 Research Article
 Open Access
 Published:
Tracking target signal strengths on a grid using sparsity
EURASIP Journal on Advances in Signal Processing volume 2014, Article number: 7 (2014)
Abstract
Multitarget tracking is mainly challenged by the nonlinearity present in the measurement equation and the difficulty in fast and accurate data association. To overcome these challenges, the present paper introduces a gridbased model in which the state captures target signal strengths on a known spatial grid (TSSG). This model leads to linear state and measurement equations, which bypass data association and can afford state estimation via sparsityaware Kalman filtering (KF). Leveraging the gridinduced sparsity of the novel model, two types of sparsitycognizant TSSGKF trackers are developed: one effects sparsity through ℓ_{1}norm regularization, and the other invokes sparsity as an extra measurement. Iterative extended KF and GaussNewton algorithms are developed for reducedcomplexity tracking, along with accurate error covariance updates for assessing performance of the resultant sparsityaware state estimators. Based on TSSG state estimates, more informative target position and track estimates can be obtained in a followup step, ensuring that track association and position estimation errors do not propagate back into TSSG state estimates. The novel TSSG trackers do not require knowing the number of targets or their signal strengths and exhibit considerably lower complexity than the benchmark hidden Markov model filter, especially for a large number of targets. Numerical simulations demonstrate that sparsitycognizant trackers enjoy improved rootmeansquare error performance at reduced complexity when compared to their sparsityagnostic counterparts. Comparison with the recently developed additive likelihood moment filter reveals the better performance of the proposed TSSG tracker.
1 Introduction
Target tracking research and development are of major importance and continuously expanding interest to a gamut of traditional and emerging applications, which include radar and sonarbased systems, surveillance and habitat monitoring using distributed wireless sensors, collision avoidance modules envisioned for modern transportation systems, and mobile robot localization and navigation in static and dynamically changing environments, to name a few; see, e.g., [1, 2] and references therein.
At the core of longstanding research issues even for singletarget tracking applications is the nonlinear dependence of the measurements on the desired state estimates, which challenges the performance of linearized Kalman filter (KF) trackers, including the extended (E)KF, the unscented (U)KF, and their iterative variants [1, 2]. This has motivated the development of particle filters (PFs), which can cope with nonlinearities but tend to incur prohibitively high complexity in many critical applications. For multitarget tracking, data association has been another formidable challenge, especially when the ambient environment is cluttered, and the sensors deployed are unreliable. This challenge amounts to determining the target associated with each measurement, where the noisy measurements typically reflect the candidate target locations acquired through signal detection in gated validation regions; see, e.g., [2, 3]. Once data association is established, targets can be tracked separately using the associated measurements, in conjunction with track fusion for improved accuracy.
The present paper investigates the multitarget tracking problem, whereby the available measurements comprise the superposition of received target signal strengths of all targets in the sensor field of view. Sensors collecting these measurements are not necessarily radars or highcost receivers, but can be generalpurpose radio units employing simple energy detectors. The measurements are nonlinearly related to target locations, but no data association issues arise because conventional rangegate operations have not yet been employed to detect, separate, and localize the targets of interest [3]. To cope with the nonlinearity issue, this paper introduces a gridbased dynamical statespace model in which the state describes signal strengths of targets traversing a preselected spatial grid (TSSG) of the tracking field. Because the locations of grid points are preset and known, both the measurement and state equations become linear. Further, data association is avoided by dynamically tracking the TSSG values rather than directly producing the target tracks. Based on TSSG tracking, however, data association and track trajectory estimation can be performed as a followup step, whereby track association and estimation errors do not propagate back to the TSSG tracker.
Similar ideas on bypassing data association at the price of tracking ‘less informative’ estimates have been exploited in recent multitarget tracking schemes, such as the probability hypothesis density (PHD) filter [4, 5] and the Bayesian occupancy filter (BOF) [6]. The PHD filter tracks the sotermed target intensity, while the BOF tracks the probability of a grid point being occupied by any target. A main advantage of the gridbased TSSG tracker here is that state estimation becomes possible via KF applied to linear state and measurement models, at considerably reduced computational burden relative to the complexity incurred by the PHD and BOF. Further, the TSSG tracker is novel in exploiting the sparsity present in the gridbased state vector, which allows one to leverage efficient solvers of (weighted) leastsquares (LS) minimization problems regularized by the ℓ_{1}norm of the desired state estimate.
Sparsityaware estimators have been studied for variable selection in static linear regression problems and have recently gained popularity in signal processing and various other fields in the context of compressive sampling (CS); see, e.g., [7–9]. However, few results pertain to the dynamic scenario encountered with target tracking. When measurements arrive sequentially in time, a sparsityaware recursive leastsquares scheme was reported in [10], but its tracking capability is confined only to slow model variations; see also [11] for a sparsitycognizant smoothing scheme which nevertheless does not lend itself to filtering, as well as [12], where a socalled KFCSresidual scheme is reported for tracking slowly varying sparsity patterns. Different from existing alternatives, the present work develops sparsityaware trackers along with their error covariances, without requiring knowledge on the number of (possibly fastmoving) targets or their signal strengths.
As alluded to in the previous paragraphs, most of the wellknown multitarget tracking algorithms such as the joint probabilistic data association filter (JPDAF) [3], multiple hypothesis tracker (MHT) [2], and the original PHD are derived with a certain measurement model in mind. This model exerts that every measurement comes from either a single target or clutter and each target generates at most one measurement. On the other hand, existing literature on superpositional sensors, which are investigated here, is considerably more sparse. A cardinalized PHD (CPHD) filter for such sensors was developed by [13] which is computationally intractable. An approximate PHD filter for superpositional sensors was developed by [14] which was referred to as the additive likelihood moment (ALM) filter. It was later extended to an approximate CPHD by [15]. Our comparison reveals that the TSSG tracker outperforms the ALM filter by a comfortable margin.
This work extends the results of [16] in four major directions:

(1)
It introduces the novel TSSGIEKF tracker which is based on the idea that sparsity can be enforced as an extra measurement.

(2)
In addition to the TSSG estimate, target positions and tracks are also computed from the TSSG. This requires tackling several issues such as clustering grid points using their TSSG values for an unknown number of targets, computing the targets’ estimated positions, and performing positiontotrack association which is needed to arrive at tracks from mere positions.

(3)
Instead of presenting the multitarget model as an ad hoc extension of the single target case, a rigorous proof is set forth in the appendix which leads to the linear target movement model on the grid. During the process, the two important assumptions that are needed for this model to remain valid are presented.

(4)
A complete set of multitarget scenarios and their corresponding performance analysis is investigated in the simulations. The performance of the newly introduced TSSGIEKF tracker is also examined, and comparisons with ALM are carried out.
The rest of the paper is organized as follows. Section 2 develops the novel gridbased sparse model, for which a sparsityagnostic KF tracker is introduced in Section 3. Two sparsitycognizant trackers are presented in Sections 4 and 5. Target position estimation and track formation are detailed in Section 6. Numerical results are presented in Section 7, followed by concluding remarks in Section 8.
2 Gridbased state space model
Consider the problem of tracking M moving targets using N active (e.g., radar) or passive (e.g., acoustic) sensors deployed to provide situational awareness over a geographical area. Targets emit power either because they passively reflect the energy of other transmitters such as radar or because they are active sources such as cell phones or transmitters mounted on smart cars. Associated with each target, say the m th one, is its position vector {\mathbf{p}}_{k}^{\left(m\right)} per time k and the signal of strength s^{(m)} that the target reflects or emits. Sensor n measures the superposition of received target signal strengths,
where n = 1,…,N, k = 1,2,…, h (•) denotes the distancedependent propagation function, {d}_{k}^{m\to n}:=\parallel {\mathbf{p}}_{k}^{\left(m\right)}{\mathbf{q}}_{n}{\parallel}_{2} is the distance between the known position q_{ n } of sensor n and the unknown position vector {\mathbf{p}}_{k}^{\left(m\right)} of target m, and ν_{n,k} is a zeromean Gaussian noise at sensor n. Function h (•) satisfies h(0) = 1, is nonnegative, decreasing, and is either assumed known from the physics of propagation or acquired through training [9].
