fault diagnosis based on multi-sensor information fusion using Dempster-Shafer evidence theory,” Information Fusion, Vol. 8, No. 4, 2007, pp. 379-386 [6] L. R. Beach, “Image theory: Decision Making in Personal & Organization Contexts,” John Wiley & Sons Inc., 1990 [7] M. Beiql, M. Beuster, D. Rohr, T. Riedel, C. Decker and A. Krohn, “S2b2: Blackboard for transparent data and control access in heterogeneous sensing systems,” the 4th Intl. Conf. on Networked Sensing Systems (INSS), 2007, pp. 126-129 [8] S. Benferhat, D. Dubois, S. Kaci, and H. Prade, “Bipolar possibility theory in preference modeling: representation, fusion and optimal solutions,” Information Fusion, Vol. 7, No. 1, 2006, pp. 135-150 [9] M. Beynon, B. Curry, and P. Morgan, “The DempsterShafer theory of evidence: an alternative approach to multicriteria decision modeling,” Omega, Vol. 28, No. 1, 2000, pp. 37-50 [10] E. Biglieri and M. Lops, “Multiuser detection using random-set theory,” EUSIPCO 2006 Special Session on 'Multi-user MIMO Communications, 2006 [11] P. K. Biswas, H. Qi, and Y. Xu, “Mobile-agent based collaborative sensor fusion,” Information Fusion, Vol. 9, No. 3, 2008, pp. 399-411 [12] S. Blackman and R. Popoli,” Design and Analysis of Modern Tracking Systems,” Artech House, 1999 [13] E. Blasch, “Assembling information fused human computer cognitive decision making tool,” IEEE Aerospace and Electronic Systems Magazine, 2000, pp. 11-17 [14] E. Blasch, “Ontological issues in higher level of information fusion: User refinement of the fusion,” the 6th Intl. Conf. on Information Fusion, 2003, pp. 634-641
Figure 9. A Proposed Multisensor Context Reasoning based on Evidential Networks
7. Conclusions We survey for data, multi-sensors, and information fusions in pervasive computing application areas to get reliable contextual information. In general, fusion is a critical technique to estimate and infer the situation of human in different smart applications. However, there is no general accepted approach in those applied areas. Therefore, we review some generated concepts, models, algorithms, theories, systems, and applications in this survey. Moreover, we discussed famous current issues and suggested research directions for fusion techniques, which reduce uncertainty and improve the accuracy of fusion in pervasive environments.
84
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
[32] J. Dezert and F. Smarandache, “An introduction to the DSm theory for the combination of paradoxical, uncertain and imprecise sources of information,” Octagon Mathematical Magazine, Vol. 15, No. 2, 2005, pp. 681722 [33] D. Dubois and H. Prade, “Possibility theory, probability theory and multiple-valued logics: Clarification,” Annals of Mathematics and AI, Vol. 32, No. 1-4, 2001, pp. 35-66 [34] D. Dubois, H. Prade, and R. Sabbadin, “Decisiontheoretic foundation of qualitative possibility theory,” European Journal of Operational Research, Vol. 128, No. 3, 2001, pp. 459-478 [35] M. R. Endsley, “Toward a theory of situation awareness in dynamic systems,” Human Factors, Vol. 37, No. 1, 1995, pp. 32-64 [36] J. Esteban, A. Starr, R. Willetts, P. Hannah, and P. Bryanston-Cross, “A review of data fusion models and architectures: Towards engineering guidelines,” Neural Computing and Applications, Vol. 14, No. 4, 2005, pp. 273-281 [37] V. K. Garg and V. Ogale, “Fusible data structures for fault tolerance,” Proc. of 27th Intl. Conf. on Distributed Computing Systems, 2007, pp. 20-44 [38] I. Goll and A. Rasheed, “Rational decision making & firm performance: the moderating role of environment,” Strategic Management Journal, Vol. 18, No. 7, 1997, pp. 583-591 [39] J. Gomes and H. Choi, “Adaptive optimization of join threes for multi-join queries over sensor streams,” Information Fusion, Vol. 9, No. 3, 2008, pp. 412-424 [40] I. Goodman, R. Mahler, and H. Nguyen, “Mathematics of data fusion,” Kluwer Academic Publisher, 1997. [41] A. A. Goshtashby, “Image fusion: Advances in the