Vol_26 No.6 JOURNAL OF ELECTRONICS(CHINA) November 2009 SIGNATURE DRIVEN MUI』TIPLE IIARGET TRACKING Sun Shuyan Huang Zhipei Ren Xiaoyi Wu Jiankang and Engineering,Graduate University of Chinese Academy of Sciences (School of Information Science Beijing 1 00190,China) Abstract Tracking multiple maneuvering targets remains a challenge due to the existence of clutter and the disturbance of spurious targets.Traditional tracking algorithms treat target measurements as points which results in the loss of information.We have propose a Signature Driven multiple target rackiTng(SDT)method which uses target signature in spectral,spatial and temporary spaces as well as the Markov property of target motion.and the data association process in SDT is very effective.The experimental results have shown its outstanding performance. Key words Tracking;Filtering;Estimation;Signature;Sensor fusion CLC index TN911.72 DOI 10.1007/s11767—009—0089—0 I.Introduction Data association for multiple target tracking has measurements increase,the number of hypotheses will increase dramatically to result in intractable computation.Furthermore,the usage of temporal information is still restricted to track score in the MHT framework.As a matter of fact.hypothe- sistesting is not an eficifent way to make use of the temporal information in target tracking—single track always been a challenging research topic,In the methodology research,tracking algorithms always treat target measurements as points.However, point—measurement assumption has dropped too much information from target signature.For ex- ample,color histogram benefits distinguishing tar— gets in video tracking,but this point measurement assumption without the usage of that will actually score has very limited representation power,and too many hypotheses have complicated the algo— rithm. drop this information.On the other hand,although posterior distribution is conditioned by all the past measurement,but the decisions are made on every time step.Presently,Multiple Hypothesis Tracking Recently,people in tracking begin to realize the importance of features.signatures and target ID in multiple—target tracking[ By doing so,they do not .need to face so many hypotheses.On the other hand the analysis of target signatures in spectral,spatial (MHT)[ J method is regarded as the best multiple target tracking algorithm,whose fundamental idea and temporal domain gives people much more freedom in using rich information from sensory data and prior knowledge. is to use the temporal information of the target measurement through the track scores accumulated over time.However,it is based 0n hypothesis—test principle.While as the number of targets and We have proposed and developed a Signature Driven multiple—target Tracking(SDT)algorithm which uses the spectral,spatial and temporary features of the target to reduce the number of hy— Manuscript received date:August 10.2009;revised date: October 16,2009. potheses and to increase the adaptability of the algorithm. Supported by the National Natural Science Foundation of China(No.607721541 and the President Foundation of Graduate University of Chinese Academy of Sciences fNo. 085l02GN0O1. Communication author:Sun Shuyan.born in 1984.male. Ph.D.student.Ro0Ill 504,Moshi Building,Institute of Automation, Chinese Academy of Sciences, No.95 In what follows,Section II provides the back— ground for multi—target tracking,while Section III describes our SDT tracking algorithm.The ex— perimental results are presented in Section IV. Finally,Section V is the conclusion. Zhongguancun East Road,Haidian District, Beijing 100190 China. Email:sunshuyan07◎mails.gueas.ac.en. II. Background This section first gives a review of single—target SUN et a1.Signature Driven Multiple Target Tracking Bayesian filtering,followed by the formulation of multi—target tracking in Random Finite Set(RFS) framework,in which the multi—target RFS Bayesian recursion and the Probability Hypothesis Density (PHD)filter will be described. 