Machine Perception in Automation:A Call to Arms
© Dietmar Bruckner et al. 2011
Received: 28 June 2010
Accepted: 16 January 2011
Published: 23 January 2011
Automation systems strongly depend on the amount, quality, and richness of their sensor information. For decades, scientists have investigated towards more accurate and cheaper sensors as well as new sensors for previously undetectable properties or substances. With these enhancements the problem of too complex sensor information and sensor fusion raised. This paper is intended for giving a retrospection on perception systems in automation, followed by reviewing state-of-the-art approaches for handling diverse and complex sensor information as well as highlighting future requirements for more human-like systems that have the ability of performing their actions in complex and unpredictable environments. For the latter requirement, a section introducing a number of agent architectures for embedding the sensor fusion process into a comprehensive decision-making unit is given.
The imaginary accumulated by science fiction through the years has always fancied future worlds full of intelligent autonomous machines that were able to perceive the reality around. This ability of sensing the environment in which they were immersed allowed robots to act consequently. Decades ago, back in the true world, automation systems designers started to follow the path that the literature had drafted realizing that intelligence and perception go hand in hand. For instance, very simple devices can only carry out tasks that do not require perceiving and interacting with the real world (say moving a certain piece 5 cm ahead in a conveyor belt). Such an easy process may be seen like driving blind a car: in case the piece falls off the belt, the device will just fail to carry out its task being also unable to find a proper reason for its failure. All living beings have incorporated diverse sensing parts and strategies in their ascent of the evolution chain, and so do machines: perception is the turning point that allows automation systems to collect the information needed to control their actions and the consequences of them.
In the following sections, we will provide an overview on how machine perception has been addressed in the past, and what the promising approaches for perception in future automation systems are in order to be able to fulfill useful tasks in more general environments—as humans can. Additionally, a section introducing the embedding of the information processing part into a whole decision-making framework is presented. Afterwards, 4 different approaches will be highlighted in detail that are currently developed in research projects. Those address mainly two types of perceptive activities: related to human activities and related to machinery routines as for example, in a factory. The first two approaches are generally in nature, but in their first description intended for the first kind of tasks, the third one is targeted to factory automation, whereas the approach detailed in Section 8 is applicable to any kind of perceptive task. This compilation of material will inspire engineers that are engaged in designing automation devices in order to help them create more intelligent or more flexible devices through better perception of their environment. As the title implies, their exist approaches to overcome limitations of previous methods of sensor data processing, they only need to be applied.
2. Developments and Visions
The term perception has been used in computer and automation systems from the 1950's onwards, since the foundation of Artificial Intelligence (AI). It was seen as one of the components of intelligence, being learning, reasoning, problem-solving, and language-understanding. Perception means acquiring, interpreting, selecting, and organizing sensory information. The topic itself was not new to automation, but has gained a new quality from the moment information processing could be separated from energy flow and performed in completely new ways.
The development of machine perception has taken two ways. The first one is related to industrial process control, where machines are designed and built in order to increase productivity, reduce costs as well as enhance quality and flexibility in the production process. These machines mostly need to perceive a well-known environment and therefore possess a selected number of dedicated (and reliable, robust, expensive, etc.) sensors. The sum of sensor views composes the machine's view of the world.
The second development path is concerned with perception of humans and human activities, on the one hand, and with implementing perception systems imitating human perception for broader application areas, on the other. Involved research fields are, among others, cognitive sciences, artificial intelligence, image processing, audio data processing, natural language processing, user interfaces, and human-machine interfaces.
The research field related to perceiving information about human users is called context-aware systems. The common view in this community is that computers will not only become cheaper, smaller, and more powerful, but they will also more or less disappear and hide integrated in normal, everyday objects [1, 2]. Thus, smart objects will communicate, cooperate, and virtually amalgamate without explicit user interaction or commands to form consortia in order to offer or even fulfill tasks on behalf of a user. They will be capable of not only sensing values, but also of deriving context information about the reasons, intentions, desires, and beliefs of the user. This information may be shared over networks—like the internet—and used to compare and classify activities, find connections to other people and/or devices, look up semantic databases, and much more.