state of the art,” Information Fusion, Vol. 8, No. 2, 2007, pp. 114-118 [42] D. L. Hall and J. Llinas, “An introduction to multisensor data fusion,” the IEEE, Vol. 85, No. 1, 1997, pp. 6-23 [43] D. L. Hall and S. McMullen, “Mathematical techniques in multisensor data fusion,” Artech House, 2004 [44] G. Hang and Y. Min, “Data fusion in distributed multisensor system,” Geo-Spatial Information Science, Vol. 7, No. 3, 2004, pp. 214-217 [45] X. Hong, “Valuation-based systems for decision analysis using belief functions,” Decision Support Systems, Vol. 20, No. 2, 1997, pp. 165-184 [46] H. Hu, S. Qian, J. Wang, and Z. Shi, “Application of information fusion technology in the remote state online monitoring and fault diagnosing system for power transformer,” 8th Intl. Conf. on Electronic Measurement and Instruments, 2007, pp. 550-555 [47] A. Hunter and W. Liu, “Fusion rules for merging uncertain information,” Information Fusion, Vol. 7, No. 1, 2006, pp. 97-134 [48] C. Kahraman, G. Buyukozkan, and N. Y. Ates, “A two phase multi-attribute decision making approach for new product introduction,” Information Science: International Journal, Vol. 177, No. 7, 2007, pp. 1567-1582
[15] N. A. Bomberger, A. M. Waxman, B. J. Rhodes, and N. A. Sheldon, “A new approach to higher-level information fusion using associative learning in semantic network of spiking neurons,” Information Fusion, Vol. 8, No. 3, 2007, pp. 227-251 [16] E. Bosse, J. Roy, and S. Paradis, “Modeling and simulation in support of design of a data fusion system,” Information Fusion, Vol. 1, No. 2, 2000, pp. 77-87 [17] E. Bosse, P. Valin, A. Boury-brisset, and D. Grenier, “Exploitation a priori knowledge for information fusion,” Information Fusion, Vol. 7, No. 2, 2006, pp. 161-175 [18] G. Brown, J. Wyatt, R. Harris, and X. Yao, “Diversity creation methods: a survey and categorization,” Information Fusion, Vol. 6, No. 1, 2005, pp. 5-20 [19] A. Bundy, “Incidence calculus: A mechanism for probabilistic reasoning,” Journal of automated reasoning, Vol. 1, No. 3, 1985, pp. 263-283 [20] D. A. Bushey and M. J. Forsyth, “The recognitionprimed decision model: An alternative to the MDMP for GWOT,” FA Journal, 2006 [21] F. Campos and S. Calvalcante, “An extended approach for Dempster-Shafer theory,” IEEE Intl. Conf. on Information Reuse and Integration, 2003, pp. 338-344 [22] G. A. Carpenter, S. Martens, and O. J. Ogas, “Selforganizing hierarchical knowledge discovery: a new framework using ARTMAP neural networks,” Neural Networks, Vol. 18, No. 3, 2005, pp. 287-295 [23] C. Cayrol, V. Royer, and C. Saurel, “Management of preferences in assumption-based reasoning,” Advanced Methods in AI (LNCS), Vol. 682, 1993, pp. 13-22 [24] S. Chen and Y. W. Jen, “Data fusion neural network for tool condition monitoring in CNC milling machining,” Intl. Journal of Machine Tools and Manufacture, Vol. 40, No. 3, 2000, pp. 381-400 [25] L. Cholvy and A. Hunter, “Information fusion in logic: A brief overview,” Qualitative and Quantitative Practical Reasoning (LNCS), Vol. 1244, 1997, pp. 86-95 [26] J. Clark and A. Yullie, “Data fusion for sensory information processing system,” Kluwer Academic Publication, 1990 [27] B. Dasarathy, “Information fusion benefits delineation under off-nominal scenarios. Proc. of the SPIE, Vol. 3719, No. 2, 1999, pp. 2-13 [28] B. Dasarathy, “Information fusion - what, where, why, when and how,” Information Fusion, Vol. 2, No. 2, 2001, pp. 75-76 [29] B. Dasarathy, “Fusion system: How good is it,” Information Fusion, Vol. 2, No. 3, 2001, pp. 161-162 [30] D. De-Brucg, O. Colot, and A. Sombo, “Identical foundation of probability theory and fuzzy set theory,” the 5th Intl. Conf. on Information Fusion, Vol. 2, 2002, pp. 1442-1449 [31] D. Desovski, Y. Liu, and B. Cukic, “Linear randomized voting algorithm for fault tolerant sensor fusion and the corresponding reliability model,” IEEE Intl. symposium on High-Assurance Systems Engineering, 2005, pp. 153162
85