At each time k, a centralized processor has available the measurement vector y_{ k }: = [y_{1,k},…,y_{N,k}]^{T}, based on which the target positions {\left\{{\mathbf{p}}_{k}^{\left(m\right)}\right\}}_{m=1}^{M} are to be tracked. Note that the measurement model (1) differs from the one typically considered in radar applications, where a measurement comes from either a single target or a clutter, usually in the form of position information obtained from range gate operations [3]. Each measurement in (1) comes from a sensor and comprises the superposition of received signal strengths emitted by or reflected from all targets in the sensor field of view. This model considers the localization and tracking problems jointly and avoids the measurementtarget association issue.
One major challenge in tracking and localization problems is that the measurements in (1) are nonlinear functions of the wanted target position vectors. A neat approach to arrive at a linear measurement model is to adopt a set of G (possibly regularly spaced) grid points at known positions {\left\{{\mathbf{g}}_{i}\right\}}_{i=1}^{G}, where target(s) could be potentially located; see also, e.g., [6], [8], and [7]. Using a sufficiently dense grid, it is possible to capture the target locations at a prescribed spatial resolution using a G×1 vector x_{ k } having most entries equal to zero except for the {\left\{{i}_{k}^{\left(m\right)}\right\}}_{m=1}^{M} entries given by {\left\{{x}_{k}^{\left({i}_{k}^{\left(m\right)}\right)}\right\}}_{m=1}^{M}, which represent the target signal strengths at time k if and only if the m th target is located at the {i}_{k}^{\left(m\right)}th grid point, that is, {\mathbf{p}}_{k}^{\left(m\right)}={\mathbf{g}}_{{i}_{k}^{\left(m\right)}}. Note that if target m is located exactly on a grid point {i}_{k}^{\left(m\right)}, then {x}_{k}^{\left({i}_{k}^{\left(m\right)}\right)}\equiv {s}^{\left(m\right)}\ne 0 will be the only nonzero entry of x_{ k } corresponding to this target. However, to account for target presence off the preselected grid points, it will be allowed for the unknown target signal strength s^{(m)} to ‘spill over’ grid points around {i}_{k}^{\left(m\right)} and thus render nonzero a few neighboring entries of x_{ k }. Let {\mathcal{G}}_{k}^{\left(m\right)} denote the spillover region on the grid corresponding to target m at time k, such that {x}_{k}^{\left(i\right)}\ne 0 is associated with s^{(m)}, \forall i\in {\mathcal{G}}_{k}^{\left(m\right)}. The following assumption on this target occupancy model is imposed:
Assumption 1. Each grid point i can be occupied by at most one target m at any given time k.
This assumption can be easily satisfied in practice by selecting a sufficiently dense grid [6, 17]. Under Assumption 1, each grid point i is associated with a unique target index {m}_{k}^{\left(i\right)} at time k; that is, i\in {\mathcal{G}}_{k}^{\left({m}_{k}^{\left(i\right)}\right)}, where {m}_{k}^{\left(i\right)}\in \phantom{\rule{0.3em}{0ex}}[\phantom{\rule{0.3em}{0ex}}1,M] if it is occupied by one of the M targets, or {m}_{k}^{\left(i\right)}=0 if it is not occupied, meaning it is associated with a dummy target m=0 with strength s^{(0)}≡ 0. Apparently, {\left\{{\mathcal{G}}_{k}^{\left(m\right)}\right\}}_{m=0}^{M} are mutually exclusive across m and their union spans the entire grid in the sense {\cup}_{m=0}^{M}{\mathcal{G}}_{k}^{\left(m\right)}={\cup}_{i=1}^{G}i, which leads to a measurement equation (cf. (1)):
Here, {\mathbf{h}}_{n}^{T}:=\left[\phantom{\rule{0.3em}{0ex}}h\right({d}^{1\to n}),h({d}^{2\to n}),\dots ,h({d}^{G\to n}\left)\right], d^{i → n}:= ∥ q_{ n } g_{ i }∥_{2} now denotes the known timeinvariant distance between the n th sensor and the i th grid point, and the noise v_{n,k} replacing ν_{n,k} in (1) captures the unmodelled dynamics in the aforementioned spillover effect. Notwithstanding, thanks to the gridbased model, the measurements in (2) have become linear functions of the unknown x_{ k }, whose nonzero entries reveal the grid points where target signal strengths are present at time k.
The next step is to model the evolution of x_{ k } in time as the targets move across the grid. Regarding their movement pattern, targets obey the following assumption:
Assumption 2. All targets move according to identical transition probabilities {\left\{\phantom{\rule{0.3em}{0ex}}{f}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}\mathit{\text{ji}}\right)}\right\}}_{i,j=1}^{G}, where {f}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}\mathit{\text{ji}}\right)}:=p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0{x}_{k1}^{\left(i\right)}\ne 0;j\in {\mathcal{G}}_{k}^{\left(m\right)},i\in {\mathcal{G}}_{k1}^{\left(m\right)}), m = 1,…,M.
In words, the homogeneity of targets under Assumption 2 refer to the probability that a target m moves from grid point i at time k  1 to point j at time k.
Consider now expressing each entry of x_{ k } as {x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}={s}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\u2022p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\phantom{\rule{0.3em}{0ex}}\ne \phantom{\rule{0.3em}{0ex}}0), where {s}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\phantom{\rule{0.3em}{0ex}}=\phantom{\rule{0.3em}{0ex}}{s}^{\left({m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\right)}\phantom{\rule{0.3em}{0ex}}\in \phantom{\rule{0.3em}{0ex}}\{{s}^{\left(0\right)},{s}^{\left(1\right)},\dots ,{s}^{\left(M\right)}\} denotes a nonnegative proportionality constant, and p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0) stands for the probability of a target to be present on grid point j at time k. Essentially, each {x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} is associated with only one of the (M + 1) targets (including the dummy target m = 0) indexed by {m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}, and {s}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} is a proportionality constant in the sense that it takes on (M+1) possible values {s}^{\left(m\right)}=\sum _{j\in {\mathcal{G}}_{k}^{\left(m\right)}}{x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}, for m = 0,1,… M.
Under Assumptions 1 and 2, it is shown in the Appendix that the state obeys the following recursion:
Concatenating (3) for j = 1,…,G and (2) for n = 1,…,N, one arrives at the gridbased model:
where the G × G state transition matrix F_{ k } has its (i,j)th entry given by {f}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}\mathit{\text{ji}}\right)}; the measurement matrix is defined as H:=[h_{1},…,h_{ n }]^{T}, likewise for the measurement noise vector v_{ k }: = [v_{1,k},…,v_{N,k}]; and w_{ k } is a zeromean process noise vector with a positivedefinite covariance matrix Q_{ k } added to account for both Assumption 1 and the natural nonnegativity constraints on x_{ k }, whose entries represent target signal strengths (magnitudes or power).
A distinct feature of model (4) is that the unknown x_{ k } is sparse ∀k, since only few of its G entries are nonzero (in fact exactly M nonzero entries if all the M targets are located on grid points). Although (3) describes the linear evolution of each x_{ k } entry under Assumption 4a) will be nonsparse, but the true state vector x_{ k } at any time k is sparse with only a few nonzero entries around the target locations. Posterior to processing the measurements, filtered and predicted renditions of x_{ k } will remain sparse as well. The noise term w_{ k } reflects the uncertainty in the state transition model under Assumption 1.
This sparsity attribute will prove to be instrumental for enhancing tracking performance. Also, it is worth noting that the state transition matrix F_{ k } reflects the transition behavior of target positions only, without revealing full information of the target movement model that may be dependent on velocity or other factors as well. In fact, F_{ k } is derived from the target movement model but does not fully reveal it, which differs from most existing track state models.
Given y_{1 : k}: = {y_{1},…,y_{ k }}, the goal of this paper is to track x_{ k } using a sparsityaware KF. Since x_{ k } represents the target signal strength on the grid (TSSG), the KFlike algorithms proposed in Sections 3 and 4 will be referred to as TSSGKF trackers, while the iterated extended Kalman filter (IEKF) algorithms of Section 5 will be referred to as TSSGIEKF trackers. Having available {\widehat{x}}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} estimates and recalling that {x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}={s}^{\left({m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\right)}p({s}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0), one can estimate the constant s^{(m)} capturing the signal strength of the m th target at time k as
and the corresponding target position vector at time k as
The following remark makes useful observations regarding the position estimate in (6).