1.Single target tracking In single target tracking,the state of a target is assumed to be a first order Markov process on the state space ∈R with transition density p( £ 】)【3].This Markov process is partially measured in the measurement space Z∈R with likelihood p(z Ixt). The posterior density at time t,p( I 1:t), which is the probability density of the state at time t given all observations 1 :=(z1,…, £)up to time t containing all the information about the state ¨can be computed using the Bayes recur— sion from the probability density P( 一1 l川)at time t一】. P ̄-1( 一 )=fp(x ̄lggt-1 P(xt-1 Zl:t_1)d ¨(1) p( I :)= p(fp(zz ,IxtI )) (鼽lf ( x.1z¨ ))dx Estimates of the state at time t can be derived rfom the posterior density p(x£I 1:£)using Mini— mum Mean Squared Error fMMSE),Maximum A Posteriori(MAP),or other criterions. In this paper,we assume each target follows a Gaussian dynamical model and the sensor follows a Gaussian measurement mode1. .e. p( )=N(x ;缸 (Xt_1), 一 ) p(Z =N(Z ; (Xt),Rt) whereⅣ(・;m,P)denotes a Gaussian density with mean 1 x 10 and covariance P;f ̄t,-1( ¨);gt(z£), 】and R are dynamic transition function,like— lihood function)process noise covariance and ob— servation noise covariance respectively. 2.Random finite set formulation of multi-target tracking In RFS formulation of multi—target tracking, multi・-target state and multi・・target measurement are respectively regarded as a set—valued state and a 755 set-valued observation[4 91Mahler[ 0】 as presented .the Bayesian recursion for the multi—target tracking problem in the presence of clutter and miss—detections under the RFS framework.In this frame work,the multiple target state at time t is represented as a finite setiaao]. .e. : ,…, )) (5) where M ft1 is the number of targets at time t. At the next time step.i.e.at time t+1.some of these targets may die,the surviving targets may evolve to their new states,and new targets may appear.Thus given a multi—target state Xt1 at time t一1.the multi—target state at time t is characterized by the union of the surviving targets S ,the spawned targets巨and the spontaneous births Ft: Xt= ‘ U U (<)lU ‘∈xt 1 U UEU (6) —At the sensor,the multi—target measurement at time t given a multi—target state Xt,is also rep— resented as a ifnite set =Zt,1…, 州)) where N(£1 is the number of measurements at time t.Only some of these measurements are ac— tually generated by targets with others by clutter. Thus,given a multi—target state Xt at time t,the multi—target measurement Z.received at the sensor is formed by the union of target generated meas— urements and clutter .i.e. Zt=哆U Both in Eq.(5)and Eq.(7),there is no ordering on the respectively set--valued state and set—-valued measurement.Mahler[ 。】has developed the multi— target Bayesian recursive formula in the form of set state and set observations.Given multi—target transition density p( l 一1),multi—target likeli— hood p( l ),and the multi—target posterior den- sity p( 一1 l :f_1)at time t一1,then the optimal multi--target Bayesian filter propagates the multi-- target posterior in time via the followed recursion 、、,、●,U、璺756 JOURNAL OF ELECTRONICS(CHINA),Vol。26 No.6,November 2009 (X ̄IZ1:H)= p( l )P(Xt ・ ( ) p(X )= p( l )p啦一 ( p( IX)p 一 (xlz,: ) ( ) where is an appropriate reference measure_3Il0】. However,the recursion Eq.(9)-Eq.(10)in— volves multiple integrals,and the computation is intractable.Thus the PHD filter was developed by Mahler which is an approximation to alleviate the computational intractability in the multi—target Bayesian filter. In the PHD filtering algorithm,the first order statistica1 moment of the posterior multi—target state is propagated in time.For an RFS on )(∈R with probability distribution P,its first— order moment,i.e.,PHD function,v,is defined as follows:the integral of v over any region S【。, 。】 