3. The Research Field of Sensor Fusion
One of the most active disciplines around autonomous perception is sensor fusion. This research area aims at combining sensorial data from diverse origins (and sometimes also other information sources) to achieve a "better perception" of the environment.
There can be found various definitions of sensor fusion differing slightly in the meaning. One states that sensor fusion is "the combining of sensory data or data derived from sensory data in order to produce enhanced data in form of an internal representation of the process environment. The achievements of sensor fusion are robustness, extended spatial and temporal coverage, increased confidence, reduced ambiguity and uncertainty, and improved resolution." , to which we fully agree.
Sensor data fusion is a relatively recent and dynamic field, and a standard terminology has not yet been adopted. The terms "sensor fusion", "sensor integration", "data fusion", "information fusion", "multisensor data fusion", and "multisensor integration" have been widely used in technical literature to refer to a variety of techniques, technologies, systems, and applications, which use data derived from multiple information sources [4–6].
Data for sensor fusion can come from single sensors taken from multiple measurements subsequently at different instants of time, from multiple sensors of identical types, or from sensors of different types. In the following, concepts, models, methods, and applications for sensor fusion will be summarized, mainly following the ideas of [7, 8].
Concepts for Fusion
Sensor fusion is generally based on the combination of redundant or complementary information. Among others, the works in [3, 5, 8] distinguish three types of sensor data fusion, which are not mutually exclusive: complementary fusion, competitive fusion, and cooperative fusion.
Complementary fusion is the fusion of incomplete sensor measurements from several disparate sources. Sensor data do not directly depend on each other, but are combined to give a more complete image of a phenomenon under observation.
Competitive fusion is the fusion of redundant sensor measurements from several sources. Each sensor delivers independent measurements of the same property. Competitive sensor configurations are also called redundant configurations.
Cooperative fusion uses the information provided by independent sensors to derive information that would not be available from the single sensors. An example for cooperative sensor fusion is stereovision. In contrast to complementary and competitive fusion, cooperative fusion generally decreases accuracy and reliability.
Models for Fusion
Regarding the models for sensor fusion, it has to be noted that sensor fusion models heavily depend on the application they are used in. So far, there does not exist a model for sensor fusion that is generally accepted, and it seems unlikely that one technique or architecture will provide a uniformly superior solution . Therefore, there exist numerous models for sensor fusion in the literature. To mention only few of them: the JDL fusion model architecture, the Waterfall model, the Intelligence cycle, the Boyd loop, the LAAS architecture, the Omnibus model, Mr. Fusion, the DFuse framework, and the Time-Triggered Sensor Fusion Model.
Methods for Fusion
There have been suggested various methods for sensor fusion. Sensor fusion methods can principally be divided into grid-based (geometric) and parameter-based (numerical) approaches whereby in the case of numeric approaches. A further distinction is made between feature-based approaches (weighted average, Kalman filter), probabilistic approaches (classical statistics, Bayesian statistics), fuzzy methods, and neural approaches. In contrast, the work in  classifies fusion algorithms into estimation methods (weighted average, Kalman filter), classification methods (cluster analysis, unsupervised or self-organized learning algorithms), interference methods (Bayesian interference, Dempster-Shafter evidential reasoning), and artificial intelligence methods (neural networks, fuzzy logic). Similar to the models of sensor fusion, there is also no one sensor fusion method suitable for all applications. Hence, new hierarchical approaches are sought to combine the advantages of the basic mathematical ones.
Areas of applications of fusion are broad and range from measurement engineering and production engineering over robotics and navigation to medicine technology and military applications. Examples for applications can be found in [4, 8, 9].
Biological Sensor Fusion
It is well appreciated that sensor fusion in the perceptual system of the human brain is of far superior quality than sensor fusion achieved with existing mathematical methods [10, 11]. Therefore, it seems to be particularly useful to study biological principles of sensor fusion.
Such studies can, on the one hand, lead to better technical models for sensor fusion and, on the other hand, to a better understanding of how perception is performed in the brain. Sensor fusion based on models derived from biology is called biological sensor fusion. Approaches to biological sensor fusion made so far can be found in [12–18].