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
[49] A. Klausner, B. Rinner, and A. Teng, “I-sense: Intelligent embedded multi-sensor fusion,” Proc. of Intl. workshop on intelligent solutions in embedded systems, 2006, pp. 1-12 [50] J. Korbicz and M. Kowal, “Neuro-fuzzy networks and their application to fault detection of dynamical systems,” Engineering Applications of AI journal, Vol. 20, No. 5, 2007, pp. 607-617 [51] H. Lee, J. Choi, and R. Elmasri, “A classification and modeling of the quality of contextual information in smart spaces,” 7th IEEE Intl. Conf. on PerCom Workshop on Context Modeling and Reasoning (CoMoRea), 2009. [52] H. Lee, J. Choi, and R. Elmasri, “A dynamic evidential network for multisensor context reasoning in home-based care,” the IEEE 8th Intl. Conf. on Systems, Man, and Cybernetics (SMC 2009), 2009 [53] H. Lee, J. Choi, and R. Elmasri, “Sensor data fusion using DSmT for activity recognition under uncertainty in home-based care,” the IEEE 23rd Intl. Conf. on Advanced Information Networking and Applications, 2009 [54] A. Lemmi and G. Betti, “Capability approach and fuzzy set theory: description, aggregation and inference issues,” Fuzzy Set Approach to Multidimensional Poverty Measurement, 2006, pp. 93-113 [55] D. Lewis, “Probabilities of conditionals and conditional probabilities,” the Philosophical Review, Vol. 85, No. 3, 1976, pp. 297-315 [56] J. Lindner, R. Murphy, and E. Nitz, “Learning the expected utility of sensor and algorithms,” the IEEE Intl. Conf. on multisensor fusion and integration for intelligent systems, 1004, pp. 583-590 [57] J. Llinas, C. Bowman, G. Rogova, A. Steinberg, and F. White, “Revisiting the JDL data fusion model II,” the 7th Intl. Conf. on Information Fusion, 2004, pp. 1218-1230 [58] C. Lu, Y. Ho, and L. Fu,” Creating robust activity maps using wireless sensor network in a smart home,” IEEE Intl. Conf. on Automation Science and Engineering, 2007, pp. 741-746 [59] R. Luo, C. Yih, and K. Su, “Multisensor fusion and integration: approaches, applications, and future research directions,” IEEE Sensors Journal, Vol. 2, No. 2, 2002, pp. 107-119 [60] D. Macii, A. Boni, M. D. Cecco, and D. Petri, “Tutorial 14: Multisensor data fusion,” IEEE Instrumentation and Measurement Magazine, Vol. 11, No. 3, 2008, pp. 24-33 [61] D. Mandic, D. Obradovic, A. Kuh, T. Adali, U. Trutschel, M. Golz, P. Wilde, J. Barria, A. Constantinides, and J. Chambers, “Data fusion for modern engineering applications: An overview,” Intl. Conf. on Artificial Neural Networks (LNCS), Vol. 3697, 2005, pp. 715-721 [62] A. Martin and C. Osswald, “Toward a combination rule to deal with partial conflict and specificity in belief functions theory. 10th Intl. Conf. on Information Fusion, 2007, pp. 1-8 [63] J. P. Matsuura and T. Yoneyama, “Learning Bayesian networks for fault detection,” Proc. of 14th IEEE Signal Processing Workshop, 2004, pp. 133-142
[64] C. McCullough, B. Dasarathy, and P. Lindberg, “Multilevel sensor fusion for improved target discrimination,” Proc. of the 35th IEEE Decision and Control, Vol. 4, 1996, pp. 3674-3675 [65] S. Mohammed and E. Ringseis, “Cognitive diversity and consensus in group decision making: The role of inputs, processes, and outcomes,” Organizational Behavior and Human Decision Processes, Vol. 85, No. 2, 2001, pp. 310-335 [66] A. Motro and P. Anokhin, “Fusionplex: Resolution of data inconsistencies in the integration of heterogeneous information sources,” Information Fusion, Vol. 7, No. 2, 2006, pp. 176-196 [67] Z. P. Mourelatos and J. Zhou, “Reliability estimation and design with insufficient data based on possibility theory,” AIAA Journal, Vol. 43, No. 8, 2005, pp. 16961705 [68] S. Mulgund, G. Rinkus, C. Illgen, and J. Friskie, “OLIPSA: On-line intelligent processor for situation assessment,” the 2nd Annual Symposium and Exhibition on situational awareness in the tactical air environment, 1997. [69] K. Murphy, “Dynamic Bayesian network: representation, inference and learning,” Ph. D. Dissertation, Univ. of