GB2510877A - Method for cluster management in multiple target tracking - Google Patents

Abstract

When tracking a large number of targets individually, it is convenient to divide the targets into a number of clusters having a similar target position and maybe velocity. Targets can leave (split from) a subgroup and proceed independently or join (merge with) another cluster. A cluster management method for the automatic generation of sequences of split and merge decisions for a multiple target tracking system comprises a set of input track clusters at time t and a set of output track clusters at time t+1 of arbitrary sizes. All possible pairings between the input and output clusters are evaluated according to a cost function to give a matrix of assignment costs. The cost function accounts for the number of common tracks, the cluster sizes, and whether there is an exact match. The assignment costs are augmented to include the cost of a dummy assignment. A two-dimensional assignment problem (minimization / optimization) based on the matrix of assignment costs is solved to yield a mapping between input and output clusters. The mapping is used to determine the sequence of split and merge decisions that transforms the set of input track clusters into the set of output clusters.

Description

Method for Cluster Management in Multiple Target Tracking

Field of the Invention

[001] The present invention relates to multiple target tracking (MTT) where the physical parameters of interest of a multiplicity of objects are to be determined on the basis of data received at discrete points in time by one or more sensors. More particularly, the invention relates to the management of clusters of target tracks that are used to decompose the MU problem into sub-problems thr computationally efficient solution.

Background of the Invention

[002] A classical example of the multitarget tracking problem arises in radar tracking of aircraft, in which a microwave or high frequency radar sensor measures, for ins tancc, range, bearing and, possibly, range rate of a number of aircraft in its surveillance volume. The measurement data from the radar at discrete points in time are referred to as scans. The problem then consists of processing the scans of received sensor data, which contain random noise and clutter, in order to estimate the position and velocity of the aircraft For historical reasons, the objects in MU are referred to as targets and the sequences of position estimates are called tracks.

[003] The MTT problem is of widespread interest in many domains. Examples include tracking in video surveillance, cell biolo,, oceanography, astronomy, robotics and navigation.

[004] The estimation of quantities that vary in time is what distinguishes tracking from other signal processing problems: tracking is fundamentally a state estimation problem. A further distinguishing feature of MU is that there is generally no a priori information about the correspondence between the measurements and the targets, i.e., given a set of measurements, it is not possible to say with certainty which measurement if any, is from a given target This lack of measurement origin information is known as the data association problem. The data association problem involves discrete uncertainties, while the state estimation problem involves continuous uncertainties, with both types of uncertainty being present in MU.

[005] It is known that multitarget tracking is a problem whose mathematically optimal solution cannot be implemented in general since its computational and memory requirements increase combinatorially with the number of targets. Many methods of approximate, suboptimal solution have ken proposed, as evidenced in the survey paper Taxonomy of multiple target tracking methods", tEE Proceedings on Radar, Sonar and Navigation, Vol. 152, No. 5, October 2005.

Current state of the art algorithms, such as track-oriented multihypothesis tracking (Miff) and multihypothesis Viterbi data association, use a process called clustering that breaks down the overall MU problem into smaller subproblems that are computationally more tractable.

[006] Clustering works by limiting joint multitarget probability computations to only those target tracks whose state estimates are close with respect to a similarity metric or correlation measure.

Clustering allows the overall tracking space to be decomposed into a number of non-overlapping clusters of target tracks at a given time. Each cluster of target tracks is treated without reference to any of the other clusters, which simplifies the subsequent association probability calculations. In prior art methods, such as US patent 5,379,044, clustering of tracks at successive instants in time is carried out but no consideration is given to the time evolution of these clusters. That is to say, the way or ways in which clusters of tracks change from one scan to the next is not a feature of the MU algorithm.

[007] In order to implement many MTT methods, the number and labelling of target tracks must be fixed before association probability calculations can be carried out As further scans of data are received, the clustering of target tracks need not remain the same because the similarity metric between various target tracks changes with time due to their differing dynamics. To apply probabilistic MIT methods, and in particular for the development of digital computer implementations thcrcof, it is uscifil to know the relationship between the track clusters from one scan to the next.