Remark 1. A TSSG filter for tracking x_{ k } avoids data association, because the TSSGbased state and measurement equations in (4) hold for any targetgrid association {\left\{{\mathcal{G}}_{k}^{\left(m\right)}\right\}}_{m}, so long as Assumptions 1 and 2 are satisfied. On the other hand, finding the target positions via (6) requires knowledge of {\left\{{\mathcal{G}}_{k}^{\left(m\right)}\right\}}_{m} and hence calls for associating targets with TSSG entries. Solution to such an association problem is provided in Section 6. Nonetheless, it is worth stressing that the association errors and resultant position estimation errors do not affect TSSG tracking that is independent of target position estimation, similar to the PHD and BOF in [4] and [6], respectively.
In addition to reduced complexity, an attractive feature of the present formulation relative to, e.g., [6], is that even for finite G, there is no need to assume that targets are located on grid points since (6) allows for interpolating the target position vectors regardless of knowing that grid point j is associated with the target {m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} occupying it. The next remark is useful to further appreciate this point:
Remark 2. Given measurements y_{1:k} and supposing that the number of targets M and their signal strengths {s^{(1)},…,s^{(M)}} are known, the maximum a posteriori (MAP) and minimum meansquare error (MMSE) optimal trackers can be derived from a hidden Markov model (HMM) filter implementing the recursions (7) derived from Bayes’ rule (cf. (34) and (35) in the Appendix), where {f}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}\mathit{\text{ji}}\right)} is the transition probability as in (3).
These HMM recursions hinge on prior knowledge of the targetgrid association {\left\{{\mathcal{G}}_{k}^{\left(m\right)}\right\}}_{m=0}^{M}, which needs to be figured out among a total of (M + 1)^{G  M}G!/(G  M)! possible combinations. A large G increases grid density and hence spatial resolution, at the expense of increasing complexity. In addition, M and {\left\{{s}^{\left(m\right)}\right\}}_{m=1}^{M} need to be known beforehand.
One additional remark is now in order:
Remark 3. Although y_{ k } in (4b) comprises scalar measurements from N geographically distributed sensors per time k, it is possible to form y_{ k } with samples of the continuoustime signal received at a single (e.g., a radar or sonar) sensor by oversampling at a rate faster than the rate x_{ k } changes, so long as the statespace model (4) is guaranteed to be observable (and thus x_{ k } is ensured to be identifiable).
3 KF for tracking TSSG
If the nonnegativity constraints for x_{ k } were absent, the optimal state estimator for (4) in the MAP, MMSE, or LS error sense would be the clairvoyant linear KF. A pertinent state estimator is pursued here in the presence of nonnegativity constraints. Suppose that the estimate {\widehat{\mathbf{x}}}_{k1k1} and its error covariance matrix P_{k  1k  1} are available from the previous time step. At time k, the KF state predictor and its error covariance are obtained as
For the KF corrector update, consider the LS formulation of the KF; see, e.g., [18]. The corrector update can be derived as a regularized LS criterion, which will also be useful to account for the sparsity attribute. To show this, view {\widehat{\mathbf{x}}}_{kk1} as a noisy measurement of x_{ k }. It follows readily from (8) that {\widehat{\mathbf{x}}}_{kk1}={\mathbf{x}}_{k}+{\mathbf{e}}_{kk1}, where e_{kk  1} has covariance matrix P_{kk  1}. Stacking {\widehat{\mathbf{x}}}_{kk1} and y_{ k } to form an augmented measurement vector yields the following linear regression model:
where the augmented noise vector has block diagonal covariance matrix denoted as diag(P_{kk  1},R_{ k }). The weighted LS (WLS) estimator for this linear regression problem is given by
where \parallel \mathbf{x}{\parallel}_{\mathbf{A}}^{2}:={\mathbf{x}}^{T}\mathbf{A}\mathbf{x}. In the absence of nonnegativity constraints, the optimal state corrector {\widehat{\mathbf{x}}}_{kk} can be found in closed form as the cost is quadratic, and likewise its error covariance can be updated as
A gradient projection algorithm is developed in Section 4 to solve (9) under nonnegativity constraints on the state vector. However, (10) will still be used bearing in mind that this update is approximate now. The TSSGKF tracker implemented by (8) to (10) is sparsityagnostic, as it does not explicitly utilize the prior knowledge that x_{ k } is sparse.
4 Sparsityaware KF trackers
Taking sparsity into account, this section develops sparsitycognizant trackers. To this end, the degree of sparsity quantified by the number of nonzero entries of x_{ k }, namely, the ℓ_{0} norm ∥ x_{ k }∥ _{0}, can be used to regularize the LS cost of the previous section. Unfortunately, similar to compressed sensing formulations for solving underdetermined linear systems of equations [19], such a regularization results in a nonconvex optimization problem that is NPhard to solve and motivates relaxing the ℓ_{0} norm with its closest convex approximation, namely, the ℓ_{1} norm. Thus, the proposed sparsitycognizant tracker is based on the state corrector minimizing the following ℓ_{1}regularized WLS cost function
The state corrector minimizing (11), together with the covariance update^{a} in (10) and the prediction step in (8), forms the recursions of the sparsityaware TSSGKF tracker. Relevant design choices and algorithms for minimizing (11) is elaborated in the next subsection.
The TSSGKF trackers in (9) and (11) involve both prediction and correction steps, which interestingly can be combined into a single estimation step. Considering that both x_{k 1} and x_{ k } are sparse and nonnegative and combining the LS terms for both the prediction and correction steps, the following optimization problem arises for some nonnegative λ_{k 1} and λ_{ k } parameters:
The performance gain of this tracker was evaluated via simulations, and no substantial improvement over the TSSGKF tracker was observed. For this reason, focus henceforth will be placed on the TSSGKF tracker in (11).
4.1 Parameter selection
The scalar parameter λ_{ k } in (11) controls the sparsitybias tradeoff [20]. The corrector {\widehat{\mathbf{x}}}_{kk} becomes increasingly sparse as λ_{ k } increases, and eventually vanishes, i.e., {\widehat{\mathbf{x}}}_{kk}=\mathbf{0}, when λ_{ k } exceeds an upper bound {\stackrel{\u0304}{\lambda}}_{k}. There are two systematic means of selecting λ_{ k }. The first one popular for variable selection in linear regressions is crossvalidation [20], pp. 241249. The second one is the sotermed absolute variance deviationbased selection that has been advocated in the context of outlier rejection setups [21]. Both approaches require solving (11) for different trial values of λ_{ k }. This can be certainly affordable for offline solvers of a linear regression problem or a fixedinterval smoothing scenario, but may incur prohibitive delays for realtime applications, even though warm starts reduce the computational burden considerably. For the tracking problem at hand, the simple rule advocated is to set {\lambda}_{k}=\alpha {\stackrel{\u0304}{\lambda}}_{k}, where α ∈ (0,1) is a fixed scaling value to avoid the trivial solution {\widehat{\mathbf{x}}}_{kk}=\mathbf{0}. The bound {\stackrel{\u0304}{\lambda}}_{k} is derived next.
Proposition 1. The solution to (11) reduces to {\widehat{\mathbf{x}}}_{kk}=\mathbf{0} for any scalar {\lambda}_{k}\ge {\stackrel{\u0304}{\lambda}}_{k} , where
Proof. Since x_{ k }≥ 0, it holds that \parallel {\mathbf{x}}_{k}{\parallel}_{1}={\mathbf{x}}_{k}^{T}\mathbf{1}, where 1 denotes the allone vector. Therefore, J(x) in (11) is differentiable and results in a convex problem. The necessary and sufficient optimality condition states that x^{∗} is an optimum point iff (yx^{∗})^{T}∇J(x^{∗}) ≥ 0, ∀ y ≥ 0. For x^{∗} = 0, this condition holds iff ∇J(x^{∗}) ≥ 0. It then follows from (11) that
Therefore, x^{∗}= 0 is an optimal solution iff (13) holds.
