您的当前位置:首页正文

Model and Algorithm Research of Multi-Sensor Information Fusion

来源:易榕旅网
Scientific Journal of Control Engineering October 2014,Volume 4,Issue 5,PE150—156 Model and Algorithm Research of Multi—Sensor Information Fusion ZhiliangZhu JingHu 3,Yan Shen1,Shaoming Chen 1.College ofElectrical&Information Engineering,Hunan University,Changsha Hunan 410082,China 2.College ofPhysic&Electronic Information Engineering,Wenzhou Universiy,Wenzhou tZhejiang 325000,China 3.School ofManagement,Zhejiang Industry and Trade Vocational College,Wenzhou Zhejiang 325000,China 4.Education Equipment Engineering Technology Research Center ofZhejing aProvince,Wenzhou Zhejing a325000,China Email:zlzhu@hnu.edu.cn Abstract As the precondition of intelligent control for automatic system,multi-sensor information fusion technology is one of he key ttechnologies of auto.control area.Aiming at the fusion model and algorithm,the research development of multi—sensor information fusion technology was well introduced and concluded in detail.Finally,the future research work tendencies were pointed out. Keywords:Multi-Sensor Information Fusion;Auto—Control INTRODUCTION With the rapid development of microelectronie technology,signal detection and processing technology,computer technology,network communication technology and control technology.multi。sensor system for the complex application background appear in a large number.Because the information acquired by various kinds of sensors may have different characteristics:time—varying or time invariant,real—time or non—real time,fast or slowly varying,fuzzy or determined.accurate or not complete,reliable or unreliable,mutua1 support or complementary,also may conflict, the system must make full use of he multtiple sensors information which is redundant and complementary in space nd taime according to a combination of criteria,in order to obtain consistency on the observation environment description nd aexplanation. Multi—sensor information fusion refers to he titegratned treatment of multiple sensors to provide united expression of he external environment.It is the ftusion of the complementarity,redundancy,real—time and low cost of the information.So it can reflect the environmental characteristics completely and accurately,and helps to make the right iudgments and decisions,ensresu the speediness,accuracy and stability of he tsystem.The ultimate objective of multi—sensor information fusion is to improve he pertformance of the whole system by using he common or tjoint operation advantage of multiple sensors.Multi—sensor information fusion technology will bring the following advantages to system[ 】: (1)Improve the reliability and robustness ofthe system. (2)The coverage extension of time. (3)Extend monitoring range. (4)Enhance data credibiliy.t (5)Reduce the reaction time. Information fusion function can be summarized as:expand the space search range,improve trget detaectability, improve detection performance;improve the spatial or temporal resolution,increase he tdimension of featm'e vector of the target.reduce the uncertainty of information,improve the confidence of information system;enhance the abiliy of ftault tolerance and adaptive of system;followed by the reduction of fuzzy degree reasoning and the -150一 http://www.sj-ce.org improvement the decision—making ability,so that the performance of the whole system is greatly improved. Fundamentally speaking,the results derived from hte redundancy and complementariyt of information.Therefore, the multi—sensor information fusion can often get the results that the single sensor is dififcult to obtain,and the performance will have a qualitative leap.From the principle of speaking,these ideas carl be further performed to equipment,system integration. The fusion model and algorithm is an important research content in multi—sensor information fusion.Aiming at the model and algorithm,the research progress of multi—sensor information fusion technology are introduced and summarized in detail in this paper,and finally the prospects of the future development direction. 1 INFORMATION FUSION MODEL In recent years,people have proposed many kinds of information ufsion model[ .The common point or central idea is multistage processing in information fusion process./n 1 980s,the typical functional model are Intelink,Bc’yd control loop(OODA loop);the typical data model is JDL.In 1 990s the waterfall model and the Dasarathy model developped,Mark Bedworth integrated several model and proposed a new hybrid model in 1 999. J.1 Intelink UK Intelink describe the information processing as a ring structure【jJ,as shown in ifgure 1.It consists of 4 satges:a1 acquisition,including the initial intelligence data from sensors and artiifcial information source;b1 sorting, association the relevant intelligence reports,some data consolidation and compression processing will be made at htis stage,for use in the next stage of ufsion;c)evaluation,in hte stage of ufsion nad analysis of intelligence data,at the saro_e time,analysts also directly delegate tasks of information gathering;d)distribution,in