[008] The physical parameters of interest of the objects may also be non-dynamical. For instance, in passive sonar, one is interested in estimating one or more narrowband frequencies emitted by one or more targets, which involves a clustering operation when multiple targets are present. In high frequency over-the-horizon radar, clustering is required to group tracks arising from the same target via different ionospheric multipath propagation modes. From an algorithmic point of view, there is no essential difference between these clustering problems and the clustering problem as it occurs in MTT for active radar and sonar systems. In the remainder of the specification, when the clustering problem is discussed in the context of radar-based multiple target tracking, it is by way of example only and should not be taken as limiting the applicability of the method.

Summary of the Invention

[009] Herein a method is disclosed by which the time evolution of the clustering relationship can be effectively described by a sequence of operations wherein each operation is either an addition or subtraction of target tracks to or from existing clusters. More precisely, the subtraction operation splits a target track from one cluster thereby creating a new cluster and reducing the size of the original cluster. The addition operation merges or combines two target tracks or clusters thereof into a larger cluster.

[0010] By performing the correct sequence of split and merge operations that is output by the method described herein, it is possible to transform the clusters at one time into the clusters at the next time. The sequence of operations then forms the basis for adding or removing target tracks from existing clusters that accords with the updated clustering arrangement. This allows a new set of MTT subproblems to be derived from an existing set of MTT subproblems simply by applying the split and merge operations appropriate to the clustering arrangements.

[0011] It is an object of the present invention to provide a means for the automatic generation of the sequences of split and merge operations reqLtired to manage the clusters that determine the subproblems in a multiple target tracking system.

[0012] The logic of the method applies regardless of whether the number of targets is constant or not, or whether target tracks appear and disappear at arbitrary times.

[0013] The cluster management decisions include: (i) merge a track to an existing input cluster; (ii) split a track front an existing input cluster: (iii) split a track from an existing input cluster and merge it to another existing input cluster; (iv) split a track from an existing input cluster and merge it to a new cluster.

[0014] The method consists of defining two sets of clusters of tracks at each scan: the input cluster set comprising the clusters formed on the previous scan and the output cluster set comprising the clusters formed on the culTent scan.

[0015] Any unmatched output clusters of two tracks or more are rcmoved front the output cluster set.

[0016] Any unmatched singletons (clusters consisting of a single track only) are removed from both the input and output cluster sets.

[001?] Matching criteria are established that account for the number of common tracks between the input and output clusters, the sizes of the input and output clusters, and whether or not the input cluster exactly matches the output cluster.

[0018] A cost thnction is defined that combines, for a given pair of input and output clusters, the matching criteria into a single numerical value.

[0019] Possible pairings between the input and output clusters are evaluated according to the matching criteria and the cost function to give a set of assignment costs.

[0020] The set of assignment costs is augmented to include the cost of a dummy assignment whereby the input cluster does not correspond to any of the output clusters.

[0021] The assignment costs for the input-output cluster pairings are arranged in a matrix.

[0022] The matrix of assignment costs is augmented to include dummy assignment costs.

[0023] The two-dimensional assignment problem defined by the augmented matrix of assignment costs is solved to yield a mapping between input and output clusters.

[0024] According to the mapping from the solution to the assignment problem, a set of companion tracks is established.

[0025] According to the mapping from the solution to the assignment problem and the set of companion tracks, a set of tracks requiring split and/or merge decisions is established.

[0026] The split and/or merge decision for each track requiring a decision is determined according to the mapping between input and output clusters and the appropriate companion tracks.

[0027] In using an optimal assignment problem formulation, one can guarantee that the number of split and/or merge operations generated is a minimum.

Brief Description of the Drawings

[0028] A preferred embodiment of the present invention is next described, by way of example only and simplified where necessary for explanatory purposes, with reference to the accompanying drawings, wherein: FIG. I is a schematic diagram of the main components of a radar tracking system, according to

the prior art;

FIG. 2 is a graph of three target trajectories in position-time space, according to the prior art; FIG. 3 is a graph of multiple position measurements versus time, according to the prior art; FIG. 4 is a graph depicting the tracking of three targets in position-time space with track

clustering shown, according to the prior art;

FIG. 5 is a flowchart of the cluster management method according to the invention.