4.2 Gradient projection algorithms
As (11) is a convex problem, convex optimization software such as SeDuMi [22] can be utilized to solve it efficiently. In addition to these solvers, lowcomplexity iterative methods are developed here, by adopting the gradient projection (GP) algorithms in [23], pp. 212217. Note that the proposed algorithms can be used to obtain the sparsityagnostic tracker from (9) too, since the latter is obtained by minimizing a special case of (11) corresponding to λ_{ k }= 0.
At each time k, the GP is initialized with {\widehat{\mathbf{x}}}_{kk}\left(0\right)={\widehat{\mathbf{x}}}_{kk1} at iteration l = 0. The state corrector iterates from l to (l + 1) as follows:
where [ x] ^{+} denotes the projection onto the nonnegative orthant, γ is the step size, and ∇J is as in (14). Here, J(x_{ k }) is differentiable because \parallel {\mathbf{x}}_{k}{\parallel}_{1}={\mathbf{x}}_{k}^{T}\mathbf{1} when x_{ k }≥ 0.
While (15) amounts to a Jacobitype iteration updating all the entries at once, one can also devise GaussSeidel variants, where entries are updated one at a time [23], pp. 218219. This is possible because the nonnegative orthant is a constraint set expressible as the Cartesian product of onedimensional sets, allowing entrywise updates per iteration (l+1) as
where {\stackrel{~}{\mathbf{x}}}_{kk}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\left(l\right)\phantom{\rule{0.3em}{0ex}}:=\phantom{\rule{0.3em}{0ex}}\left\{{\widehat{\mathbf{x}}}_{kk}^{(1:j1)}(l\phantom{\rule{0.3em}{0ex}}+\phantom{\rule{0.3em}{0ex}}1),{\widehat{\mathbf{x}}}_{kk}^{(\phantom{\rule{0.3em}{0ex}}j:G)}\left(l\right)\right\} has its first (j1) entries already updated in the (l+1)st iteration. Convergence of the iterations in (16) to the optimum solution of (11) is guaranteed under mild conditions by the results in [23], p. 219. Specifically, J(x_{ k }) should be nonnegative and its gradient should be Lipschitz continuous, both of which hold for the objective in (11).
Proposition 2. Any limit point of the sequence generated by (16), with arbitrary initialization{\widehat{\mathbf{x}}}_{kk}^{\left(0\right)}, is an optimal solution of (11) provided that the step size γ is chosen small enough. In practice, only a few gradientprojection iterations are run per time step k to allow for realtime sparsityaware KF tracking.
5 Enhanced sparsityaware IEKF tracking
The proposed sparsityaware tracker employs the KF covariance recursion in (10) to update the error covariance of the corrector state estimate. As it does not account for the ℓ_{1}norm regularization, this update is approximate. In order to incorporate the prior knowledge of sparsity when updating the corrector covariance, this section develops an EKFbased approach, which leads to enhanced tracking performance.
Toward this objective, the prior information on sparsity is viewed as an extra measurement μ_{ k }= ∥ x_{ k }∥ _{0}, rather than as a regularizing term in the LS cost function. When the number of targets M is known, an apparent choice is to set μ_{ k }= M. Accordingly, tracking will be carried out based on an augmented (N+1) × 1 measurement vector, given by
5.1 Viewing sparsity as an extra measurement
The added measurement can be modeled in a general form as
where ρ (x_{ k }) is a differentiable function approximating the sparsityinducing ℓ_{0}norm, and u_{ k } denotes zeromean noise with variance {\sigma}_{k}^{2}. The noise term captures both the uncertainty in approximating ∥x_{ k }∥ _{0}, as well as the error in attaining the desired degree of sparsity. As to ρ (x_{ k }), three wellknown approximants of the ℓ_{0} norm are the ℓ_{1} norm, the logarithm, and the inverse Gaussian functions:
where δ and σ_{ p } are tuning parameters, and only x_{ k }≥ 0 is considered. These nonlinear functions are plotted along with the ℓ_{0}norm function for a scalar x_{ k } in Figure 1. It can be seen that they all have relatively sharp edges around the origin to approximate the ℓ_{0}norm.
Adding the extra measurement μ_{ k }, the state space model in (4) is augmented to
where \stackrel{\u0304}{\mathbf{h}}\left({\mathbf{x}}_{k}\right):={[\phantom{\rule{0.3em}{0ex}}{\left(\mathbf{H}{\mathbf{x}}_{k}\right)}^{T},\phantom{\rule{1em}{0ex}}{\rho}_{k}({\mathbf{x}}_{k}\left)\right]}^{T} consists of N+1 scalar measurement functions that can be nonlinear in general, and {\stackrel{\u0304}{\mathbf{v}}}_{k}:={[\phantom{\rule{0.3em}{0ex}}{\mathbf{v}}_{k}^{T},\phantom{\rule{1em}{0ex}}{u}_{k}]}^{T} has covariance {\stackrel{\u0304}{\mathbf{R}}}_{k}:=\text{diag}({\mathbf{R}}_{k},{\sigma}_{k}^{2}). Similar to (11), the model in (17) leads to a nonlinear LS (NLS) problem
Compared with (11), (18) replaces the ℓ_{1}norm of x_{ k } with an alternative LSerror regularization involving the extra measurement which accounts for the sparsity present. Because (18) directly results from (17), the error covariance of state estimates can be updated using the KFlike recursions developed next.
5.2 IEKF algorithm for nonlinear measurement models
Since the augmented {\stackrel{\u0304}{\mathbf{y}}}_{k} in (17b) is a nonlinear function of the wanted TSSG state, the EKF approach is adopted here to update the error covariance along the lines of, e.g., [1], Chap. 10. Specifically, an iterated EKF (IEKF) algorithm is employed, which is tantamount to applying GaussNewton iterations to a relevant NLS regression problem [24].
The prediction step of the IEKF is similar to KF; hence, {\widehat{\mathbf{x}}}_{kk1} and {\widehat{\mathbf{P}}}_{kk1} follow directly from the state space model in (17) and coincide with (8). For the correction step per time k, IEKF recursions are initialized with {\widehat{\mathbf{x}}}_{kk}\left(0\right)={\widehat{\mathbf{x}}}_{kk1} for l = 0, and subsequent iterations proceed as follows [25], Appendix C:
where \mathit{\Phi}\left(l\right):=\nabla \stackrel{\u0304}{\mathbf{h}}{\left({\widehat{\mathbf{x}}}_{kk}\left(l\right)\right)}^{T} denotes the Jacobian matrix of \stackrel{\u0304}{\mathbf{h}}(\u2022) evaluated at {\widehat{\mathbf{x}}}_{kk}\left(l\right). After the IEKF iterations are completed at l = L, the corrector’s error covariance matrix is updated as
The ensuing proposition establishes the link between IEKF and GaussNewton iterations for the related NLS problem.
Proposition 3. Consider the NLS problem (cf. (17) and (18))
Solving (21) via GaussNewton iterations initialized with{\widehat{\mathbf{x}}}_{kk}\left(0\right)={\widehat{\mathbf{x}}}_{kk1}amounts to the IEKF recursions in (19).
Proof. The quadratic terms in (21) can be rewritten as
GaussNewton iterations for (22) become
where \mathit{\Psi}\left(l\right):=\nabla \mathbf{g}\left({\widehat{\mathbf{x}}}_{kk}\right(l\left)\right) is the Jacobian transpose evaluated at {\widehat{\mathbf{x}}}_{kk}\left(l\right). Substituting g(.) from (23) into (24) and applying the matrix inversion lemma to invert the matrix in (24) yield (19) after straightforward algebraic manipulations.