this sent the fusion of information to the user(usually miliatry commander),in order to make the decision of action,includes the next step of acquisition work. FIGURE 1 UK INTELINK 1.2 JDLmodel In 1984,JDL model was proposed by the data ufsion joitn command laboratory of USA Department of Defense[ , after a gradual improvement and application,this model has become a standard of USA defense information fusion system.JDL mode1 divided the data fusion into 3 stages:first stage as the target optimization.1ocation and recognition;second stage for situation assessment,construction situation map according to the first levels of processing information;third stage for threat assessment,explains second stage treatment results according to the possible actions,and analyses the advantages and disadvantages of actions.Process optimization is a repemed process,and can be called the fourth stage,it will monitor the system performance in the fusion process,identify potential increased information sources. 1.3 Boydcontroltoop The Boyd control loop‘ (OODA)is first applied to the military command processing,now has been widely used in information fusion.It consists of 4 stages:first stage as observation,acquisition of target information;second stage for orientation,determine the direction,recognize situation;third stage for decision,making reaction the plna distribution behaviour,as well as logistics management and planning;fourth stage for action,implementation .151 http://www.sj—ce.org plan,and the decision making efficiency problem is considered in practical only in this stage.The Boyd control loop makes the problem of feedback iterative characteristics very obvious.The advantage of OODA model is that it enables each stage ring form a closed loop,show the cyclic data fusion.As can be seen,the data level transferring to next fusion stage decreasing along with the continuous fusion stage..But the shortcoming of OODA model is he tlack of capacity of influence to other stages by decision and action sage,and teach stage is sequential execution. 1.4 Wa ̄rfaH model The waterfall model iS put forward in 1 994 by Bedworth[O1.and now is widely used in the British defense information fusion system:it emphasizes he functtion of the lower level,concrete frame as shown in igure 2.Ifts sensing and signal processing,feature extraction and pattem processing correspond to JDL level first,and the situation assessment and decision making respectively correspond to JDL level 2,3,4.As Can be seen,although usifon process of waterfall model divided into he tmost detailed,but it was not clear feedback process。which should be regarded as its main shortcoming. Decision making Situation assessment Pattern processing 彳 Feature extraction Signal processing 彳 Signal acquisition FIGURE 2 WATERFALLMODEL 1.5 DASARATHY model The Dasarathy model includes 5 fusion levels[7】. TABLE l DASARATHYMODEL As shown in table 1:So as you Call see,the waterfall model makes a clear distinction about he botttom function, while JDL model makes a clear division of middle functions,and the Boyd loop has a detailed explanation of the high level processing.Intelink covers all the processing level,but do not describe in detail.While the Dasarathy model is based on he tusifon ask tor function to he tconstruction,SO it Can effectively describe all usfion behavior. Mixed model integrated the cycle characteristies of ntIelink nd afeedback iterative characteristics of Boyd control loop nd tahe definition of waterfall model is applicated.which is associated with JDL and Dasarathy mode1.speciifc framework was shown in figure 3.The feedback Call be seen clearly in a mixed mode1.The model retains the Boyd control loop to make he tcycle characteristic of information fusion processing clear.The description of main processing tasks in the model has better precision of reproduction.In addition.the position of fusion behavior is also more easily find in he mode1.t .152。 http://www.sj—ce.org FIGURE 3 MIXED MODEL 2 INFORMATION FUSION ALGORITHM Fusion is a form of ramework,whose fgoal is to integrate different source information to obtain high quality useful information by using mathematical methods and technical tools.Mathematical tools is the most basic and multiple function in information fusion,all input data will be effectively described in a public space,properly integrated at the SalTle time,and ifnally output in a suitable orfm nd athe data. Commonly used methods of multi—sensor information fusion can basically be summarized as random and artiifcial intelligence,the application of hese metthods can perform data layer,feature layer and decision layer fusion in different levels,and accurately,fully understand and describe the measured object and environment.Can be foreknow,new concepts and technology such as neural network nd aartiifcial intelligence will play a more and more important role in multi—sensor information usion.f In the fusion technology and calculation method,the multi—sensor information fusion method mainly has:the weighted average,Calman filtering,Bayesian estimation,statistical decision theory,the Demspter-shafer evidence reasoning,production rule;while the calculation method mainly includes:the fuzzy set theory,neural network, rough set theory. 