Detailed Description of the Preferred Embodiment

[0029] High performance multiple target tracking systems as used in radar, sonar, video surveillance and other fields rely on the inclusion of target tracking algorithms that explicitly account for the presence of multiple targets in the surveillance area. Due to the combinatorial nature of the MTT problem, such systems rely for their computer software implementation on a segmentation operation called clustering. The clustering operation simplifies the MTT processing by defining at each scan sets of estimated target tracks that are deemed to be non-interacting. Each cluster of target tracks is effectively an MTT subproblem. The clListers however are not static in time as the dynamical motion of the targets may cause them to be closer to or further from each other at various times in terms of a similarity metric.

[0030] The drawings in FIGS. I -4 are intended to clarify the connection between the various processing stages that are relevant to the formation of clusters of target tracks. This particular application is chosen by way of example only and does not limit the applicability of the cluster management method to other systems or technologies. For clarity, a one-dimensional presentation is used to avoid issues such as different coordinate system choices between the state and measurement variables. This does not limit the scope of the invention.

[0031] FIG. 1 depicts the main components of a prior art microwave radar target tracking system that includes a sensor 002 (in this case, a radar transmit-receive antenna), linked to a signal and data processing unit 006 and associated display console 00K An object 004 inside the surveillance region, an aircraft in this case, reflects or emits signals in the bandwidth of the receive subsystem that impinge on the receiver antenna 002. These signals are processed by die processing unit 006 to form target tracks on the display console 008.

[0032] The aph in FIG. 2 shows three target trajectories in position-time space. For clarity, only one dimension of position (x) has been chosen for this graph, with time (t) on the horizontal axis.

This should not be taken as a limiting factor in the invention, which applies to measurements of arbitrary dimension. The target trajectories 102, 104 and 106 arc shown, by way of example, as sfraight lines to accentuate the fact that these rcprescnt the physical motion of the targets without the corrupting effects of sensor measurement noise.

[0033] FIG. 3 shows a graph of multiple position measurements versus time in the same coordinate system as the graph in FIG. 2. As before, the choice of a single measurement dimension is for clarity of presentation only. The discrete or sampled nature of the measurements is evident on the graph, with three sampling instants k-i, k and ld-l, indicated as 204, shown on the time axis (t). The initial position measurements of the three targets 202 are shown but the remaining measLirements, or detections, have been deliberately not indicated explicitly to outline the nature of the raw sensor measurement data. Two false alarms, or clutter detections, 206 are designated on thc graph; other false alarms are also present in the measurement data in FIG. 3. For clarity, the probability of detection for all three targets has deliberately been taken as unity in this example, althoLigh this is not a necessary restriction.

[0034] The graph in FIG. 4 depicts the tracking of the three targets from FIG. 2 in position-time space. The tracks 302, 304 and 306 have been initiated on the target detections 202 and have been successfully continued to time k-I-i. The presence of sensor measurement noise means that the target tracks, which are estimates, do not exactly follow the true target trajectories even when the correct detections for each target are used in the formation of each track.

[0035] The clustering operation is also exhibited in FIG. 4. Track clustering is generally based on the predicted motion of each target one sean ahead in time. Thus., when the predicted states or predicted measurements of two targets are close in some statistical sense, the two target tracks may be placed in the same cluster. Common measures used for this purpose include the Mahalanobis metric and the chi-squared statistic. Alternatively, the clustering may be based the sharing of measurements by the so-called validation gates of two or more targets. In FIG. 4, target tracks 304 and 306, which are close in position space at times k-i and k, arc in the same cluster, shown by the dashed box 310. At time k-I-i, track 304 is further separated from track 306 than track 302. As a result, the clustering arrangement has changed, with tracks 302 and 306 now being in the same cluster.