gives the expected number of elements of X which are in S.i.e. f xNse(ex)=f ̄ ( )d where fXf gives the cardinality of X.If some cer- tain assumptions are hold] ,then the posterior in— tensity can be propagated in time via the PHD recursion: vtlt-1 xt)=-fP xt一 ( f墨一 vt xt一 )dxt一 + ( xt vt xt_dxt一 +7(xt)(12) ( )=I1一p。(xt)Iv ( ) + (13) where denotes the intensity of clutter RFS at time£; is the multi—target observation available at time ;7(t)denotes the intensity of spontaneous target birth RFS; ㈠is the intensity of the target RFS spawned by a target with previous state at time ;Ps is the probability that a target still exists at time t given its previous state;P『J is the probability of detection given a state at time t. The PHD recursion Eq.(12)-Eq.(13)requires much less computational power than the multi— target recursion Eq.(9)-Eq.(10),for it avoids the combinatorial computations arising from the asso— ciation of measurements with targets.and the PHD intensity is a function on the single—target state space.However,the PHD recursion does not have closed form solutions in genera1.Thus.Vo and Ma_3. proposed the closed form PHD formulation.i.e.. Gaussian Mixture Probability Hypothesis Density fGMPHD 1 in the case of linear Gaussian dynamic and likelihood function,especially,they have proved the convergence of this close—ofrm formula- tion[121. III. SDT Algorithm This section first presents the definition of the signature for a target,and then the information fusion framework will be derived based on this definition,which includes the gating of existed signatures,signature formation,and signature con— ifrmation and termination,etc. 1.Signature of a target As mentioned above,the analysis of target sig- natures in spectral,spatila and temporal domain gives people much more freedom in using rich in— formation from sensory data and prior knowledge]13】. Fusing a wealth of information from spectral,spa- tial and temporal domain will reduce the number of hypotheses and will increase the adaptability of the data association process in the meantime.Herein the signature of a target is proposed as a framework ofr the fusion of this rich information. The signature of a target o is defined as a data structure from time a to time t as follows: Signature(o,t)={ , ,K , ,4}={ Features at time t< Measurement sequence:Lt=(Y。,…,Y£); Feature sequence: ={仉,…, ); Temporal signature:Kt= , , ); Spectral/spatial signature:C£:{ ,Tt); Overall belief of the signature(o)at time t: ={01,Pt); = L p(,t I厶)+ p( f ),= P(TI厶)+ F。P(Tl ), £+ F1 .); where a=max(t—N+1,s);8,N,J(J=a:t)and t are the emergence time of the target,the length of time series to be analyzed,the index of time and the SUN et a1.Signature Driven Multiple Target Tracking 757 current time respectively;the equivalent status£ at time t containing motion information Call be derived from N slices of measurement data Lt; signatures,a new signature ol will be formed with as the first measurement of OL. the kinetics Markov belief =p(毛l厶)and the feature Markov belief t=p(叩f l )are used to represent the temporal property and the spec— At time t+1,the measurement set{ztJ+1,J= 1,…,mn}within the neighborhood of oL will be settled after gating.Associating these measure— ments with ,new m +1 signatures,oL,, tral/spatial property of the target; =P(T 1Lt) and 7_£=P(T l )are the probabilities that o is J=0,1,…,m ,where J=0 indicates that all of the measurements are clutter,will be formed and regarded as a target based on the respective prop— erties above 2.Signature gating In this paper,we assume a target will keep its continuity in spectral,spatial and temporal domain during its movement. .e.the motion and spec— tral/spatial signatures possess Markov property. For example,if a target flies at a certain speed at time t一1.then at time t it will continue moving at the speed which nears that of time t一1,i.e.,the speed of time t will falls into the neighborhood of the speed at time t一1.Thus the gating method is necessary here rfom which the neighborhood of each signature can be obtained. If the equivalent status of the signature at time t is£,the neighborhood of this target is defined as follows: G={zI(z一9 ( + (毒 )))T s ( 一9 ( + ( ))) E} (14) where ft+llt(x£+l lx£)and ( Ix )are respectively dynamic transition function and likelihood function, whose forms are problem dependent,while different choice of these functions presents different Markov properties. 