Although there have already been introduced a number of models for biological sensor fusion, yet success of research efforts incorporating lessons learned from biology into "smart algorithms" has been limited . One reason therefore might be that the use of biological models in actual machines is often only metaphorical, using the biological architecture as a general guideline .
4. Agent Architectures
The development in AI as briefly sketched in Section 2 can be summarized to have taken four main scientific directions, the so-called symbolic, statistical, emotional, and behavior-based AI . In symbolic AI sensor inputs are abstracted to "symbols" and then processed. Symbolic AI's major concern is knowledge representation and the modeling of search algorithms for identifying situations. Statistic AI is used for applications where the problem space cannot be defined and in dynamic or unrestricted environments. The claim  that human decision-making is influenced by subjective evaluation based on emotion is taken into account by emotional AI, while behavior-based AI focuses on observable system world interaction.
The different theories overlap in practice. Based on these theories a number of control architectures and frameworks have been developed. They are applied for systems which must be able to accomplish tasks by harking back on predefined and learned knowledge.
The embodied approach to AI pioneered by Brooks and his subsumption architecture follows the paradigm that mobile agents need to have a body as origin for decisions . From a cognitive scientist's perspective, it contributes to the idea that intelligence can arise or emerge out of a large number of simple, loosely coupled parallel processes [23, 24].
5. Recognizing Scenarios with Statistical Methods
The following four approaches to machine perception have been selected because they give a representative overview about the principles in advanced machine perception methods. All of them have a layered data processing architecture that allows hierarchical information processing. This is necessary for complex processes. (Additionally, there are approaches for formalizing the hierarchical representations in taxonomies or ontologies. Their introduction is not within the scope of this paper since we would like to give an introduction to the mechanisms of perception and not to focus on the organization of results.) For introductory purposes, they are not presented including parameter learning capabilities (except the Automatic Scenario Learning approach from this Section, which, however, is also not intended to change parameters after the initial structural training phase).
Scenario recognition tries to find sequences of particular behaviors and groups it in a way humans would according to similarities. Similarity in this case can be in time, in space, or via similar events. The range of scenarios is application dependent, such as "a person walking along a corridor", or "there happens a football match in a stadium". An additionally important aspect of scenarios is the possible time span between some of them belonging to the same operation (please note that the concept operation is something very abstract and time consuming, such as scenarios like "starting an operation", "waiting for something to happen", "do something", etc.). Moreover, related scenarios can be discontinued by others, which are not concerned with the mentioned operation. Therefore, a system which has the target of detecting human scenarios must be capable of dealing with a multitude of operations like those a human can perform.
Still, it is not within the scope of this work to deal with human operations. On the one hand, the computational effort would be far too large because of the huge number of possibilities. On the other hand, the presented approach is not intended to observe single persons in all aspects of their lives. Quite the opposite: the system will be installed in a (e.g., public) building and therefore sees only small time frames out of a particular person's life. The detected scenarios and operations refer more to the "life" of the building rather than that of people.
An approach to scenario recognition based on fully learned models is summarized below. This approach  can be used to learn common models for scenarios which can slightly vary in their generated sensor data. The approach is based on hidden Markov models (HMMs) . The states of the model are interpreted as events of the scenario . The approach is mainly targeted for surveillance systems (e.g., Ambient Assisted Living ) to model trajectories of persons or to model routines within sensor environments. One application uses motion detector sensor data to learn about daily routines in the occupation of rooms.
A hidden Markov model consists of a transition matrix (it gives the probability of going from one particular state in the model at time to another state at time . Usually, the transition matrix is time independent, which is no hard restriction, since implicit time dependency can be incorporated via self-transitions and parallel paths within the model), an emission or confusion matrix (which models the probability of outputting symbols from an alphabet or some value from a continuous distribution), and an initial state distribution vector. The latter gives the probabilities of being in all the states at the first point in time. In the presented approach it can be omitted with the introduction of an initial and a final state, which have to be passed by each scenario.