[0036] FIG. 4 shows how the clustering arrangement can vaiy from one scan to the next. In this case it is easy to appreciate that a multiple target tracking algorithm that had processed target tracks 304 and 306 together up to time k, should instead at time k+l jointly process target tracks 302 and 306. The corresponding cluster management operations from time k to k+1 are: split track 304 from cluster 310 and merge track 302 to the cluster. When the number of targets is small, it is a relatively simple matter to determine the relationship between the clusters from one time to the next.

[0037] We provide two slightly more complicated examples to thrther motivate the development of the cluster management method. Suppose there are two clusters at time k: {[l, 2,3], [4, 5. 6,7, 8. 9fl and S clusters at time k+1: ff8, 9], [1, 2], [3, 4], 5, [6, 7]}. It can be verified that the following sequence of split and merge operations transforms the clusters at time k to those at time k-I-I: split track 5 from cluster 2 to new cluster 3 split track 3 from cluster ito new cluster 4 split track 4 from cluster 2 and merge to cluster 4 split track 6 from cluster 2 to new cluster 5 split track 7 from cluster 2 to merge to cluster S [0038] The next example is the converse of the previous one. Thus at time k we have 5 clusters: {[8, 9], [i, 2], [3, 4], 5, [6, 7]} and at time k-I-i, 2 clusters: {[1, 2, 3], [4, 5, 6, 7, 8, 9]}. The following sequence of split and merge operations transforms the clusters at time Ic to those at time k+ I: split track 3 from cluster 3 and merge to cluster 2 merge track 4 to cluster 5 split track 8 from cluster I and merge to clusters merge track 5 to cluster S mcrgc track 9 to cluster 5 [0039] As the number of targcts increases, for instance for tcns or cvcn hundreds of targets, an automatic procedure is required to determine the relationship between clusters at successive times.

It is advantageous to provide such a generally applicable method that, given two successive clustering arrangements, produces a sequence of merge and split operations that transforms the first into the second.

[0040] The sequence of merge and split operations required to transform the clusters from time k to time lc+i is not generally unique. For this reason, it is of interest to consider methods that reduce or minimise the number of operations. It is advantageous to provide a means of generating a sequence of operations that is efficient in respect of the number of operations that are so generated.

[0041] Before proceeding with the detailed specification of the cluster managcment method in the general case, we note some of the particular features of the problem. Note that a swapping of one target track from one cluster to another is equivalent to a split followed by a merge so that there is no need to expand on the two basic operations (split and merge). Note also that the order of clusters is not important, only the tracks they contain. The same set of split and merge operations would apply with a corresponding relabelling of the clusters if the problem were posed with the same clusters presented in a different order. Likewise, the order of tracks in a cluster is also unimportant: clusters are unordered sets. Furthermore, since the elements of the clusters are distinct target tracks, the clusters at a given time are mutually disjoint sets.

[0042] In FIG. S a flowchart is given for the cluster management method according to the invention. The detailed description of the processing stages in the method is now addressed. In the interests of readability, the cluster management method is henceforth referred to simply as the method. The inputs to the method are the input cluster set 402 and the output cluster set 404. As clarified in the preceding examples, the input cluster set is the set of track clusters "at time k" while the output cluster set is the set of traclc clusters "at time k+l ", where Ic is the scan number.

[0043] The set of targets contained in the input clusters does not have to be the same as the set of targets contained in the output clusters. The method applies regardless of whether the number of targets is constant or not, or whether target tracks appear and disappear at arbitrary times. To handle such cases, the processing block 406 in FIG. 5 detects the presence of unmatched output clusters, that is, clusters of more than a single track that are present in the output cluster set but whose constituent tracks arc not present in the input cluster set. While such a condition persists, the processing block 408 removes the unmatched output clusters. On exit from block 406 there are no further unmatched output clusters.

[0044] The input and output cluster sets may still contain unmatched single tracks, which are referred to as trivial singletons. The processing block 410 in FIG. 5 detects the presence of trivial singletons in the input cluster set, that is, any tracks that are present in die input cluster set but not in the output cluster set. If trivial singletons exist in the input cluster set, they are removed by processing block 412.