3.Signature formation and management Traditional tracking methods emphasize data, while our method emphasizes information fusion rfom organized sensory data and prior knowledge in different levels,i.e.the signature of a target. Therefore,the formation and management principle for a signature should be properly defined. The existing signature processes continuity in spectral,spatial and temporal domain and its neighborhood can be obtained using gating method as mentioned above.If there exists a position measurement z+at time t which does not faU into any neighborhood of all existing confirmed oL will be deleted.Thus one signature extends to m +1 new signatures and all of them represent only one target with the same emergence time t. The same implementation will be repeated at time t+2. t +1 +2 Fig.1 Signature formation Hence we form a tree consisting of signatures. Each signature will be confirmed if it is generated from true target measurement or wil1 be terminated if it is composed of clutter in a tree fsee Subsection III.61.A tree is confirmed as long as any signature in it has been confirmed.If one confirmed tree has been existed for consecutive N time steps,the confirmed leaf signature with the highest overall belief 0 will be sent to the state filter as the output of this confirmed tree, while 0 has been defined as above. 4.Temporal signature The continuity of a target in temporal domain mainly refers to the Markov property of the target’s state,i.e.the kinetics Markov property,which in— eludes position,velocity and acceleration informa- tion.Given a tree structure《三)at time t.while ={厶, , , , )belongs to .At time t+1,with the measurement set Zt+1={ l, i=1,…,mn)in the neighborhood of oL,the meas— urement set that falls into the neighborhood of《三) js: 758 =JOURNAL OF ELECTRONICS(CHINA),Vo1.26 No.6,November 2009 U = i∥= ・ =P( 扎 ) 刈 ={厶,J, ,,,Kt Ct ,J}.J=0,1,…,m ,the kinet— 6.Bayesian confirmation and termination lPoGiven a signature,the decision that whether this P c 扎, ¨ =P( ̄+I’; : 二≠。 signature is generated by the measurements of a target should be made.This can be done using the Bayesian criterion[ . ∑(1+ nco As for the temporal information,only N time slides of data up to the current time t will be considered in our method,thus we have =P( 扎 fLt扎 )=P(毒 ,f + ) (18) P(岛扎 l毫)=Ⅳ( 一g ( ( )),s) (19) 5.Spectral/spatial signature Be ̄re launching the tracking algorithm,a tar— get model base should be properly defined and established.This may involve the study of sensory system as well as the characteristics of target sen— sory data with consideration of reflectance,target dynamics and the characteristics of the clutter,etc. In order to derive the target model parameters,we may need to collect training data samples and de- sign proper training algorithms.For example,cer— tain type of aircraft may fly at certain speed and acceleration,which can be characterized by Doppler signa1.This Doppler signal can be used as the spectral/spatial signature.The continuity of a target in spectral/spatial domain mainly rears to the Markov property of the signal like this,which is problem dependent and should be properly defined. After the deifnition of spectral/spatial signature, we obtain the spectral/spatial Markov belief in the same way as that of kinetics Markov belief Let P( JT)to be the probability that the sig- nature is formed based on true observations of a target,then through Bayes Rule[ , P( 1 )= P(o ̄IT)Po(T) P(OL) P( ̄lr)Po(T) P(alT)Po(T)+P(o ̄IF)Po(F) ( )+P0(F)=1 (22) where P0( )and P0(F)are respectively the initial probability that oL generated by a true target or by clutter.Let P=P(aIT)/P( ̄IF)to be the like- lihood ratio,then P( l )= pro(T)+(PP0(1一 (T) )) (23) Given P( 1T)at time t, will extend to m。 +1 new signatures J,J=0,1,…,m ,at time t+1,with the corresponding likelihood ratio PⅢ,J,then we get P(TlaJ) P T ,)= (24) With confirmation and termination threshold G and弓 ,the Bayesian confirmation and ter— mination logic is given below. P(T o ̄,)< ,signature terminated <P( )< ,continue test (25) P( I ,)> ,signature c。nfirmed Given and P(T la)= P(TIL,)+ F tP(TIF ̄) at time t,where ={厶, ,Ks,Ct, )and u + 。=1.At time t+1,OL extends to ,,J= L SUN et a1.Signature Driven Multiple Target Tracking ● J 759 0 m We propose to use RFS tracking as the state ilfter[。,10, .and at the state update stage,we allow ●ltⅣ h C 鲁P( the state estimation to be updated by any possible signature,as a set operation.The complete algo— o Pt+lJ (26) rithm is presented below: r ,r 1一 e S p 1一 o n d ●l n P g r 7t+1,J 1一 (27) ta ●l o 1一 where P。iL J s the detection probability;8 is the clutter spatia al density;P£+1are re— ,spectively tn J and +1,,d emporal and spectral/spatial Markov likelihood ratio. Then we have P(T Lt ,)= ,尸( I厶) +l, P(r I厶)+ (1-P(rlL ̄)) (28) P( J 扎 )= : 萎 (29) while P(TILt+l_J)and P(TIFt_1.J)are the prob— abilities that the signature is regarded as a target based on the respective properties above. 7.Overall belief of the signature Given temporal signature Eq.(16),Eq.(28)and spectral/spatial signature Eq.(20),Eq.(29),the overall belief of a signature is derived by =02L, p(专I厶)+ p( I ) f30) =czL,tP(TILt)+a%f(TlF,) (31) where £and F t are weights on the condition .L.£+ F£=1 which are problem independent. .Hence,given a sequence of observation data,as well as prior knowledge of targets of interest and application domain knowledge,we are going to estimate the number of targets and their states (1ocation,speed,and acceleration),while the targets maybe emerge,disappear,spawn and move with various type of motion models at the state space, and the targets may be detected partially with the existence of clutter due to interference of other reasons at the observation space. 一 lGiven tree set三={ "l =1,…, )at time t,measurement set Zt+1 at}ime t+l,and multi— ple target state set ={ . li=1,…,os)at time 8=t—N+1,where )={ l2=1,…, 礼 , ={ ’ , , , , 。). Step 1 Prediction = _l l( ’);for l=1,…, and =1, …・佗 ・ Step 2 Gating For =l,…, ,do gating using Eq.(14)to get l=1,…,n and derive for for through Eq.(15). =Zm} U4fl Step 3 Signature formation Associate 0 with z f]U and form new (kd signatures …,where =I 1. 血 = r!k,0 , , };J=0,l, …,m ;l=1,…,竹 =1,…,n£;{ derive J Jfrom ,7 by polynomial fitting; derive a‰){ and 7- respectively through Eqs.(16),(2S),(20),and(29)I) If +1≠ and +1={ 1,J=1,… initialize m +1]signatures ={Q‰ Adjust the superscript and subscript. 三刊 ={ Ik=1,…, 刊 ),where = 十m', +1)・ l1 Step 4 Signature confirmation and termi..r nation Initialize置+1= ;rt…=0; oFr k=l,・-.,仃£+1I ,do signature confirmation, termination and management through Eq.(25)and Subsections II1.3,6,and adjust the superscript and subscript. 三 + ={O;fllk=1,…, + ), ={ ?If=1j…, ) Step 5 State filtering . 、oFr =1,…, , with the forming time s < If s t—N+2 with been confirmed, ifnd in with m=argmax ( ): if the cOrresponding track has been initial— ized find it in x s and put it along with ,川) 760 JOURNAL OF ELECTRONICS(CHINA),Vo1.26 No.6,November 2009 into the state filter to get a new state als+1。,; else set 抖1)as the position initial value and initialize state as _l_。;, + ={ +1.。li=1,…,Os+1); Step 6 Output 三+ ={ =1,…, + ); ={ f-1,…,礼 ); X刚={ +1l。 =1,…,Os+1) IV. Experimental Results 1.Comp ̄ison with MHT In this experiment,a two—dimensional scenario with unknown and time varying number of targets observed in clutter is simulated for illustration purposes.The targets are tracked in the surveil— lance region of i一1000 m,1000 m1 x 1—1000 m,1000 m],while the state =[P ,P ,vx Vy, r of each target consists of position(P ,P )and velocity ( £,Vy.t),while the measurement is the position with noise.Again we assume each target follows a linear Gaussian dynamical model,i.