In the motion detector application the initial state represents 0:00 in the morning, while the final state represents midnight. In between these two there are different paths which represent one particular daily routine. That sensor sends a data packet with value 1 in case of detected motion. When the sensor permanently detects moving objects, it sends packets at a maximum speed of five seconds. After detecting no moving object for more than 1 minute, the sensor sends a packet with value 0. The system is not directly supported with the motion detector's sensor values, but with averaged sensor values. The 24 hours of a day are divided into 48 time slots, each 30 minutes long. In those time slots, the mean of the sensor values is computed and rounded. If no value is available during 30 minutes, the mean is set to 0 which is synonymic to "no motion". The chains of 48 values are then fed into the (empty) model and during a procedure of several merging steps the structure of the model is learned (see also ). Merging in combination with the averaging of the sensor values will produce HMMs with a manageable number of states. The number of states of HMMs is a compromise between generalization (low number of states, the model is applicable for a wide range of different scenarios, but not able to distinguish between particular ones) and specialization (rather high number of states, not every possible scenario is depicted in the model and quite similar scenarios can have different paths).
For another level of abstraction, models of single days can be easily put together with their initial and final states to create a model for a longer period, for example, a week. For such purpose the transition probabilities from the initial state to particular days can be modified with respect to their position within the week. Hence, the first five models can omit the weekend part (and renormalize the rest), while the latter two could be modeled with only state 14 between initial and final connection state.
6. Processing and Symbolization of Ambient Sensor Data
Symbols can be created, their properties can be updated, and they can be deleted. Their level of sophistication increases with each layer. The number of symbols is different at each layer. At the lowest layer, a large number of microsymbols occur. At the representation layer, only a few symbols exist, where each symbol represents a lot of information of a higher quality. The three types of symbols are defined as follows.
Microsymbols are extracted from sensory input data. They form the basis of the symbol alphabet and bear the least amount of information. A microsymbol is created from a few single sensor inputs at a specific instant of time. Examples for microsymbols in the scenario of Figure 3 are motion (detected by motion sensors), footsteps (detected by tactile floor sensors), objects or a person (detected by video cameras), and so forth.
A group of microsymbols is combined to create a snapshot symbol. They represent how the system perceives the world at a given moment in time. Whenever the system perceives a situation or an object of interest, it creates an according snapshot symbol. The information is provided either by the presence of microsymbols or the absence of specific microsymbols. Examples for snapshot symbols in the scenario of Figure 3 are a gait, a standing person, a falling person, a lying person, and so forth. When the symbol is activated it is determined by either if-then rules or fuzzy rules. The if-then rule used for activating, for example, the symbol gait looks as follows. In the other two layers, the same type of rules are applied.
if (motion==true &&
The third level of symbolization is the representation of the world. Similar to snapshot symbols, representation symbols are used to represent what the system perceives. The fundamental difference is that representation symbols are created and updated by establishing associations between snapshot symbols. The representation level contains not only the information how the world is perceived at the current instant but also the history of this world representation. Compared to the lower levels of symbols, there exist only a few representation symbols, and these are seldom created or destroyed. Only their properties are updated regularly. Following the example mentioned above, on this level, it is detected that a person fell down and cannot get up any more by integrating the information coming from the snapshot symbols. It is important to note that the world representation does not hold the entirety of all sensory information available but just what is defined as relevant. If for instance a person walks around, the world representation does not present information at which exact positions the person has placed its feet. Rather than that, it presents just a position for this person, which may be more or less accurate.
The representation layer can be regarded as the interface to applications. Applications are required to monitor the world representation in order to obtain the information needed to fulfill their specific tasks. This approach relieves applications from handling large amounts of sensory information and provides a condensed and filtered composition of all this information in a highly reusable way. When an application is running, it searches the existing world representation for scenarios that the application knows (e.g., an elderly person has collapsed on the floor) . The events that are required for the scenario to take place can be found on the representation level. Therefore, the application augments the representation by noting that it has found a scenario. It does so by creating a scenario symbol. This makes it possible to study the output of applications later. Additionally, an application can create higher-level scenarios by linking together lower-level scenarios of other applications. That way, the hierarchy can be even further extended by having lower-level applications looking for simple scenarios and higher-level applications using these scenarios to find more complex scenarios.
7. Perception in Factory Automation
There are many applications in which perception can be a key success factor in factory automation. Traditionally, automated machines have carried out quite simple tasks in factories. At most, devices doing simple tasks work together and, after a proper coordination, may execute more difficult enterprises. Still, as already mentioned, perception enables them to go beyond that turn point and start fulfilling more complex activities.