[0045] In a similar maimer, processing block 414 in FIG. 5 detects the presence of trivial singletons in the output cluster set, that is, any tracks that are present in the output cluster set but not in the input cluster set. If trivial singletons exist in the output cluster set, they are removed by processing block 416.

[0046] As noted at 417 in FIG. 5, at this point in the processing the input cluster set and output cluster set have the same set of tracks. Without loss of generality the set of tracks can be indexed as 1 I where T is the number of unique tracks. The number of input clusters at this point is denoted by M while the number of output clusters is denoted by N.T*he numbers M and N are arbitrary and do not have to be equal.

[0047] At 418 in FIG. 5 the cost matrix for the two-dimensional assignment problem is constructed. This is a NI x (N+M) matrix of non-negative numbers. The (i, j) element of the principal M x N submatrix corresponds to the painvise assignment costs between the i-th input cluster and the j-th output cluster. The leftmost M x M block of the cost matrix is referred to as the dummy cost matrix.

[0048] The thnction of the dummy cost matrix is to allow any of the input clusters to be allocated to a dummy variable. This is useful since, in cases where a new output cluster is created front tracks that have been split from one or more of the input clusters, it is advantageous not to direetly assign an input cluster to an output cluster.

[0049] The entries of the principal M x N submatrix of the cost matrix are chosen to favour assignments between input and output clusters that are similar in a certain sense. In the preferred embodiment, discussed below, three different matching criteria are used to determine the costs.

The matching criteria are: the number of common tracks between the input and output clusters, the sizes of the input and output clusters, and whether or not the input cluster exactly matches the output cluster.

[0050] A cost thnction is defined that combines the matching criteria, for a given pair of input and output clusters, into a single numerical value that can be inserted into the cost matrix. For an exact match between the input and output clusters, the cost is defined to be zero. A higher cost is assigned to pairs of clusters that contain some common tracks, with prefcrenee given to pairs of clusters that have more elements. Finally, pairs of clusters that have no elements in common are given the highest cost.

[0051] The diagonal entries of the dummy cost matrix are given by the dummy cost. This is chosen to be less than the highest cost but more than the cost of any assignments that involve a partial match (i.e., that have at least one track in common). The off-diagonal costs of the dummy cost niatrix are taken to be equal to the highest cost.

[0052] To complete the speciflcation of the preferred embodiment of block 418 in FIG. 5, there is provided next a pseudo-code description for the determination of the cost matrix. Recall that at this stage of die processing, there arc no unmatched input or output clusters and no trivial singletons.

The pseudo-code relies on certain definitions and notation that are givcn below.

[0053] Pseudo-code for determination of assignment problem costs.

[0054] Vectot of elements are denoted by square brackets. Vectors include scalars (single elements). Elements of vectors are denoted by parentheses with a single index, e.g., V(i). Elements of matriecs arc denoted by parentheses with a double index, e.g., Q(i,j).

[0055] Tracks are denoted by track indexes that are positive integers.

[0056] A cluster of tracks is a vector of track indexes.

[0057] A singleton is a cluster containing only a single track index.

[0058] An equals symbol is used both in its usual mathematical sense and to denote assigned equality as in software engineering, e.g., j = j + 1.

[0059] Curly braces {} arc used to denote the elements ofa set of vectors. Thus ifC = ff1, 2], [3, 4.5]), then Cf2) = [3,4,5].

[0060] I S is the cardinality or number of elements in the set S. [0061] = set of input clusters containing the vectors of track indexes C11{1}. ..., C0{M}.

[0062] ICH=M [0063] nj11(i) = C11{i}I, i = 1 M [0064] Coot = set of output clusters containing the vectors of track indexes Corn 1} Corn N}.