巴 p( fIxt-1)=N(xt; 一 , 一 ) (32) p( l )=Ⅳ HtXt, ) (33) whereⅣ(・;m,P)denotes a Gaussian density with mean钉 and covariance P,and 1 and Ht are respectively dynamic transition matrix and obser— vation matrix.The dynamic transition matrix F 1 and process noise covariance 一1 are I2△I I, 2 = 02 = : (34) △ I、 where j『2 and 02 denote 2 x 2 identity and zero matrices.△=2 s is the sampling interva1.and =0.3 m/s。is the standard deviation of the process noise. The observation matrix Ht and measurement noise covariance matrix R.are given by H = [ ,02].Rt= : ,where =5 m is the stan— dard deviation of the measurement noise.The de— tection probability of each target is 0.9886,and the clutter are modeled as independent identically distributed fi.i.d.)measurements with uniform spatial distribution while the spatial density is =1×10 m- .i.e.there are 40 clutter returns over the surveillance at each time step.The maximum speed of the targets is 100 m/s.The total simulating time is 200 8,while the whole number of time steps is 100. The MHT algorithm was first proposed by Reidp5]at 1978,and was greatly developed by Blackman[1,21.The algorithm used in this paper was given by Cox and Hingoranip6】,and more details can be found in the literature. Fig.2 shows the true target trajectories,while Fig.3 shows the measurements over whole simu— lating time,and Fig.4 shows the measurements and true trajectories against time respectively along x axis and Y axis.Tab.1 gives the born/emergence time steps and the death/disappearance time steps of these targets.Fig.5 and Fig.6 respectively show the tracking results of the SDT method and the MHT method. Tab.1 The born time and death time of each target 1000 800 600 400 200 0 —200 400 -600 —800 1000 -800 400 0 400 800 zfm1 Fig.2 Target true trajectories,while the circle marks denotes the target emergence locations,and the square marks represents the target disappearance locations SUN et a1.Signature Driven Multiple Target Tracking 1000 800 600 400 200 0 200 400 600 800 1000 -800 400 0 400 800 f】111 Fig.3 Measurements over whole simulating time,while a cross mark indicates a position measurement 1000 800 600 400 200 0 200 400 600 800 1000 10 30 5O 7O 90 Time(step) 1000 800 600 400 200 0 200 400 600 800 1000 l0 3O 5O 70 90 Time(step) Fig.4 Measurements and true trajectories against time,while the lines denote trajectories and the cross marks represent measurement including clutter and measurement generated by targets The Wasserstein distance and errors in target No.estimating are used here for the evaluation and comparison between the two algorithms.The Wasserstein distance penalizes when its estimate of the number of targets jS incorrect while adopted as multi—targets distance[ .Given the multi—target truth 761 = …, (35) and its estimate 2= …, (36) the Lp Wasserstein distance is given by ( , )= n f I"1 Ilxal I (37) 1000 800 600 400 200 0 200 400 600 800 1000 -800 400 0 400 800 X(m) Fig.5 SDT method tracking results,where the solid lines denote true trajectories,and the circle mark lines represent SDT estimated trajectories 1000 800 600 400 200 0 200 400 600 800 400 0 400 800 fm) Fig.6 MHT method tracking results,where the solid lines denote true trajectories,and the circle mark lines represent MHT estimated trajectories while the minimum iS taken over the set of all transportation matrices C={ ,j);while the ma- trix C satisfies 762 fJOURNAL OF ELECTRONICS(CHINA),Vo1.26 No.6,November 2009 ∞u∑ f38) (39) 伽 蓦i Here in this paper P=2.The errors in target No ∑芦 estimating = is given by = E{IXI一 From Figs.5~8.we can see that the SDT filter provides excellent estimates of both target position and number,and picks up less false alarms corn— paring with that of MHT.From Tab.2.we can see the Wasserstein distance of MHT is much larger than that of SDT.and