When GaussNewton iterations in (24) are adopted in lieu of IEKF, the resulting error covariance matrix is a function of ∇g at the last iteration L given by
The sparsityaware EKF formulation in (18) is a special case of the general NLS problem in (21) corresponding to \stackrel{\u0304}{\mathbf{h}}\left({\mathbf{x}}_{k}\right):={[\phantom{\rule{0.3em}{0ex}}{\left(\mathbf{H}{\mathbf{x}}_{k}\right)}^{T},\phantom{\rule{1em}{0ex}}{\rho}_{k}({\mathbf{x}}_{k}\left)\right]}^{T}. As a result, the error covariance for the state estimate of (18) can be derived from (25) as
Compared with (10) for the sparsityagnostic KF, the last summand in (26) captures the effect of the sparsitypromoting penalty term on the error covariance. To enforce the nonnegativity constraints in (18), one can project each GaussNewton iterate in (24) onto the nonnegative orthant. Unfortunately, this may not generate a convergent sequence [23], p. 215. To ensure convergence, the projection should be with respect to a different distance metric than the usual Euclidean distance. Upon defining B(l):=(Ψ(l)Ψ^{T}(l))^{1}, one implements
where [.] B + denotes projection onto the nonnegative orthant, which minimizes the \parallel .{\parallel}_{\mathbf{B}}^{2} distance instead of the usual \parallel .{\parallel}_{2}^{2}. If \rho \left({\mathbf{x}}_{k}\right)={\mathbf{x}}_{k}^{T}\mathbf{1}, which is equivalent to the ℓ_{1} norm for x_{ k }≥ 0, then (18) becomes convex, and generalpurpose convex solvers such as SeDuMi can also be utilized to solve it [22].
The iterative updates in (27) and (26), along with the prediction step (8), constitute the sparsityaware TSSGIEKF tracker.
5.3 Enhanced sparsityaware KF tracker
As a final note, the sparsityaware TSSGKF tracker in Section 4 can be enhanced by also casting the ℓ_{1}regularized WLS cost in (11) as an NLS cost. The ℓ_{1}norm term in (11) can be equivalently expressed as an extra LS error term for the extra measurement 0=\sqrt{2\lambda}\sqrt{{\mathbf{x}}_{k}^{T}\mathbf{1}}+{u}_{k}, where u_{ k } is zeromean noise with unitvariance. The corresponding covariance update can be derived from (25) as
In all, the state update in (11), together with the prediction step in (8) and the refined covariance update in (28) with x_{ k } replaced by {\widehat{\mathbf{x}\phantom{\rule{0.3em}{0ex}}}}_{kk}, forms the recursions of the enhanced sparsityaware TSSGKF tracker.
6 Position estimation and track formation
The TSSG filters developed so far produce a dynamic TSS map of the operational environment. Such information is adequate to describe the targets’ distribution and spatial occupancy over the sensing field of interest, similar in spirit to the PHD filter which portrays the targets’ intensity function and the BOF that depicts their occupancy map. In many tracking applications, however, more informative estimates such as target positions and trajectories are desired. This section provides TSSGbased solutions to these estimation tasks too.
For the PHD approach, methods performing these extra steps have been reported using particle PHD filters [26–28] or Gaussian mixture (GM)PHD filters [29]. Target positions are typically identified by peakpicking the target intensity function being tracked, and the estimated target positions are treated as measurements for the ensuing data association and track recovery tasks. PHD filters view each particle or each Gaussian component involved as a target [4, 5] and employ conventional target movement models to describe the state transition. As a result, most of the wellknown data association methods can be run after PHD filtering [3], [2], Chapters 67. Examples include the auction algorithm proposed in [27] and the joint probabilistic data association algorithm [30]. Likewise, for the BOF, the target movement model is employed in updating the HMM filter, which makes it feasible to be combined with a wellestablished data association method such as the JPDAF [30].
In contrast, the TSSG state equation only models the dynamic behavior of the TSS distribution on the grid, in which grid points are not treated as targets, and hence do not directly obey the conventional target movement model. As remarked in Section 2, only partial information about position changes is explicitly captured by the state transition matrix F_{ k }, while other factors such as velocity are implicit. Due to this major difference, conventional data association methods cannot be directly adopted as a followup to TSSG filtering. This section develops estimators of target positions and tracks for multitarget scenarios, based solely on the limited information regarding target transition probabilities on the grid.
6.1 Target position estimation
Given the output \widehat{\mathbf{x}}{}_{kk} of the TSSG filter, target positions can be obtained from (6) provided that the subset of grid points associated with each target is known in the form of {\mathcal{G}}_{k}^{\left(m\right)}, ∀ m.
Starting from {\widehat{\mathbf{x}}}_{kk}, one can apply appropriate clustering techniques to identify {\mathcal{G}}_{k}^{\left(m\right)}. When the number of targets M is known, simple parametric clustering methods such as the k  means can be used [31], pp. 424–429. When M is unknown, one can perform joint clustering and model order selection. Such algorithms utilize some global model order selection criteria such as Akaike’s information criterion to determine the best number of clusters \widehat{M}, as well as the clusters {\left\{{\widehat{\mathcal{G}}}_{k}^{\left(m\right)}\right\}}_{m=1}^{\widehat{M}} themselves [32]. Other nonparametric clustering methods can be employed as well, without assuming or estimating the number of clusters. For example, hierarchical clustering techniques either aggregate or divide the data based on some proximity measure, while density estimationbased nonparametric approaches identify clusters and their number from the modes of the empirical density function of the unknowns (see, e.g., [33], for a survey).
Having acquired \widehat{M} and {\left\{{\widehat{\mathcal{G}}}_{k}^{\left(m\right)}\right\}}_{m=1}^{\widehat{M}}, and based on (6), the target positions can be obtained individually from the TSSG estimates on the associated clusters of grid points \forall i\in {\widehat{\mathcal{G}}}_{k}^{\left(m\right)}, as follows:
6.2 Positiontotrack association
Suppose that there are M_{ t } tracks from time slot 1 up to k1 and {\widehat{\mathbf{p}}}_{k1}^{\left(m\right)} has been associated with track t and hence alternatively expressed as {\widehat{\mathbf{p}}}_{k1}^{\left(t\right)}, t = 1,…,M_{ t }. The goal of track association is to assign the position estimates {\left\{{\widehat{\mathbf{p}}}_{k}^{\left(m\right)}\right\}}_{m=1}^{M} of the M targets at time k to one of the established M_{ t } tracks. For clarity in exposition, suppose first that M = M_{ t } and there is no target birth or death. This assumption will be removed later on. Evidently, there are M! different assignments, which must be examined to find the best possible association.
Given y_{1 : k  1}, the first step is to establish a track prediction model to be used for computing the predicted track positions {\left\{{\widehat{\mathbf{P}}}_{kk1}^{\left(t\right)}\right\}}_{t=1}^{{M}_{t}} and their error covariances. Note from (29) that the target position estimates conditioned on the TSSG are independent of the persensor measurements. Hence, it suffices to predict {\left\{{\widehat{\mathbf{p}}}_{kk1}^{\left(t\right)}\right\}}_{t} solely from the TSSG vector {\widehat{\mathbf{x}}}_{k1k1}. To do so, focus on track t and form a G × 1 vector {\stackrel{\u030c}{\mathbf{x}}}_{k1,t} that only retains the entries of {\widehat{\mathbf{x}}}_{k1k1} belonging to the t th cluster of grid points in {\mathcal{G}}_{k1}^{\left(t\right)}; that is, {\stackrel{\u030c}{x}}_{k1,t}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}={\widehat{x}}_{k1}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} for j\in {\mathcal{G}}_{k1}^{\left(t\right)} and {\stackrel{\u030c}{x}}_{k1,t}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}=0 otherwise, ∀ j.
Given {\stackrel{\u030c}{\mathbf{x}}}_{k1,t} at time k  1, the predicted TSSG belonging to track t at time k becomes
and correspondingly, the predicted track position is
The normalized quantities {\stackrel{\u030c}{x}}_{kk1,t}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}/\left({\sum}_{j=1}^{G}{\stackrel{\u030c}{x}}_{kk1,t}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\right) in (30) play the role of fractional weights when the corresponding grid positions g_{ j } are used to estimate the track position. Viewing {\widehat{\mathbf{p}}}_{kk1}^{\left(t\right)} as the weighted average of G positionsamples {\left\{{\mathbf{g}}_{j}\right\}}_{j=1}^{G}, it is straightforward to estimate the covariance of {\widehat{\mathbf{P}}}_{kk1}^{\left(t\right)} using the sample covariance as
The process in (30) to (31) is repeated for all target tracks t = 1,…,M_{ t }, so that the prediction estimates and covariances become available for all tracks. Now, the aim is to associate the predicted track positions {\left\{{\widehat{\mathbf{p}}}_{kk1}^{\left(t\right)}\right\}}_{t} in (30) with the target position estimates {\left\{{\widehat{\mathbf{P}}}_{k}^{\left(m\right)}\right\}}_{m} in (29). To this end, define the decision variables a (t,m) ∈ {0,1} for t = 1,…,M_{ t } and m = 1,…,M, where a (t,m) = 1 amounts to deciding that target m measured at {\widehat{\mathbf{p}}}_{k}^{\left(m\right)} is assigned to track t. The pairwiseassociation cost can be quantified using the Mahalanobis distance between track t’s prediction and {\widehat{\mathbf{P}}}_{k}^{\left(m\right)} as a measurement, that is,
The following optimization problem is formulated to minimize the total association cost subject to linear constraints that ensure onetoone tracktomeasurement mapping:
It is worth mentioning that (33) is a special case of the socalled assignment problem, which is a wellknown data association algorithm [2], pp. 342–349. Its solution can be efficiently computed in polynomial time using integer programming solvers such as the Hungarian algorithm [34].