In this way, here we present a Bayesian-network-based model that allows error detection and prediction in high-precision foundries. Basically, information queries are processed by a trained Bayesian network, which issues its prediction on whether the piece to be casted is going to be valid or not. That is, as seen in Figure 3, the sensor values are gathered into the microsymbol layer and the Bayesian network, based upon that representation, produces an snapshot symbol layer (error or not) that may be used in the upper layer, the representation symbol one, to call a reaction. By analyzing more representation symbols, a surveillance application might deduce, for instance, that a quality control is needed (in case more errors are detected or predicted), try to infer their cause, update the production plan (to reschedule pieces that will not be produced), and so on. Later on, we will give an example of such an application running on top of the representation layer.
Nowadays, the only used methodology to guarantee a failure-free casting production (up to a given probability) consists in performing random quality controls (which is a common practice in many other industries). Such controls proceed in the following manner: moulds considered to be representative of a certain production time are extracted and examined with ultrasounds to detect microshrinkages (which is the error targeted by this application). In case it is failure-free, the whole lot is labeled as correct and dispatched. Yet, if a microshrinkage is found, then the failure procedure starts. The first step is the assessment of the damage, depending on the number of the pieces involved, the position of the defect, its size, and so on, a microshrinkage can be acceptable (i.e., the flaw is minor) and, therefore, the piece must not be discarded. Otherwise, the responsible person decides whether analyze the whole lot or discard it.
Against this background, the alternative presented here combines the power of Bayesian networks with the perception architecture described in Figure 3. Bayesian networks  are probabilistic models that are very helpful when facing problems that require predicting the outcome of a system consisting of a high number of interrelated variables. After a training period, the Bayesian network learns the behavior of the system and, thereafter it is able to foresee its outcome.
This Bayesian network was fed with real data of the foundry and the training consisted in the simulation of manufacturing situations whose output had been registered beforehand. After the Bayesian network was tuned up properly, it was applied to predict the outcome of several normal production lots that were also double checked by ultrasound techniques afterwards (see [41, 42] for more accurate description of the training process followed, experiments done, and results obtained).
In a first version presented in , the Bayesian network concentrated on distinguishing pieces containing microshrinkages. Therefore, there was only one symbol at the snapshot layer. A second version (reported in ), extended the number of symbols at that layer to define risk models, which increased the accuracy of the predictions. This time, the Bayesian network was able to distinguish between valid and not valid microshrinkages. The risk levels modeled the sensitivity of the system and, in this way, helped better classify the outcome of each production situation (i.e., whether a microshrinkage will appear and whether it will be valid or invalid).
The definition of these risk levels was performed as follows: the Bayesian network used the analysis on the first lot of the production series to infer the behavior of the rest. According to this result, the risk of every lot was classified into "Risk 0" (no microshrinkages foreseen), "Risk 1" (less than 5 valid microshrinkages expected), "Risk 2" (more than 5 valid microshrinkages predicted), and "Risk 3" (invalid microshrinkages foreseen). Thus, the prediction was more accurate and gave more detailed information.
Still, the real power lies on the use of the information, not on the information itself. Having the Bayesian network issuing predictions on castings' validity, that would not be enough without giving those forecasts a proper use. In this way, the Bayesian network predictor architecture was fitted with an additional application (the so-called Sensitivity Module, (SM) ), operating on top of the representation symbol layer.
The SM studied the different values that each variable (i.e., microsymbols) adopted in order to trace the influence of such values in the apparition of the different microshrinkage risks (i.e., snapshot symbol). Note that a variable may represent for instance using one or another product in a certain phase of the process, applying one certain methodology or not, and so on. In this way, if a variable showed the type of cleaning method used and there were 3 choices, the sensitivity module was able to determine which one was the most convenient in terms of preventing the apparition of microshrinkages. That is, the SM evaluated the results obtained by the Bayesian network and calculated the causal relationship between each type of cleaning method (i.e., value of the variable cleaning method) and the probability that a certain microshrinkage risk appeared. Hence, the SM was able to recommend using only the one that presented the smallest probability, that is, prevent a certain (not desired) scenario to appear.