[0065] I C011 =N [0066] n0i)= COUt(i)L i= I N [0067] N1, = max(M, N) [0068] Nimaxsize = rnax(n1(1), . .., n(M), n011(1) [0069] The M x N matrices, E, S and P. of metrics for the input and output cluster pairs are computed as follows: [0070] fori= 1:M [0071] forj=l:N [0072] n = i} n C011 U) [0073] if n = njn(i) AND n = n001('j) [0074] E(i,j) = [0075] elseifn >0 [0076] S(ij) = rnin(n1(i), n,UEU)) [0077] P(ij) = n [0078] else [0079] E(ij) = 0, S(i.j) = 0, P(iJ) = 0 [0080] end if [0081] endfor [0082] end for [0083] Q = (Nni.axsize + 1) P + S [0084] Q. = set of unique enfries in Q sorted in increasing order (in order of decreasing cost) [0085] NQ = I I [0086] y = vector of assignment cost elements [0087] y(NQ) = 1 [0088] forj= 1:NQ-1 [0089] y(NQ-j) = (Nnaxl) y(NQ-j+l) + 1 [0090] end for [0091] Insertion of dummy costs and reassignment of the highest cost y(l) [0092] S=y(l) [0093] y(l) = (INmaxl) y(1) -1 [0094] The assignment cost matrix [is an v1 x (N+M) matrix of non-negative numbers [0095] The principal Mx N submatrix of the assignment cost matrix is computed as follows: [0096] fori= I:M [0097] forj=1:N [0098] k = rank of Q(i,j) in Q. [0099] if E(i,j) = 1 [00100] F(i,j)=0 [00101] else [00102] [(i,j) = y(k) [00103] end if [00104] end for [00105] end for [00106] The dummy part of the assignment cost matrix is computed as follows: [00107] fori=1:M [00108] forj =N+1:N-{-M [00109] ifl-1 = [00110] f(i,D=o [00111] else [00112] = y(1) [00113] endif [00114] end for [00115] endfor [00116] This completes the description of the pseudo-code for the determination of the costs for the assignment problem. The next stage of the method, shown at 420 in P10.5, involves solving the two-dimensional assignment problem with the M x (N+M) cost matrix obtained as previously described. Numerous prior art approaches are available for this problem including the generalised Munices algorithm (for rectangular cost matrices), the auction algorithm and the Jonker-Volgenant-Castanon algorithm. Alternatively, a suboptinial algorithm such as a greedy algorithm may be used. The solution of the assignment problem can be written as an Mx 2 matrix of indexes that we denote by 1. The entries of the mapping Z. contain the M assignments whose total cost (the sum of the component costs from the cost matrix) is a minimum compared with all other possible assignments. (The minimum need not be strict in the sense that other solutions may also achieve the same minimum cost, but this does not affect the efficacy of the method.) [00117] The next stage of the method, 422 in FIG. 5, involves the determination of the set of companion tracks. A companion track is a track that is in both an input cluster and the output cluster that has been assigned to the input cluster under the mapping Z4. The set of companion tracks is of interest because it contains all the tracks for which no split and/or merge decisions are required. Conversely, the set of tracks for which decisions are required, called the track decision set 424 in FIG. 5, is the complement of the set of companion tracks in the set of all non-trivial tracks labelled 1 to T. Processing block 426 loops over the tracks in the track decision set and for each one of these determines the split and/or merge decision that is requ red. The loop 430 terminates when all tracks in the track decision set have been processed.

[00118] The cluster management decisions, i.e., the sequence of individual split and/or merge decisions required to transform the input cluster set into the output cluster set, are obtained in the processing block 428 in FIG. 5.

[00119] The following pseudo-code, which is part of thc prefelTed embodiment of the method, outputs the cluster management decisions for each track in the track decision set. The mapping of input clusters to output clusters Z should already have been obtained as the solution of the two-dimensional assignment problem the cost matrix of which was previously elaborated. All decisions are in terms of operations on the input clusters. At this stage of the processing there are T non-trivial tracks that arc assumed to be labelled from 1 to I. [00120] Cluster management decisions are of the form: (i) merge (singleton) track to existing input cluster; (ii) split track from existing input cluster (as a new singleton); (iii) split track from existing input cluster and merge to existing input cluster; (iv) split track from existing input cluster and merge (with other split tracks) to new cluster.