The track association problem in (33) can be modified to handle track birth and death scenarios [2]. Toward this objective, introduce a dummy target m=0 and a dummy track t = 0. The onetoone constraints in (33) are modified as follows: each track is assigned to at most one target position measurement, but the dummy track can be associated with any number of targets; meanwhile, each position measurement is assigned to at most one track, but the dummy measurement can be assigned to multiple tracks. Further, the dummy target cannot be associated with the dummy track. Such a modified association problem resembles the auction algorithm [2, 27], along with the corresponding association costs defined in (32). The computational burden of this combinatorial problem can be reduced by removing some unlikely association pairs in advance. Essentially, if for a track t all the association costs {MD(t,m)}_{ m } exceed a large threshold, then this track is considered ‘dead’ and is associated with the dummy target. Similarly for a target m, if all the association costs {MD(t,m)}_{ t } are too large, then this target is considered ‘born’ and is associated with the dummy track.
Once the positiontotrack association is completed, velocity estimates can be obtained too. This is possible by subtracting target position at time k1 from its position at time k and dividing by the sampling period.
Finally, it is worth noting that in formulating (33), only the state transition probability matrix F_{ k } is needed, regardless of the underlying target movement model. It is possible however to utilize each target’s movement model to develop other (more effective) data association schemes and refine the track estimates as well. Such association and track refinement steps will take place after every TSSG update, using the output of the TSSG tracker to form the positionmeasurements (29) for the ensuing parallel target trackers, one for each target. The results will not be fed back to the TSSG trackers, thus ensuring resilience of TSSG estimates to data misassociation and track estimation errors.
7 Numerical tests
Consider a 300×300square meter surveillance region along with a 10×10 rectangular grid with equally spaced grid points. Therefore, each grid cell is of size 30×30. Simulations are performed for both single and multitarget scenarios.
7.1 Singletarget case
A single target starts at the southwest corner of the grid at time k=1 and moves northeast according to a constant velocity model
where \stackrel{\u0304}{\mathbf{v}} denotes the target’s constant velocity assumed known and given by \stackrel{\u0304}{\mathbf{v}}=(15,15)m/s, p_{k 1} is the previous target position, T_{ s }= 1 is the sampling time in seconds, and n_{ k } represents the modeling noise of zeromean and variance {\sigma}_{n}^{2}{\mathbf{I}}_{2}. Given this model and ignoring n_{ k }, if the target starts at the center of the grid cell it is currently in, then at the next time instant, it will arrive at the northeast corner of this grid cell conjoining the north, east, and northeast grid cells. Due to the symmetrically distributed noise, the target will have equal probability of falling inside each of the 4 grid cells. It is assumed that σ_{ n } is small enough so that the probability of a target moving into grid cells other than its four adjacent ones is negligible. The resultant movement model is as follows: a target stays on the current grid point with probability 1/4 and moves north, east, or northeast with probability 1/4. Whenever the target moves outside the boundaries of the surveillance region, tracking stops. One random realization of this movement model is plotted in Figure 2 and is considered for the ensuing simulations starting with the singletarget case. The target’s signal strength is s=10, and there are N=20 sensors distributed randomly over the surveillance region measuring the received TSS. The measurement noise v_{ k } is zeromean Gaussian white with unit variance. The propagation function h(x) in (1) is given by h (x) = c/(c + x^{2}) for x ≥ 0, where c is chosen so that h (60) = 0.5. Apparently, h (0) = 1 and h (x) is monotonically decreasing as x increases.
The proposed sparsityagnostic and sparsityaware TSSGKF trackers in Sections 3 and 4 are employed to estimate the target signal strengths and position vectors over time. The position estimation accuracy is measured by the average rootmeansquare error (RMSE) in the form of \text{RMSE}=\sqrt{\frac{1}{{K}_{\text{max}}}\sum _{k=1}^{{K}_{\text{max}}}\parallel {\widehat{\mathbf{p}}}_{k}{\mathbf{p}}_{k}{\parallel}_{2}^{2}}, where K_{max} is the tracking duration, and {\widehat{\mathbf{p}}}_{k} is obtained as in (6). The covariance matrix of the process noise w_{ k } is set to Q_{ k }= I_{ G } in (8), and 1,000 Monte Carlo runs over the random measurement noise are performed to compute the RMSE. Figure 3 depicts the RMSE performance with respect to the sparsitycontrolling coefficient λ_{ k } as a fraction of {\stackrel{\u0304}{\lambda}}_{k} in (13). The sparsityagnostic tracker corresponds to setting λ=0 in (11) and is also plotted for comparison. It is seen that the sparsityaware KF tracker outperforms the sparsityagnostic one for a large range of λ_{ k }≠ 0 values, and {\lambda}_{k}=0.1{\stackrel{\u0304}{\lambda}}_{k} appears to yield the lowest RMSE for this test. The optimal HMM filter exhibits the best performance but requires accurate knowledge of the target signal strength.
Figure 4 depicts the RMSE of the sparsityaware TSSGIEKF tracker of (18), with μ_{ k }= 1 and for different values of σ_{ k }. This tracker incorporates sparsity as an extra measurement and selects the sparsity model ρ(x_{ k }) as the ℓ_{1} norm function. Evidently, this extra measurement is effective in promoting sparsity, which leads to improved performance relative to the sparsityagnostic tracker. The noise variance {\sigma}_{k}^{2} of the sparsity measurement in (17b) is a design parameter chosen in accordance with the sensor measurements (here having unit variance). As Figure 4 indicates, there is an optimal value of σ_{ k } that attains the most effective tradeoff between the sensor measurements and the sparsityinduced measurement. As σ_{ k } becomes larger, the tracker collects less information from the extra measurement and eventually becomes sparsityagnostic when σ_{ k } is too large. On the other hand, when σ_{ k } is too small, the tracker is predominantly enforcing a sparse solution without considering much the sensor measurements, which also degrades tracking performance.
Both sparsityaware TSSG trackers, the TSSGKF tracker with {\lambda}_{k}=0.1{\stackrel{\u0304}{\lambda}}_{k} and the TSSGIEKF tracker with σ_{ k }= 2, are compared in Figure 5 in terms of their RMSE performance versus time. The curves are generated using 1,000 Monte Carlo runs. These two sparsityaware trackers exhibit similar performance, both outperforming the sparsityagnostic tracker. The clairvoyant optimal HMM filter is also tested as the benchmark.
Finally, Figure 6 demonstrates the dynamic behavior of the sparsityaware estimator in (11) with {\lambda}_{k}=0.9{\stackrel{\u0304}{\lambda}}_{k}. Even though the sparsityaware TSSGKF performs worse than sparsityagnostic TSSGKF for this value of λ_{ k }, it is chosen to demonstrate how sparsity affects the tracking process. The estimated TSSG state vectors are depicted over time, with a circle representing a nonzero TSS at the corresponding grid point. The true and estimated tracks are plotted as well. For clarity, only the projection of the target track on the ydirection is depicted. It is seen that the ‘cloud’ of nonzero target signal strengths follows the true track. The estimated target profile is seen to be indeed spatially sparse. The size of the nonzero support indicates the uncertainty in target position estimates, which apparently does not grow over time, even when using a simple gridinduced linear KF tracker to follow the state transition pattern.