8. Bionics for Human-Like Machine Perception
Machine perception deals with designing machines that can sense and interpret their environment. For restricted and well-known environments, they can be already achieved, quite promising results. However, the situation changes when shifting to the so-called real-world environments with a seemingly infinite number of possible occurring objects, events, and situations. The problems that scientists are currently confronted with here show that this research area is still in its infancy . In contrast, it is well accepted that humans are equipped with a preceptory system that enables them to apprehend their environment within reasonable time and accuracy . This inspired several research groups to use biology as archetype for perceptual model development . Success of most existing approaches, however, has been limited so far. One reason might be that in many cases, engineers just "grab" some fancy sounding terms and concepts from biology for model development without considering the overall functioning of the biological system taken as archetype . In contrast to this, one quite promising approach to human-like machine perception, which actually sticks to neuroscientific and neuropsychological research findings about the structural organization and function of the perceptual system of the human brain, was made by the work in . The basic idea of this approach will briefly be sketched in the following.
In order to perform complex perceptive tasks, a certain number of neurosymbols are connected to a so-called neurosymbolic network. The structural organization of this network is similar to the modular hierarchical organization of the perceptual system of the human brain as described by [24, 47, 48]. Information of different sensory modalities is first processed separately and in parallel and then merged in higher hierarchical levels. In a first processing step, simple so-called feature symbols are extracted from sensory raw data. Information processing in this level correlates with information processing performed in the primary cortex of the brain. In the next two steps, feature symbols are combined to subunimodal and unimodal symbols. These two levels correspond to the function of the secondary cortex of the brain. Afterwards, information of all sensory modalities is merged to a multimodal perception, which is in accordance with the function of the tertiary cortex of the human brain. For application examples of this model, see . In an application of the model, the meaning of the neurosymbols has to be predefined, whereas the weigths can be learned. This is done in a hierarchical way layer by layer, where first the forward connections from the lower to the higher layer are trained with the help of examples. After finalizing the forwards, the feedbacks to the lower layers are trained, again with examples. This procedure may generate slightly different weights compared to learning forward and backward connections at once but ensure stability and effectivity of the learning approach.
Within a neurosymbolic layer, information is processed in parallel, which allows high performance. Like in artificial neural networks, connections and correlations between neurosymbols can be acquired from examples in different learning phases. Despite some similarities, neurosymbolic networks show many differences to artificial neural networks. In both cases, weighted input information is summed up and compared with a threshold in the basic processing units. Both combine basic processing units to perform complex tasks and process information in parallel. However, unlike in neural networks, where information is represented in a distributed and generally not interpretable form via weights of connections, every single neurosymbol has a certain interpretable semantic meaning as each neurosymbol represents a certain perceptual image. In artificial neural networks, only the structure and function of a single nerve cell serves as biological archetype. In contrast to this, in neurosymbolic networks, also the structural organization of the perceptual system of the human brain is used as archetype of their architecture. Hence, neurosymbolic networks combine advantages of neural and symbolic systems. For a more detailed discussion of this topic, see .
9. Conclusion and Outlook
This paper has outlined four current approaches to overcome the problem of complexity in sensor systems. Future automation systems will perceive their environment with myriads of sensors (the so-called smart dust), having available a quality of perception that may reach or even exceed human perception. This situation implies some basic problems related to initialization, reliability, and sensor fusion. The presented approaches tackle the problem of sensor fusion from different perspectives. Hierarchical systems are introduced—as are used in the human brain—in order to reduce the complexity and amount of data layer by layer while on the other hand enriching the semantic meaning of data.
