Cooperative Tracking of Vessel Trajectory by Multi-Static Passive Stations Using an MC-RMPF Authors: Bingzhuo Liu, Lingqi Kong, Panlong Wu Published: 2026 ABSTRACT Highlights What are the main findings? Dynamic station subsets improve localization efficiency while keeping accuracy.RMPF with vessel-motion constraints enhances tracking accuracy and continuity. What are the implications of the main findings? Geometry-aware station cooperation enables scalable passive maritime surveillance.Physics-informed particle filtering improves robust non-cooperative target tracking. AbstractTraditional maritime vessel tracking methods based on multi-static passive radar stations typically process all available observations, leading to substantial computational overhead and estimation variance. Furthermore, discrepancies in refresh rates and noise levels among stations often cause significant jumps in estimated positions between updates, resulting in trajectory discontinuities. To mitigate these issues, this paper introduces a multi-station cooperative vessel tracking framework based on a motion-constrained resample–move particle filter (MC-RMPF). In the proposed method, systematic resampling is first used to alleviate particle degeneracy, and a markov chain monte carlo (MCMC) move step is subsequently applied to rejuvenate the resampled particles under vessel-motion feasibility constraints. Additionally, a distributed detection network is constructed using directional data from multiple stations, dynamically selecting optimal observation subsets to balance localization accuracy with computational load. The experimental results demonstrate that, compared to the baseline methods, our method reduces the Root Mean Square Error and Circular Error Probability of position tracking by 23.5% and 21.7%, respectively. It exhibits strong reliability in challenging scenarios such as target maneuvers and temporary observation loss. 1. Introduction Maritime domain awareness is critical for a wide range of applications, including maritime traffic management, naval defense, illegal fishing monitoring, and search and rescue operations [1]. The cornerstone of this awareness is accurate and continuous vessel trajectory tracking. While the Automatic Identification System (AIS) is widely used, it is susceptible to intentional disabling and spoofing, or may simply be unavailable [2]. Therefore, non-cooperative tracking techniques that do not rely on vessels’ active transmission have become an indispensable complement. The traditional radar systems utilized for this task are predominantly active, emitting electromagnetic waves and detecting the reflected signals. However, active radar systems suffer from several drawbacks: they are energy-intensive and susceptible to Anti-Radiation Missiles, and their emissions can be easily detected, compromising operational stealth [3]. In contrast, passive radar systems exploit existing illuminators of opportunity (e.g., FM radio, digital TV, and cellular signals) or, in the specific context of vessel tracking, the signals emitted by the vessels’ own radar systems [4]. This passive nature grants them significant advantages in terms of low cost, covert operation, and immunity to electronic countermeasures, making them an increasingly attractive solution for persistent surveillance [5]. However, passive trajectory tracking techniques still face many challenges. Radar measurements are highly susceptible to multipath propagation and atmospheric refraction in marine environments, leading to significant angular errors [6]. Vessel motion often exhibits irregular and dynamic trajectories that can hardly be captured by entirely statistically based filters [7]. Kinematic maneuvers like turns and speed changes lead to a progressive accumulation of vessel position errors [8]. Many existing studies employ predetermined, static configurations and neither optimize the station geometry nor leverage the potential of multi-station cooperation [9]. The fundamental principle of passive tracking relies on angle-of-arrival (AOA) [10], time-difference-of-arrival (TDOA) [11], or frequency-difference-of-arrival (FDOA) [12]. Although long-range maritime localization remains a key topic, short-range indoor tracking has recently gained attention, within which AOA-enabled millimeter-wave sensors have emerged due to their high positioning accuracy. However, noise introduced by low-cost hardware significantly degrades AOA estimation precision. To mitigate this issue, Ref. [10] proposed a data-driven deep learning regression model that effectively reduces AOA estimation errors. A similar challenge arises in medium- and long-range passive tracking. In this context, Ref. [13] addressed the impact of sensor position uncertainty on estimation performance and proposed a bias-compensated method to overcome the limitations of pseudo-linear estimators. Numerous studies have also focused on TDOA-based passive localization methods, in which the intersections of several hyperbolic surfaces correspond to the emitter’s position. Considering systems with more than three measurement nodes, Ref. [11] performed combinatorial analysis and compared a Cramér–Rao Lower Bound (CRLB)-based weighted averaging approach with the traditional Weighted Least Squares (WLS) method, demonstrating the superior positioning performance of the former. Nevertheless, this advantage diminishes when the number of TDOA nodes is limited. Recent research has also explored hybrid AOA–TDOA measurements [14], in which the intersection of hyperbolic surfaces and bearing lines determines the source position. Since these intersections are inherently nonlinear and may yield multiple solutions, the Maximum Likelihood Estimation (MLE) is commonly employed. However, MLE performance strongly depends on the accuracy of the initial estimate during the