2.1 Weighted average The weighted average is he tmost simple and most intuitive method,which is suitable for dynamic environment.The method weighted average the redundant information provided by a group of sensors,and use value as the ifnal result offusion. Calman filter iS used for real—time fusion of lOW level redundant dynamic multi—sensor data.The method use the statistical properties of measurement model to recursive determine data fusion estimation of optimal statistical signiicafnce If he tsystem dynamics model is linear.and he tsystem noise and sensor noise is white noise model of Gauss distibutrion,provide the only optimal estimation of statistical sinigifcance for fusion data,the recursive characteristics of Calman filter make he systtem data processing does not need a large amount of data storage and computation. 2.3 Bayesian estimation .153一 http://www.sj—ce.org Bayesian estimation is a commonly used method for multi—sensor infc)rmation fusion in static environment of 1ow layer.The information described as a probability distribution,suiable for uncerttainty that has additive Gauss noise. When the sensor group has consistent observation coordinate,the sensor data can be fused by direct method.In most cases,multiple sensors describe the same object from diferent coordinate frames,then the sensor measurement data should be fused indirectly using Bayesian estimation.The problem of indirect method is to find the rotation matrix and translation vector which is consistent with multiple sensor readings …. 2.4 Multiple Bayesian estimation Durrant Wh、,te express the task environment representation as multi—sensor system mode1 of uncertain geometric object set,and put forward the multiple Bayesian estimation method .Each sensor in he systtem is represented by the useful static description abiliy of tthese objects.Multiple Bayesin estaimation take each sensor as a Bayesian estimation,combine the associated probabiliy disttribution of each individual object into a posterior probabiliy tdistribution function.and provide the final value of multi—sensor information fusion by minimizing the 1ikelihood function ofjoint distribution function. 2.5Statisticaldecision theory Literature[12 using statistica1 decision theory rSDT)to propose two step generalized method for the fusion of multi sensor redundancy positioning information.Sensor noise is modeled as a probability distribution of possible variety”e-Contaminated”.The sensor mode1”e-Contaminated”is used to increase the robustness of he decitsion making process,by separating he ditsritbution function to determine the separation coeficifent of£,to represent heavy aitled deviations caused by possible nturue sensor readings.Compared with he tmultiple Bayesian estimation, ncerutainty of statistica1 decision theory is additive noise.thus has uncertainty of wider adaptation.The observed data of different sensors must pass all integrated robust test for consistency;the data of consistency validate is fused by robust extreme decision rules. 2-6 Demspter-shafer evidence reasoning Dempster-Shafer evidence reasoning is extended Bayesian method .In the Bayesian method,specifies the characteristics of all he tlack infc}rmation of tI1e environment as an equivalent prior probabiliy.When the number tof useful additional information or unknown premise of sensor is greater than the number of known premise,the known premise probabiliy becomes tnstuable,it is the obvious disadvantage of Bayesin amethod.In he tDempster—Shaller method,it can be avoided by don’t specify he tprior probabiliy tof unknown precondition. In multi.sensor system.information of the environment provided by each information source has a certain de ̄ee of ncerutainty,the uncertain information fusion process is a process of ncertuainty reasoning.Literature using fuzzy logic to fuse image analysis and target recognition.Fuzzy logic is multiple valued logic,it allows direct representation he uncerttainty in he multti—sensor information fusion process in reasoning process,by specifying a real between 0 to 1 to represent he ttruth degree.which is equivalent to he tpremise of implicit operator.If using a systematic modeling approach to integrate the uncertainty in he tprocess,it will produce consistent fuzzy reasoning. 