- Hainich RR: The End of Hardware, A Novel Approach to Augmented Reality. Booksurge; 2006.Google Scholar
- Mattern F: Ubiquitous computing: schlaue altagsgegenstände—die vision von der informatisierung des alltags. Bulletin des SEV/VSE 2004, 19: 9-13.Google Scholar
- Elmenreich W: Sensor Fusion in Time-Triggered Systems, Ph.D. thesis. Vienna Univertiy of Technology; 2002.Google Scholar
- Beyerer J, Puente Leon F, Sommer K-D (Eds): Informationfusion in der Mess- und Sensortechnik. Universitätsverlag Karlsruhe; 2006.Google Scholar
- van Dam J: Environment modelling for mobile robots: neural learning for sensor fusion, Ph.D. thesis. University of Amsterdam; 1998.Google Scholar
- Vadakkepat P, Lim P, De Silva LC, Jing L, Ling LL: Multimodal approach to human-face detection and tracking. IEEE Transactions on Industrial Electronics 2008,55(3):1385-1393.View ArticleGoogle Scholar
- Elmenreich W: A review on system architectures for sensor fusion applications. In Software Technologies for Embedded and Ubiquitous Systems. Springer, Berlin, Germany; 2007:547-559.View ArticleGoogle Scholar
- Ruser H, Leon FP: Informationfusion—eine üebersicht. Technisches Messen 2007,74(3):93-102. 10.1524/teme.2007.74.3.93View ArticleGoogle Scholar
- Luo RC, Yih CC, Su KL: Multisensor fusion and integration: approaches, applications, and future research directions. IEEE Sensors Journal 2002,2(2):107-119. 10.1109/JSEN.2002.1000251View ArticleGoogle Scholar
- Perlovsky LI, Weijers B, Mutz CW: Cognitive foundations for model-based sensor fusion. Proceedings of the International Society for Optical Engineering: Signal Processing, Sensor Fusion, and Target Recognition, April 2003, Proceedings of SPIE 494-501.Google Scholar
- Velik R, Lang R, Bruckner D, Deutsch T: Emulating the perceptual system of the brain for the purpose of sensor fusion. Proceedings of the Conference on Human System Interaction (HSI '08), May 2008 657-662.Google Scholar
- Costello MC, Reichle ED: LSDNet: a neural network for multisensory perception. Proceedings of the 6th International Conference on Cognitive Modeling, 2004 341.Google Scholar
- Davis J: Biological sensor fusion inspires novel system design. Proceedings of the Joint Service Combat Identification Systems Conference, 1997Google Scholar
- George D, Jaros B: The HTM Learning Algorithms. Numenta; 2007.Google Scholar
- Harvey RL, Heinemann KG: Biological vision models for sensor fusion. Proceedings of the 1st IEEE Conference on Control Applications, 1992 392-397.Google Scholar
- Hawkins J, George D: Hierarchical Temporal Memory—Concepts, Theory, and Terminology. Numenta; 2006.Google Scholar
- Kammermeier P, Buss M, Schmidt G: A systems theoretical model for human perception in multimodal presence systems. IEEE/ASME Transactions on Mechatronics 2001,6(3):234-244. 10.1109/3516.951361View ArticleGoogle Scholar
- Murphy RR: Biological and cognitive foundations of intelligent sensor fusion. IEEE Transactions on Systems, Man, and Cybernetics Part A 1996,26(1):42-51.View ArticleGoogle Scholar
- Kam M, Zhu X, Kalata P: Sensor fusion for mobile robot navigation. Proceedings of the IEEE 1997,85(1):108-119. 10.1109/JPROC.1997.554212View ArticleGoogle Scholar
- Russell SJ, Norvig P: Artificial Intelligence: A Modern Approach. Pearson Education; 2003.Google Scholar
- Damasio A: Descartes' Error: Emotion, Reason, and the Human Brain. Penguin; 1994.Google Scholar
- Brooks RA: A robust layered control system for a mobile robot. IEEE Journal of Robotics and Automation 1986,2(1):14-23.MathSciNetView ArticleGoogle Scholar
- Pfeifer R, Scheier C: Understanding Intelligence. MIT Press; 1999.Google Scholar
- Solms M, Turnbull O: The Brain and the Inner World: An Introduction to the Neuroscience of Subjective Experience. Karnac/Other Press, Cathy Miller Foreign Rights Agency, London, UK; 2002.Google Scholar
- Newell A: Unified Theories of Cognition. Harvard University Press, Cambridge, Mass, USA; 1994.Google Scholar
- Byrne MD: ACT-R/PM and menu selection: applying a cognitive architecture to HCl. International Journal of Human Computer Studies 2001,55(1):41-84. 10.1006/ijhc.2001.0469MATHView ArticleGoogle Scholar