California, Berkeley, 2002 [70] E. F. Nakamura and A. Loureiro, “Information fusion in wireless sensor networks,” Intl. Conf. on Management of data (ACM SIGMOD), 2008, pp. 1365-1372 [71] G. W. Ng and K. H. Ng, “Sensor management - what, why and how,” Information Fusion, Vol. 1, No. 2, 2000, pp. 67-75 [72] N. C. Oza and K. Tumer, “Classifier ensembles: Select real-world applications,” Information Fusion, Vol. 9, No. 1, 2008, pp. 4-20 [73] G. Pallapa, N. Roy, and S. Das, “Precision: privacy enhanced context-aware information fusion in ubiquitous healthcare,” 1st Intl. Workshop on Software Engineering for Pervasive Computing Applications, Systems, and Environments, 2007, pp. 1-10 [74] J. Pansiot, D. Stoyanov, D. McIlwraith, B. Lo, and G. Z. Yang, “Ambient and wearable sensor fusion for activity recognition in healthcare monitoring systems,” 4th Intl. Workshop on Wearable and Implantable Body sensor network, 2007, pp. 208-212 [75] L. Paradis and Q. Han, “A survey of fault management in wireless sensor networks,” Journal of Network and Systems Management, Vol. 15, No. 2, 2007, pp. 171-190 [76] Y. Peng, C. Lin, and M. Sun, “A distributed multimodality sensor system for home-used sleep condition inference and monitoring,” 1st Trans-Disciplinary Conf. on Distributed Diagnosis and Home Healthcare, 2006, pp. 20-23 [77] P. Perez, J. Vermaak, and A. Blake, “Data fusion for visual tracking with particles,” the IEEE, Vol. 92, No. 3, 2004, pp. 495-513 [78] L. Perlvsky, “Concurrent learning and fusion,” Information Fusion, Vol. 8, No. 3, 2007, pp. 224-226
86
International Journal of Digital Content Technology and its Applications Volume 4, Number 2, April 2010
[97] S. Thiriot and J. Kant, “Modeling one human decision maker with a multi-agent system: the CODAGE approach. Intl. Conf. on Autonomous Agents, 2006, pp. 50-57 [98] J. Thomas, W. R. Center, and Y. Height, “Visual data fusion for applications of high-resolution numerical weather prediction,” the IEEE Visualization Conference, 2000, pp. 477-480 [99] P. Todd and G. Gigerenzer, “Putting naturalistic decision making into the adaptive toolbox,” Journal of Behavioral Decision Making, Vol. 14, No. 5, 2001, pp. 353-384 [100] V. Torra and Y. Narukawa, “Modeling decisions,” Springer, 2007 [101] P. Varshney, “Multi-sensor data fusion,” Electronics and Communication Engineering Journal, Vol. 9, No. 6, 1997, pp. 245-253 [102] M. Vatas, R. Singh, A. Noore, and M. Houch, “Quality-augmented fusion of level-2 and level-3 fingerprint information using DSmT,” the Intl. Journal of Approximate Reasoning, Vol. 50, No. 1, 2007, pp. 51-61 [103] T. Vercauteren, D. Guo, and X. Wang, “Joint multiple target tracking and classification in collaborative sensor networks,” the IEEE journal on Selected Areas in Communications, Vol. 23, No. 4, 2005, pp. 714-723 [104] F. E. White, “A model for data fusion,” the 1st National Symposium on Sensor Fusion, Vol. 2, 1988, pp. 149-158 [105] H. Wu, M. Siegel, R. Stiefelhagen, and J. Yang, “Sensor fusion using Dempster-Shafer theory,” 19th IEEE Conf. on Instrumentation and Measurement Technology, Vol. 1, 2002, pp. 7-12 [106] J. K. Wu and Y. F. Wong, “Bayesian approach for data fusion in sensor networks,” 9th Intl. Conf. on Information Fusion, 2006, pp. 1-5 [107] N. Xiong and P. Svensson, “Multi-sensor management for information fusion: issues and approaches,” Information Fusion, Vol. 3, No. 2, 2002, pp. 163-186 [108] X. Xu, G. Ge, and C. Wen, “A fusion algorithm of multisensor system based on sequential filtering in distributed network,” the IEEE Intl. Conf. on Automation and Logistics, 2007, pp. 2325-2328 [109] Y. Xu, L. Wang, and Y. Sheng, “Rough set theorybased multi-class decision attribute reduction algorithm and its application,” 5th IEEE Intl. Conf. on Cognitive Informatics, Vol. 2, 2006, pp. 722-725 [110] Z. Yunmin and L. X. Rong, “Extended DempsterShafer combination rules based on random set theory,” Proc. of SPIE, Multisensor and Multisource Information Fusion: Architectures, Algorithms, and Applications, Vol. 5434, 2004, pp. 112-120 [111] L. A. Zadeh, “Toward a unified theory of uncertainty,” IPMU conference, 2004 [112] Y. Zhang and Q. Ji, “Active and dynamic information fusion for multi-sensor systems with dynamic Bayesian networks,” IEEE Transaction on Systems, Mans, and Cybernetics, Vol. 36, No. 2, 2006, pp. 467-472 [113] X. Zhao, S. Yuan, Z. Yu, W. Ye, and J. Cao, “Designing strategy for multi-agent system based large structural health monitoring,” Expert Systems with