[00121] As before, curly braces {} are used to denote the elements of a set of vectors. S is the cardinality of the set S. 5' is the complement of the set S. 0 is the empty set. A\B is the set difference of A with respect to B. [00122] Pseudo-code for track decision logic.

[00123] To determine the set of companion tracks D' (shown as 422 in FIG. 5) [00124] Initialise D' = 0.

[00125] Fori= 1:M [00126] if Z*(i,1) «= M AND Zi,2) «= N [00127] c = Cj{Z'(i,1)} n [00128] D'=D'ur [00129] end if [00130] end for [00131] The track decision set (shown as 424 in FIG. 5) is: [00132] D = {l T}\ D' [00133] for each trackj ED [00134] c = input cluster of track] [00135] n = [00136] c0 = output cluster of track] [00137] CoutCco} [00138] qi = companion track of c E D' oa 0 if none [00139] r0 = companion track of c0 E D' OR 0 if none [00140] ifr0!=0 [00141] r = input cluster of r0 under f [00142] end if [00143] ifr0øi\Nnq=ø [00144] if n1 = 1 [00145] merge track] to input cluster r [00146] elseifn> 1 [00147] split trackj from input cluster c and merge to cluster ri [00148] end if [00149] elseifr0 = 0 AND qi!= 0 [00150] ifn0= 1 [00151] split track] from inpLtt cluster c to new singleton [00152] clscifn0> 1 [00153] split track] from input cluster c and merge to new cluster c0 [00154] end if [00155] e1seifr0!=0ANDq!=0 [00156] split track j from input cluster c and merge to existing input cluster r [00157] end if [00158] end for [00159] This completes the description of the pseudo-code for the determination of the track decisions.

[00160] Many modifications will be apparent to those skilled in the art without departing from the scope of the present invention as herein described with reference to the accompanying drawings.

Claims (13)

1. Claims What is claimed is: I. A duster management method for a multiple target tracking system providing a set of input track clusters and a set of output track clusters pertaining to different times, the method comprising: evaluating each possible pairing between the input and output clusters according to a cost function to give a matrix of assignment costs; solving a two-dimensional assignment problem based on the matrix of assignment costs to yicld a mapping bctween input and output clusters; using the mapping to determine the sequence of clustcr management decisions.
2. 2. A cluster management method according to claim 1, wherein the assiment costs are augmented to include the cost of a dummy assignment whereby the input cluster does not correspond to any of the output clusters.
3. 3. A cluster management method according to claim 1, wherein the cost thnction combines, for a given pair of input and output clusters, a plurality of matching criteria into a single numerical val Lie.
4. 4. A cluster management method according to claim 3, wherein the matching criteria account for the number of common tracks between the input and output clusters, the sizes of the input and output clusters, and whether or not the input cluster exactly matches the output cluster.
5. 5. A cluster management method according to claim 1, wherein the cluster management decisions include: merging a track to an existing input cluster; splitting a track from an existing input cluster; splitting a track from an existing input cluster and merging it to another existing input cluster; splitting a track from an existing input cluster and merging it to a new cluster.
6. 6. A cluster management method according to claim 1, wherein, according to the mapping from the solution to thc assignment problem, a set of companion tracks is established.
7. 7. A cluster management method according to claim 6, wherein, according to the set of companion tracks, a set of tracks requiring decisions is established.
8. S. A cluster management method according to claim 1, wherein track decisions are detennined according to the mapping from the assignment problem and the appropriate companion tracks.
9. 9. A cluster management method according to claim 8, wherein the number of split and/or merge decisions generated is a minimum at each time among all possible sets of such decisions that map the input clustei to the output clusters.
10. 10. A cluster management method according to claim 1, wherein any unmatched output clusters of two tracks or more are removed from the output cluster set.
11. II. A cluster management method according to claim I, wherein any unmatched clusters consisting of a single track only are removed from both the input and output cluster sets.
12. 12. A cluster management method according to claim 1, wherein the assignment problem is linear.
13. 13. A cluster management method according to claim 1, further comprising the application of the cluster management decisions to configure the processing of data in the multiple target tracking system.