tracking process. To alleviate this dependency, Ref. [15] introduced a Weighted Pseudo-Linear Estimator with Fusion (WPLE-F), which constructs an Instrumental Variable matrix to provide reliable initialization, thereby reducing sensitivity to initial conditions and achieving error performance close to the CRLB. Early research mainly focused on bistatic or two-station configurations, which can resolve a target’s position through triangulation [16]. While simpler to implement, these systems are often limited by geometric dilution of precision (GDOP) and provide poor coverage and robustness, especially in complex coastal environments [17]. The concept of GDOP has been incorporated into inertial navigation and synthetic aperture radar systems to analyze the influence of external factors on navigation accuracy [18]. Mathematically, GDOP is derived from the inverse of the Fisher Information Matrix (FIM), and the D-optimality criterion optimizes the observation geometry—such as sensor placement—by maximizing the determinant of the FIM. Building upon this framework, Ref. [19] employed the FIM to propose a localization accuracy metric for optimal sensor deployment. Nevertheless, these approaches remain limited in long-range, non-line-of-sight scenarios that require multi-static cooperation. The natural evolution is towards multi-station cooperative networks [20]. By fusing information from multi-static distributed passive radars, such networks can significantly improve localization accuracy, extend coverage, and enhance system reliability through redundancy [21]. The paradigm is shifting from simple data fusion to intelligent, cooperative tracking frameworks. State estimation is the core component of a tracking system. The journey began with the Kalman Filter (KF) and its nonlinear extensions, such as the Extended KF (EKF) and Unscented KF (UKF), which are efficient for Gaussian linear or mildly nonlinear problems [22]. For more complex, highly nonlinear, and non-Gaussian scenarios encountered in real-world tracking, Sequential Monte Carlo methods, particularly the Particle Filter (PF), have gained prominence [23]. The PF represents the posterior probability density function by a set of random particles, offering remarkable flexibility. However, standard PFs suffer from particle degeneracy and impoverishment [24]. A well-established remedy is the resample move strategy, in which importance resampling is followed by an MCMC move step to rejuvenate particles while preserving the target filtering distribution. This type of method combines importance sampling and MCMC sampling to mitigate progressive sample degeneration in dynamic Bayesian inference. It should be distinguished from MCMC-based particle filtering or sequential MCMC filtering methods, in which MCMC is used more fundamentally for sequential inference over the filtering distribution, often in high-dimensional multi-target tracking problems. For example, an MCMC-based particle filter was proposed for tracking a variable number of interacting targets, replacing the traditional importance sampling step with an MCMC sampling step [25]. Recently, a Rao–Blackwellised sequential MCMC sampler was developed for joint detection and tracking in clutter [26]. Despite these advances, significant challenges remain in the context of multi-station passive tracking:High computational cost: Most existing methods process all observation data from all stations. As the number of stations and targets grows, this “brute-force” fusion leads to prohibitive computational overhead, making it unsuitable for large-scale networks [27].Physical inconsistency: Many positioning algorithms merely rely on statistics and fail to incorporate domain knowledge such as velocity, tonnage inertia, and course heading. This often produces physically implausible trajectories with erratic jumps between updates, leading to discontinuous and unreliable tracks [28]. High computational cost: Most existing methods process all observation data from all stations. As the number of stations and targets grows, this “brute-force” fusion leads to prohibitive computational overhead, making it unsuitable for large-scale networks [27]. High computational cost: Most existing methods process all observation data from all stations. As the number of stations and targets grows, this “brute-force” fusion leads to prohibitive computational overhead, making it unsuitable for large-scale networks [27]. Physical inconsistency: Many positioning algorithms merely rely on statistics and fail to incorporate domain knowledge such as velocity, tonnage inertia, and course heading. This often produces physically implausible trajectories with erratic jumps between updates, leading to discontinuous and unreliable tracks [28]. Physical inconsistency: Many positioning algorithms merely rely on statistics and fail to incorporate domain knowledge such as velocity, tonnage inertia, and course heading. This often produces physically implausible trajectories with erratic jumps between updates, leading to discontinuous and unreliable tracks [28]. To bridge these critical gaps, this paper proposes a novel multi-station cooperative vessel tracking framework based on a physically constrained RMPF. Unlike traditional methods that treat data fusion and state estimation in isolation, our approach unifies geometric optimization with kinematic regularization to simultaneously enhance tracking efficiency and physical plausibility. The main contributions of this work are summarized as follows:A unified geometric modeling framework is established for multi-station AOA observations, in which the effects of station number and spatial configuration on positioning accuracy are rigorously analyzed within the GDOP framework, thereby providing a theoretical foundation for sensor layout optimization in passive tracking