- Ramamurthy U, Baars BJ, D'Mello SK, Franklin S: Lida: a working model of cognition. Proceedings of the 7th International Conference on Cognitive Modeling, 2006 244-249.Google Scholar
- Sloman A, Chrisley R, Scheutz M: The architectural basis of affective states and processes. In Who Needs Emotions? The Brain Meets the Robot. Edited by: Arbib M, Fellous J-M. Oxford University Press, Oxford, UK; 2005:203-244.View ArticleGoogle Scholar
- Goertzel B: Opencogprime: a cognitive synergy based architecture for artificial general intelligence. Proceedings of the 8th IEEE International Conference on Cognitive Informatics, 2009 60-68.Google Scholar
- Bruckner D: Probabilistic models in building automation: recognizing scenarios with statistical methods, Dissertation Thesis. University of Technology, Vienna, Austria; 2007.Google Scholar
- Rabiner LR, Juang BH: An introduction to hidden Markov models. IEEE ASSP Magazine 1986,3(1):4-16.View ArticleGoogle Scholar
- Takeda T, Hirata Y, Kosuge K: Dance step estimation method based on HMM for dance partner robot. IEEE Transactions on Industrial Electronics 2007,54(2):699-706.View ArticleGoogle Scholar
- Tashiro S, Murakami T: Step passage control of a power-assisted wheelchair for a caregiver. IEEE Transactions on Industrial Electronics 2008,55(4):1715-1721.View ArticleGoogle Scholar
- Bruckner D, Sallans B, Russ G: Probabilistic construction of semantic symbols in building automation systems. Proceedings of the IEEE International Conference on Industrial Informatics (INDIN '06), 2007 132-137.Google Scholar
- Burgstaller W: Interpretation of Situations in Buildings, Ph.D. thesis. Vienna University of Technology; 2007.Google Scholar
- Joyce D, Richards L, Cangelosi A, Coventry KR: On the foundations of perceptual symbol systems: specifying embodied representations via connectionism. Proceedings of the 5th International Conference on Cognitive Modeling, 2003 147-152.Google Scholar
- Richtsfeld A: Szenarienerkennung durch symbolische Datenverar-beitung mit Fuzzy-Logic, M.S. thesis. Vienna University of Technology; 2007.Google Scholar
- Pratl G: Processing and symbolization of ambient sensor data, Ph.D. thesis. Vienna University of Technology; 2006.Google Scholar
- Goetzinger SO: Scenario recognition based on a bionic model for multi-level symbolization, M.S. thesis. Vienna University of Technology; 2006.Google Scholar
- Pearl J: Reverend Bayes on inference engines: a distributed hierarchical approach. Proceedings of the National Conference on Artificial Intelligence, 1982 133-136.Google Scholar
- Penya YK, Bringas PG, Zabala A: Advanced fault prediction in high-precision foundry production. Proceedings of the IEEE International Conference on Industrial Informatics (INDIN '08), 2008 1672-1677.Google Scholar
- Penya YK, Bringas PG, Zabala A: Efficient failure-free foundry production. Proceedings of the 13th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA '08), 2008 237-240.Google Scholar
- Velik R: A model for multimodal humanlike perception based on modular hierarchical symbolic information processing, knowledge integration, and learning. Proceedings of the 2nd International Conference on Bio-Inspired Models of Network, Information, and Computing Systems (BIONETICS '07), December 2007 168-175.Google Scholar
- Velik R, Lang R, Bruckner D, Deutsch T: Emulating the perceptual system of the brain for the purpose of sensor fusion. Proceedings of the Conference on Human System Interaction (HSI '08), 2008 657-662.Google Scholar
- Velik R: A bionic model for human-like machine perception, Ph.D. thesis. Vienna University of Technology; 2008.Google Scholar
- Velik R: A Bionic Model for Human-Like Machine Perception. VDH; 2008.Google Scholar
- Bruce Goldstein E: Sensation and Perception. Wadsworth Publishing; 2007.Google Scholar
- Luria AR: The Working Brain—An Introduction in Neuropsychology. Basic Books; 2001.Google Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.