the error in target No.es— timating of MHT is also bigger than that of SDT. The reason of this is that the MHT method miss—takes0葛鲁=∞∞0Z 磊lJ口【J0基 1 several1 1 1 fal0s e alO Oarms as targets,and the 0 Wasserstein distance penali2 8 6 4 2 1 8 6 4 zes when the estimate of 2 the number of targets is incorrect. As we can see from the table and the figures above,SDT shows its capability of adaptation to accurately estimate the changing number of targets, and fast response to the changes of the position and speed,and therefore,the smaller tracking errors. Tab.2 Tracking evaluation between the two algorithms …MHT O 10 2O 30 40 50 60 70 80 90 i00 Time(step) Fig.7 Comparison of Wasserstein distance between MHT and SDT against time,while the solid line denotes that of SDT,and the dashed line represents that of MHT 0 10 20 30 40 50 60 70 80 90 100 Time(step) Fig.8 Comparison of error in target No.estimates between MHT and SDT against time,while the solid line denotes that of SDT,and the dashed line represents that of MHT 2.Tracking with the Doppler signal In our second experiment.a two—dimensiona1 scenario in which everything is quite the same as that of the first one is simulated.except that the clutter spatial density is set as =1×10 m_ . =2.0408×10一 m一。, =3.0246×10一 m一。 and =4 x 10 m in succession.The maximum speed of the targets is i00 m/s.The total simulating time is 200 s,while the whole number of time steps is 100. The Doppler frequency shift is used as the spectral information in this experiment.The sensor is assumed to be settled at the center of the sur— veillance region.i.e.the origin of the coordinate system.Each moving target has a frequency shitf while the movement is in the way of the sensor[ . which is governed by 厶I= 2v ̄ft C where lD is the Doppler rfequency;vr is the radial speed of the target;c is the speed of the light;and is the transmitters frequency.In this paper,only non-maneuvering targets are simulated whose speeds do not change much within one time step. Thus it is assumed that the Doppler signal follows a Gaussian model, .e.given the Doppler signal,d-f_1 at time t一1.the Doppler signal of time t is gov— erned byⅣ( £; -f_1,Rv),while the covariance is given by (42) where the standard deviation takes the value D SUN et a1.Signature Driven Multiple Target Tracking =4 m/s;while the transmitters frequency is = 6 GHz.On the other hand the clutter iS assumed static. While tracking these targets,the MHT method makes use of only point measurements,and the SDT method utilizing both the point measurement and spectral information iS mentioned as FeaSDT for the distinguishing purpose from the SDT method which does not make usage of Doppler signa1.The Wasserstein distance iS also used here for the evaluation of different algorithms. Tab.3 Average of Wasserstein distance of diferent algorithms Clutter spatial Average of Wasserstein distance(in) density (×10 in ) MHT SDT FeaSDT Tab.3,as well as Figs.9-12 show the per— formance of these three algorithms,from which we can see obviously that our SDT method with the infc)rmation fusion of spectral signal tracks multiple targets more accurately both in the state of multi—target and in the No.of targets comparing with that of MHT algorithm.Based on hypothe- sis—test principle and the usage of track score,the advantage of MHT has been limited while dealing with false alarms. 1000 800 600 400 200 0 200 400 600 800 1000 -800 400 0 400 800 fn11 Fig.9 MHT tracking results OD the condition of =3.0246× 10 m,where the solid lines denote true trajectories,and the circle mark lines represent estimated trajectories 763 i000 800 600 400 200 0 200 400 600 800 1000 800 400 0 400 800 z,m1 Fig.10 SDT tracking results on the condition of口=3.0246× 10 m一,where the solid lines denote true trajectories,and the circle mark lines represent estimated trajectories i000 800 600 400 螂 ㈣ 湖 娜200 0 200 400 600 800 1000 -800 400 0 400 800 fm1 Fig.11 FeaSDT tracking results on the condition of口=3.0246 ×1O一 m.where the solid lines denote true trajectories.and the circle mark lines represent estimated trajectories 0 10 20 30 40 50 60 70 80 90 100 Time(step) Fig.12 Comparison of Wasserstein distance on the condition of =3.0246×1O一 m,while the solid line denotes that of SDT,the dashed line represents that of MHT,and the dashed line with circle marks indicates that of FeaSDT 00《 苫764 JOURNAL OF ELECTRONICS(CHINA),Vo1.26 No.6,November 2009 on Intelligent Control,Singapore)Singapore,Oct.1-3 2007,458—463. 