2.8 Neural network Neural network determine the classiifcation standard based on similariy of tcurrent system accepted sample,this determination method mainly perorfmance in the network weight distribution,and also can obtain knowledge to get the uncertainty reasoning mechanism by using a speciifc learning algorithm of neural network.Stndy of neural network provides a good method for multi.sensor integration and fusion modelling.Foreign scholars have done some pioneering work in multi.sensor integration and information fusion by using he tneural network.Literature proposed a fault.tolerant adaptive reconfiguration method when a sensor failre iun multi—sensor system based on neural network.Multi—sensor integration and fusion based on neural network has the following characteristics:a niuifed imemal knowledge representation.fusing network sensor information through learning method to obtaine -154一 http://www.sj-ce.org network parameters(such as the connection matrix,node offset vector),and can convert knowledge rules to digital form;easy to establish knowledge base to use external environment informationeasy to realize automatic ,nowledge acquisition and associatikve inference.The complex and uncertain environment will fused to accurate signal which system could understand through learning and reasoning. 3.SUMMARY Information usifon is he tnew direction of system science based on he cross,comprehensitve and extension between modern information technology and multidisciplinary.Due to its broad application prospects in military and civil ifelds,it has been highly concemed by many domestic and foreign scholars and relevant departmentsthe system ,review of usion system modeling,ffusion algorithm,existing problems and the idea to solve these problems is given in this paper. Although the multi—sensor information fusion technology has been developed greatly at presentbut there are still ,many problems:how to reduce the uncertainty of sensor information,reduce the error rate of information fusion, improve the real—time fusion,optimal information fusion algorithm,rational allocation of sensor resourcesIn .addition,the multi-sensor information fusion method in dynamic environments is also a very meaningful research direction. REFERENCES Pan Q,Yu X,Cheng Y M,et a1.Essential methods and progress of information fusion theory[J].Acta Automatica Sinica ̄2003, 29(4):599-615.(in Chinese) [2] Wu Xia,et a1.Research Actualities nd aProblems on Multi—sensor Target Recognition System Model[J],Journal ofAstronautics, 2010,31(5):1413—1420.(in Chinese) [3】 Lin K,Chen M,Rodrigues J P C,et a1.System Desin gnd aData Fusion in Body Sensor Networks[J].Telemedicine and E.Health Services,Policies,and Applications:Advancements and Developments,2012,1:1-25 Cilia Patricio M A,Berlanga A,et a1.A Data Fusion Perspective on Human Motion Analysis Including Multiple Camera [4】 Applications[M]//Natural and Artiifcial Computation in Engineering nd aMedical Applications.Springer Berlin Heidelberg.2013: 149.158 [5】 Lan H,Pan Q,Yang F,et a1.Design and implementation of algorithm—level testbed for multi—sensor information fusion[C]// Control Conference(ccc),2013 32nd Chinese.IEEE,2013:8546-8551 [6】 Bedworth M,O’Brien J.The Omnibus model:a new model ofdata usifon?[J].IEEE Aerospace and Electronic Systems Magazine, 2000,1 5(4):30—36 [7】 Faouzi N E E,Leung H,Kurin A.Dataa fusion in intelligent transportation systems:Progress and challenges-A survey[J]. Information Fusion,201 1,12(1):4-10 【8】 Hong Zhaoyi,Gao Xunzhang,Li Xing.Researach on Temporal-Spatial Information Fusion Model Based on DS heorty[J].Signal Processing,201 1,27(oO1):14-19.(in Chinese) 【9】Cui P,Huang X.Multi‘sensor information fusion algorithm based on federal Kalman filter and its application[J].Electric Machines and Control,200 1,5(3):204-207 Wang Yaonan,Li Shutao.Multi sensor information usfion and its application:A Survey[J].Control And Decision,2001,I6(5): 5 1 8—522.(in Chinese) Jing aM L,Liu S Y.Fault Diagnosis Approach Study of Bayesin aNetworks Based on Multi.characteristic Information Fusion[J]. Zhongguo Jixie Gongcheng(China Mechanical Engineering),20 1 0,2 1(8).(in Chinese) Wang J L,Zhang J Y.Multisensor Target Identiifcation Based on Mass Function of Statistical Evidence and D.S Evidence Theory [J].Chinese Joumal ofSensors nd aActuators,2006,19(3):862—864.(in Chinese) Xu C F,Geng W D,et a1.Review of Theory nd Applaications of Dempster--Shafer Evidential Reasoning Mehod[tJ】.Pattern Recognition and Artiifcial Intelligence,1999,12(4):424-430.(in Chinese) Wu X P,Zheng Z S,Fu Y.Fault diagnosis method based on uzzfy logic nd aevidence theory[J].Joumal ofNaval University of Engineering,20 1 2,1:002.(n iChinese) Chen M,Zhu W T,et a1.The application of information fusion technology based on BP neural network in water monitoring[J]. Microcomputer Information,2010(1O):15-17.(in Chinese) .155一 http://www.sj—ce.org rHoRS  ̄Zhiliang Zhu(1982一)j male,the Hart nationality.He obtained his Bachelor degree in Electrical Engineering and Yah Shen(1985一),male,the Han nationality,He iS a PhD student of Hunan university,and his research interests focus on embedded systems, machine learning and computer vision. He obtained his Bachelor degree in communication engineering from Hunan Automation from University of Electronic Science and Technology of China fUESTC)in 2005;and MasteI in Signal and Information Processing of ,,cling University of Posts and Telecommunications in Normal University in 2008.Email:shenyan0712@hnu.edu.cn Currently,being a lecture of Wenzhou Universiy,he its ng his Ph.D degree in Hunan University,China.His ch interest focuses on the field of intelligent robot l1.Email".zlzhu@hnu.edu.ca 一156. http://www.sj—ce.org 

因篇幅问题不能全部显示,请点此查看更多更全内容