7.2 Multitarget case
Two targets are respectively located at the southcenter and westcenter of the grid at time k = 1. They start moving according to the same movement model used for the singletarget case. Figure 7 plots one random realization of these target trajectories used for the ensuing multitarget test cases. Adhering to Assumption 1, these two trajectories do not overlap on the same grid point at the same time.
The target signal strengths are set to be s^{(1)}= s^{(2)}= 10. It is assumed that the trackers know the number of targets unless otherwise stated. There are 100 sensors deployed randomly over the surveillance region to measure the total received signal strengths.
First, the position estimation method presented in Subsection 6.1 is tested. Figure 7 depicts the position estimates as circles along with the true target trajectories, for both the sparsityagnostic TSSGKF and the sparsityaware TSSGKF trackers with {\lambda}_{k}=0.1{\stackrel{\u0304}{\lambda}}_{k}. When the ℓ_{1}norm sparsitypromoting regularization term is not present (cf. Figure 7), position estimates are rather inaccurate and some of them fall far from either of the two targets. In contrast, the sparsityaware TSSGKF in Figure 7 results in quite accurate position estimates. One can clearly associate each position estimate with one of the two targets and readily visualizes target tracks from the position estimates. Before the position estimates are associated with individual targets, a pertinent performance metric quantifying estimation accuracy is the socalled Wasserstein distance (WD) that measures the distance between two finite sets [35]. Let {P}_{k}={\left\{{\mathbf{p}}_{k}^{\left(m\right)}\right\}}_{m} denote the finite set of the true target positions at time k and {\widehat{P}}_{k}={\left\{{\widehat{\mathbf{p}}}_{k}^{\left(n\right)}\right\}}_{n} the set of position estimates, respectively. Let d(.,.) stand for the Euclidean ℓ_{2} norm, and • for set cardinality. The L^{p} WD between these two sets is defined as
Figure 8 depicts the L^{1} WD for both sparsityaware TSSGKF and TSSGIEKF trackers, in comparison with the sparsityagnostic TSSGKF tracker. The TSSGIEKF tracker is implemented with μ_{ k }= 2 and σ_{ k }= 2. The WD is evaluated by averaging over 1,000 Monte Carlo runs for each tracker. Evidently, both sparsityaware designs are effective and improve the WD performance.
The track formation algorithm of Subsection 6.2 is investigated next for the same target realization. The target tracks formed using the position estimates of a single Monte Carlo run are plotted in Figure 9, for the sparsityagnostic TSSGKF tracker. The estimated track for target 1 is not even plotted because it deviates too much from the true trajectory. The estimated track for target 2 shows some erratic behavior. As will be discussed shortly, the unsatisfactory performance is not due to the proposed track formation algorithm itself; rather, it is a manifestation of inaccurate clustering that results from badly shaped TSSG estimates to begin with. The accuracy of the TSS map provided by the TSSG filters is essential in ensuring good performance of position estimates and track formation algorithms. Figure 9 illustrates the track estimates obtained after processing the sparsityaware TSSGKF output. It can be seen that both targets are closely tracked. To compare these methods quantitatively, the RMSE curves for the two targets are plotted versus time in Figure 10, for 1,000 Monte Carlo runs. It is evident that exploitation of sparsity markedly improves performance of the TSSG filters. In addition, sparsityaware TSSGKF seems to outperform the TSSGIEKF for this specific setting and choice of parameters.
To further illustrate the importance of TSSG estimation for subsequently forming position and track estimates, Figure 11 depicts two snapshots of the TSSG heat maps after the KF prediction and correction steps at times k=2 and 3. For the sparsityagnostic TSSGKF tracker, the correction heat map at k=2 seems to contain three clusters while there are only two targets. In the correction heat map at k=3, there is a single point in the lower right which is nonzero and far from both targets. This spurious point can have a detrimental effect during the clustering phase as it can greatly shift mean positions of the two clusters. These malign effects do not show up in the TSSG heat maps for the sparsityaware TSSGKF in Figure 11, where heat maps exhibit two compact clusters in both KF correction steps.
A comparison with the ALM filter proposed by [15] is performed next. Like PHD, ALM requires evaluating an integral which does not have a closed form. While [15] uses a random grid (or PF) to evaluate this integral, we use a deterministic grid instead. The performance metric is the WD between position estimates of TSSG/ALM and the true target positions. To obtain position estimates from intensities in TSSG/ALM, we use a common clustering scheme. Firstly, the TSSG and ALM intensity are converted into distributions by appropriate scaling. Then, 100 points are sampled from these distributions. Finally, kmeans clustering is performed on the sampled points, and the corresponding clusterheads are set as target position estimates. The same twotarget model considered in the previous simulations is considered. The true number of targets is assumed known during the clustering phase, but they are not needed during the TSSG/ALM intensity evaluations. Results are plotted in Figure 12, where TSSGKF is shown to outperform ALM by a comfortable margin.
Lastly, simulations for an unknown number of targets are performed on a 15×15 grid with the true and estimated target tracks plotted in Figure 13. In this setup, targets 1 and 2 begin their movement at time k=1; at k=5 target 3 is born, and at k=10 target 1 disappears. The sparsityaware TSSGKF is utilized in both simulations. Various clustering options are available when the number of clusters is unknown [32]. Here, a simple MATLAB routine called ‘silhouette’ is used to determine the best number of natural clusters in the TSS maps. After kmeans clustering is performed, silhouette returns a value between 1 and 1 for every point that has participated in the clustering phase. The value that silhouette returns measures how well every point is explained by the cluster it belongs to, compared to other clusters. A value close to 1 is desirable. Therefore, silhouette values averaged over the clustered points offer a good measure of how well clusters explain the points which belong to them. The number of clusters with the largest average silhouette value is selected as the most appropriate number of clusters. It can be seen that the three targets are accurately tracked. However, a small erroneous track emerges close to target 1 for two time periods. Unfortunately, performance of the case with unknown number of targets is not always as accurate as shown here and more than one inaccurate track may arise. On the other hand, when applied to the twotarget example previously considered in the absence of target births or deaths, the algorithm with unknown number of targets is always successful in recovering accurate target tracks.
8 Conclusions
The problem of tracking multiple targets on a plane using the superposition of their received signal strengths as measurements has been investigated. A gridbased state space model was introduced to describe the dynamic behavior of target signal strengths. This model not only renders the nonlinear estimation problem linear, but also facilitates incorporation and exploitation of the gridinduced sparsity present. Two sparsityaware Kalman trackers were developed to exploit this sparsity attribute: TSSGKF promoting sparsity of the state estimates through ℓ_{1}norm minimization and TSSGIEKF effecting sparsity by viewing it as an extra measurement. To address the challenge of updating the state estimation error covariances under sparsity constraints, a novel approach based on iterative extended KF and measurement augmentation was also developed to provide tractable and accurate covariance updates. Position estimation and positiontotrack association issues were considered as well. The proposed trackers do not require knowing the number of targets or their signal strengths and considerably reduce complexity when compared to the optimal hidden Markov model filter. They offer improved tracking performance at reduced sensing and computational cost, especially when compared to sparsityagnostic trackers.
Endnotes
^{a} A more accurate covariance update will be derived in (28).
^{b} It holds trivially for the dummy target m = 0 as well, because p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0j\in {\mathcal{G}}_{k}^{\left(0\right)})=0 and p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0,j\in {\mathcal{G}}_{k}^{\left(0\right)})=0.