[79] R. W. Pew and A. S. Mavor, “Modeling human and organizational behavior: application to military simulations,” the National Academy Press, 1998 [80] G. Phillips-Wren, N. Lchalkaranje, and L. Jain, “Intelligent decision making: An AI-based approach,” Springer, 2008 [81] R. Polikar, D. Parikh, and S. Mandayam, “Multiple classifier systems for multisensor data fusion,” the IEEE Sensors Applications Symposium, 2006, pp. 180-184 [82] R. Polikar, A. Topalis, D. Parikh, D. Green, J. Frymiare, J. Kounios, and C. Clark, “an ensemble based data fusion approach for early diagnosis of Alzheimer's disease,” Information Fusion, Vol. 9, No. 1, 2008, pp. 83-95 [83] S. Pop, O. Lavialle, R. Terebes, and M. Borda, “Lowlevel fusion: a PDE-based approach,” 10th Intl. Conf. on Information Fusion, 2007, pp. 1-8 [84] H. Z. Qiu, H. Y. Zhang, and H. Jin, “Fusion algorithm of correlated local estimates,” Aerospace Science and Technology, Vol. 8, No. 7, 2004, pp. 619-626 [85] J. Quiggin, “Regret theory with general choice sets,” Journal of Risk and Uncertainty, Vol. 8, No. 2, 1994, pp. 153-165 [86] R. Rajagopal, X. Nguyen, S. C. Ergen, and P. Varaiya, “Distributed online simultaneous fault detection for multiple sensors,” Intl. Conf. on Information processing in sensor networks, 2008, pp. 133-144 [87] G. L. Rogova and V. Nimier, “Reliability in information fusion: Literature survey,” 7th Intl. Conf. on Information Fusion, 2004, pp. 1158-1165 [88] E. Shahbazian, G. Rogova, and P. Valin, “Data fusion for situation monitoring, incident detection, alert and response management,” NATO Science Series, 2005 [89] L. Shoemaker, R. Banfield, L. Hall, K. Bowyer, and W. Kegelmeyer, “using classifier ensembles to label spatially disjoint data,” Information Fusion, Vol. 9, No. 1, 2008, pp. 120-133 [90] A. Shuster, “Intelligent computing everywhere,” Springer, 2007 [91] F. Smarandache and J. Dezert, “Qualitative belief conditioning rules,” the 10th Intl. Conf. on Information Fusion, 2007, pp. 1-8 [92] P. R. Smart, N. R. Shadbolt, L. Carr, and M. Schraefel, “Knowledge-based information fusion for improved situational awareness,” the 8th Intl. Conf. on Information Fusion, Vol. 2, 2005, pp. 1-8 [93] P. Smets, “Data fusion in the transferable belief model,” the 3rd Intl. Conf. on Information Fusion, 2000, pp. 21-23 [94] L. D. Stone, C. A. Barlow, and T. L. Corwin, “Bayesian multiple target tracking,” Artech House, 1999 [95] Y. Sun and L. Bentabet, “A sequential Monte-Carlo and DSmT based approach for conflict handling in case of multiple targets tracking,” Image Analysis and Recognition (LNCS), Vol. 4633, 2007, pp. 526-537 [96] R. J. Tait, A. Hopgood, and G. Sheaefer, “Registration and fusion of multi-sensor data using multiple agents,” Proc. of SPIE, the Intl. Society for Optical Engineering, Vol. 5823, 2005, pp. 104-112
87
Fusion Techniques for Reliable Information: A Survey Hyun Lee, Byoungyong Lee, Kyungseo Park and Ramez Elmasri
Applications: An Intl. Journal, Vol. 34, No. 2, 2008, pp. 1154-1168 [114] J. Zhou, Y. Zhu, Z. You, and E. Song, “An efficient algorithm for optimal linear estimation fusion in distributed multisensor systems,” IEEE Transaction on Systems, and Man and Cybernetics, Vol. 36, No. 2, 2006, pp. 1000-1009 [115] X. Zhou, D. Comaniciu, and A. Gupta, “An information fusion framework for robust shape tracking,” IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol. 27, No. 1, 2005, pp. 115-129
88