1 2 3 4 5 6 V. Conclusion We have presented a new multiple target tracking algorithm,Signature Driven multiple [8 Ya-Dong Wang,Jian—Kang Wu,Weimin Huang,and A.A.Kassim.Gaussian mixture probability hypothe- sis density for visual people racking.10th Interna- target Tracking(SDT).It makes use of the target signatures in the earliest possible time,and has shown the superior performance.Further work will be on the testing and improvement of the algo— rithm. tiona1 Conference on Information Fusion,Qu ̄bec, Canada,July.9 12,2007,1-6. [9 Ba-Tuong Vo,Ba-Ngu Vo,and A.Cantoni.Analytic implementations of the cardinalized probability Acknowledgment rhe authors acknowledge the support by the members of Sensor Networks and Applications Research Center(SNARC),especially,Ji Lianying, for his guidance and suggestions about radar knowledge,and Li Shuangquan,for his support of MHT implementation. References S.s.Blackman.Multiple hypothesis tracking for multiple target tracking.IEEE Aerospace and Elec— tronic Systems Magazine,19(2004)l,5—18. J.Lancaster and S.Blackman.Joint IMM/MHT tracking and identification for multi—sensor ground target tracking.9th International Conference on Information Fusion,Florence,Italy,Ju1.10—13,2006, 1-7. Vo Ba-Ngu and Ma Wing-Kin.The Gaussian mixture probability hypothesis density filter.IEEE Transac— tions on Signal Processing,54(2006)11,4091-4104. Ba-Tuong Vo,Ba-Ngu Vo,and A.Cantoni.Bayesian ifltering with random finite set observations.IEEE Transactions on Signal Processing,56(2008)4,1313— 1326. R.Mahler.PHD filters of higher order in target number.IEEE Transactions on A erospace and Electronic S stems{43(2o07)4,1523—1543. D.E.Clark and J.Bel1.Multi—target state estimation and track continuity ofr the particle PHD filter.1EEE Transactions on A erospace and Electronic Systems, 43(2007)4,1441—1453. Ba-Tuong Vo,Ba-Ngu Vo,and A.Cantoni.A Bayes— ian approach to target tracking with finite—・set—-valued Observations.IEEE 22nd International Symposium hypothesis density filter.IEEE Transactions on Signal Processing,55(2007)7,3553—3567. [10] Ronald P.S.Mahler.Multitarget Bayes filtering via ifrst—order multitarget moments.IEEE Transactions on Aerospace and Electronic Systems,39(2003)4, 1152 1178. [11 K.Panta,Ba-Ngu Vo,and D.E.Clark.Data associa- tion and track management for the Gaussian mixture probability hypothesis density Filter.IEEE Transac— tions on A erospace and Electronic Systemsj 45(2009)3 1003-1016. [12] D,Clark.B.Vo.Convergence analysis of the Gaussian mixture PHD filter.IEEE Transactions on Signal Processing,55(2007)4,1204—1212. [13] Y.Wang,K.一F.Loe,T.Tan)and J.-K.Wu.Spatio— temporal video segmentation based on graphical models.IEEE Transactions on Image Processing. 14(2005)7,937-947. [14] Jing Zhongliang,Zhou Hongren,and Wang Peide. Tracking initiation and termination of multiple maneuvering targets in a dense multi—return environment.IEEE Conference on Decision and Control,Honolulu,USA Dec.5-7,1990,2270-2275. [15] Donald B.Reid.An Algorithm for tracking multiple targets.IEEE Transactions on Automatic Contro1.24 (1979)6,843—854. [16] I. J. Cox and S. L.Hingorani. An efficient implementation of Reid’s multiple hypothesis tracking algorithm and its evaluation for the purpose of visual tracking.IEEE Transactions on Pattern Analysis and Machine Intelligence,18(1996)2,138—150. 【17 Christian wo1ff.Doppler Effect.http://www.radar— tutoria1.eu/1 1.coherent/co06.en.htm1.July 2009.