Appendix
State transition model
From the total probability argument, it holds that
which leads to the following equality after invoking Assumption 2 in Bayes’ rule^{b}:
Any grid point j = 1,…,G with a nonzero {x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0 is associated with a single target index {m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\in \phantom{\rule{0.3em}{0ex}}[\phantom{\rule{0.3em}{0ex}}1,M] at time k, which means p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0,j\in {\mathcal{G}}_{k}^{\left({m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\right)})\ne 0 for {m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\in \phantom{\rule{0.3em}{0ex}}[\phantom{\rule{0.3em}{0ex}}1,M], and according to assumption 1, p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0,j\in {\mathcal{G}}_{k}^{\left(m\right)})=0, \forall m\ne {m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)} or m={m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}=0. Invoking p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0)=\sum _{m=0}^{M}p({x}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\ne 0,j\in {\mathcal{G}}_{k}^{\left(m\right)}) and noting that p(j\in {\mathcal{G}}_{k}^{\left({m}_{k}^{\left(\phantom{\rule{0.3em}{0ex}}j\right)}\right)})=1 yields
Similarly for a grid point i at time (k1), there exists a target index {m}_{k1}^{\left(i\right)}\in \phantom{\rule{0.3em}{0ex}}[\phantom{\rule{0.3em}{0ex}}0,M] such that p({x}_{k1}^{\left(i\right)}\ne 0)=p({x}_{k1}^{\left(i\right)}\ne 0,i\in {\mathcal{G}}_{k1}^{\left({m}_{k1}^{\left(i\right)}\right)}), and p({x}_{k1}^{\left(i\right)}\ne 0,i\notin {\mathcal{G}}_{k1}^{\left({m}_{k1}^{\left(i\right)}\right)})=0, ∀i∈ [ 1,G]. Under Assumptions 1 and 2, (36) follows from (34) and (35) which completes the proof.
References
BarShalom Y, Li XR, Kirubarajan T: Estimation with Applications to Tracking and Navigation. Wiley, Hoboken; 2001.
Blackman S, Popoli R: Design and Analysis of Modern Tracking Systems. Artech House, Boston; 1999.
BarShalom Y, Daum F, Huang J: The probabilistic data association filter: Estimation in the presence of measurement origin uncertainty. IEEE Control Syst. Mag 2009, 29(6):82.
Mahler R: Multitarget Bayes filtering via firstorder multitarget moments. IEEE Trans. Aerospace Electron Syst 2003, 39(4):1152. 10.1109/TAES.2003.1261119
Vo BN, Ma WK: The Gaussian mixture probability hypothesis density filter. IEEE Trans. Signal Process 2006, 54(11):4091.
Coue C, Pradalier C, Laugier C, Fraichard T, Bessiere P: Bayesian occupancy filtering for multitarget tracking: An automative application. Intl. J. Robot. Res 2006, 25(1):19. 10.1177/0278364906061158
Bazerque JA, Giannakis GB: Distributed spectrum sensing for cognitive radio networks by exploiting sparsity. IEEE Trans. Signal Process 2010, 58(3):1847.
Cevher V, Duarte MF, Baraniuk RG: Distributed target localization via spatial sparsity. Paper presented in the European signal processing conference. Switzerland, Laussanne; 2008.
Ling Q, Wu G, Jiang C, Tian Z: Joint multisource localization and environment perception in wireless sensor networks. Paper presented in Chinese control and decision conference. Xuzhou, China; May 2010.
Angelosante D, Bazerque JA, Giannakis GB: Online adaptive estimation of sparse signals: Where RLS meets the ℓ_{1}norm. IEEE Trans. on Signal Processing 2010, 58(7):3436.
Angelosante D, Roumeliotis SI, Giannakis GB: LassoKalman smoother for tracking sparse signals. Asilomar Conference on signals, systems, and computers. Pacific Grove, CA, USA; 2009.
Vaswani N: Kalman filtered compressed sensing. Paper presented in the international conference on image processing. San Diego, CA, USA; 2008.
Mahler R: CPHD filters for superpositional sensors. In Signal and Data Processing of Small Targets. Edited by: Drummond RD, Teichgraeber OE. SPIE, Bellingham; 2009.
Thouin F, Nannuru S, Coates M: Multitarget tracking for measurement models with additive contributions. In Proceedings of the International Conference on Information Fusion. Chicago, IL; July 2011 (see also the corrected version on ) http://www.ece.mcgill.ca/~mcoate/publications/publications.html July 2011 (see also the corrected version on )
Nannuru S, Coates M, Mahler R: Computationallytractable approximate PHD and CPHD filters for superpositional sensors. IEEE J. Select. Topics Signal Process 2013, 7(3):410.
Farahmand S, Giannakis GB, Leus G, Tian Z: Sparsityaware Kalman tracking of target signal strengths on a grid. Paper presented in the international conference on information fusion. Chicago, IL, USA; 2011.
Erdinc O, Willett P, BarShalom Y: The binoccupancy filter and its connection to the PHD filters. IEEE Trans. Signal Process 2009, 57(11):4232.
Tanaka M, Katayama T: Robust fixedlag smoother for linear systems including outliers in the system and observation noises. Intl. J. Syst. Sci 1988, 19(11):2243. 10.1080/00207728808964116
Candes EJ, Tao T: Decoding by linear programming. IEEE Trans. Inform. Theory 2005, 51(12):4203. 10.1109/TIT.2005.858979
Hastie T, Tibshirani R, Friedman J: The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York; 2009.
Farahmand S, Angelosante D, Giannakis GB: Doubly robust smoothing of dynamical processes via outlier sparsity constraints. IEEE Trans. Signal Process 2011, 59(10):4529.
Department of Industrial and Systems Engineering, Lehigh University. Sedumi . Accessed 9 Jan 2014 http://sedumi.ie.lehigh.edu/
Bertsekas DP, Tsitsiklis JN: Parallel and Distributed Computation: Numerical Methods. Athena Scientific, Nashua; 1999.
Bertsekas DP: Incremental leastsquares methods and the extended Kalman filter. SIAM J. Optimization 1996, 6(3):807. 10.1137/S1052623494268522
Taylor G, Kleeman L: Visual Perception and Robotic Manipulation. Springer, Berlin; 2006.
Clark DE, Bell J: Data association for PHD filter. Paper presented in the international conference on intelligent sensor, sensor networks and information processing. Melbourne, Australia; 2005.
Lin L, BarShalom Y, Kirubarajan T: Track labeling and PHD filter for multitarget tracking. IEEE Trans. Aerospace and Electronic Syst 2006, 42(3):778.
Panta K, Vo BN, Singh S: Novel data association techniques for the probability hypothesis density filter. IEEE Trans. Aerospace Electron. Syst 2007, 43(2):556.
Panta K, Clark DE, Vo BN: Data association and track management for the Gaussian mixture probability hypothesis density filter. IEEE Trans. Aerospace Electron Syst 2009, 45(3):1003.
Mekhnacha K, Mao Y, Raulo D, Laugier C: The “fast clusteringtracking” algorithm in the Bayesian occupancy filter framework. Paper presented in international conference on multisensor fusion and integration for intelligent systems. Seoul, Korea; 2008.
Bishop CM: Pattern Recognition and Machine Learning. Springer, New York; 2006.
Xu R, Wunsch II D: Survey of clustering algorithms. IEEE Trans. Neural Netw 2005, 16(3):645. 10.1109/TNN.2005.845141
Jain AK, Duin RPW, Mao J: Statistical pattern recognition: A review. IEEE Trans. Pattern Anal. Mach. Intell 2000, 22(1):4. 10.1109/34.824819
Kuhn HW: The Hungarian method for the assignment problem. In 50 Years of Integer Programming 1958–2008. Edited by: Junger M. Springer, Berlin; 2010:2947.
Clark DE, Panta K, Vo BN: The GMPHD filter multiple target tracker. Paper presented in the international conference on information fusion. Florence, Italy; 2006.
Acknowledgements
The first two authors were supported by the NSF grants CCF0830480, CCF1016605, ECCS0824007, and ECCS1002180. Geert Leus was supported in part by NWOSTW under the VICI program (project 10382). Zhi Tian was partly supported by the NSF grant ECS0925881. Part of the results in this paper has appeared in Proceedings of the International Conference on Information Fusion, Chicago, IL, USA, July 2011. The first two authors wish to thank Dr. D. Angelosante of ABB, Switzerland, and Prof. Stergios I. Roumeliotis of the CS&E Department at the University of Minnesota for helpful discussions and suggestions on the early stages of this work.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 2.0 International License ( https://creativecommons.org/licenses/by/2.0 ), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
About this article
Cite this article
Farahmand, S., Giannakis, G.B., Leus, G. et al. Tracking target signal strengths on a grid using sparsity. EURASIP J. Adv. Signal Process. 2014, 7 (2014). https://doi.org/10.1186/1687618020147
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/1687618020147
Keywords
 Kalman Filter
 Target Position
 Data Association
 Probability Hypothesis Density
 Extra Measurement