systems;A dynamic optimal subset selection strategy is introduced to obtain the most informative subset from the distributed stations, which can drastically reduce computational load while maintaining high-precision tracking.A physics-informed refinement mechanism is introduced, in which vessel kinematic constraints are embedded into the MCMC resampling procedure to suppress unrealistic state transitions and ensure physically consistent and dynamically reliable trajectory estimates. A unified geometric modeling framework is established for multi-station AOA observations, in which the effects of station number and spatial configuration on positioning accuracy are rigorously analyzed within the GDOP framework, thereby providing a theoretical foundation for sensor layout optimization in passive tracking systems; A unified geometric modeling framework is established for multi-station AOA observations, in which the effects of station number and spatial configuration on positioning accuracy are rigorously analyzed within the GDOP framework, thereby providing a theoretical foundation for sensor layout optimization in passive tracking systems; A dynamic optimal subset selection strategy is introduced to obtain the most informative subset from the distributed stations, which can drastically reduce computational load while maintaining high-precision tracking. A dynamic optimal subset selection strategy is introduced to obtain the most informative subset from the distributed stations, which can drastically reduce computational load while maintaining high-precision tracking. A physics-informed refinement mechanism is introduced, in which vessel kinematic constraints are embedded into the MCMC resampling procedure to suppress unrealistic state transitions and ensure physically consistent and dynamically reliable trajectory estimates. A physics-informed refinement mechanism is introduced, in which vessel kinematic constraints are embedded into the MCMC resampling procedure to suppress unrealistic state transitions and ensure physically consistent and dynamically reliable trajectory estimates. The remainder of this paper is structured as follows. Section 2 formulates the AOA positioning model and introduces the dynamic optimal station subset selection strategy. Section 3 presents the proposed physics-constrained RMPF framework, detailing the integration of vessel kinetic constraints. Section 4 describes the implementation of this framework for trajectory tracking. Section 5 details the experimental setup and presents a comparative performance evaluation against baseline methods across multiple station configurations. Finally, Section 6 concludes the paper. 2. Preliminaries 2.1. Signal Discrimination The surveillance network considered in this paper is a receive-only passive direction-finding system [5]. Each shore-based station intercepts vessel-emitted radio-frequency signals (e.g., navigation radar leakage or communication signals) and estimates the bearing of the source. No station transmits signals, and no reflected paths are utilized. Thus, the vessel acts as a non-cooperative emitter, and all stations operate as passive receivers, fundamentally differing from bistatic or multi-static radar systems. As illustrated in Figure 1, signal discrimination is the initial step in passive vessel tracking. Upon reception, raw AOA measurements are first calibrated to remove systematic biases. Subsequently, filtering and detection algorithms are applied to separate valid target signals from environmental noise. Specifically, detection algorithms evaluate signal strength, frequency characteristics, and temporal stability; this is followed by feature extraction and classification to identify candidate target signals. Adaptive thresholding and time–frequency analysis further refine signal selection at each station. A probabilistic data association method based on Mahalanobis distance [29] is then employed for cross-station matching, effectively resolving ambiguities by incorporating station geometry, measurement accuracy, and historical trajectory information. Finally, triangulation is performed to estimate target positions after successful signal association. Detailed signal identification methods are beyond the scope of this paper. 2.2. Position Tracking Principle As shown in Figure 1b, the multi-static passive system determines the position of a maritime target by fusing angular measurements from multiple stations, forming the basis of the proposed WLS model. Consider a 2D scenario where the target is located at (px,py) and M passive stations are positioned at (sxj,syj), j=1,2,…,M. Under ideal conditions, the bearing angle αj measured by the jth station satisfies [13](1)sinαj(px−sxj)−cosαj(py−syj)=0 In practice, angular measurements are subject to systematic and random errors, causing the measured value α^j to deviate from the true value αj by:(2)α^j=αj+Δαj where Δαj∼N(0,σj2). Rearranging (1) yields(3)sxjtanαj−syj=pxtanαj−py Substituting α^j and introducing noise ηj, we obtain(4)sxjtanα^j−syj=pxtanα^j−py+ηj where ηj is the equivalent linearized measurement error induced by the angular error Δαj. Its variance is not identical to σαj2, but is obtained through first-order error propagation and depends on the target station geometry. For all M stations, the following linear system can be established:(5)z=Hp+η where(6a)z=[sx1sinα^1−cosα^1sy1sx2sinα^2−cosα^2sy2⋮sxMsinα^M−cosα^MsyM],H=[sinα^1−cosα^1sinα^2−cosα^2⋮⋮sinα^M−cosα^M],(6b)p=[pxpy],η=[η1η2⋮ηM]. To account for the differing measurement accuracies across stations, a WLS scheme is adopted by constructing a diagonal weight matrix W as:(7)W=diag(w1,…,wM) where wj=1/σj2 is inversely proportional to the angular measurement accuracy of the corresponding station. The WLS estimate [15] of the target position is then(8)p^WLS=(HTWH)−1HTWz A unique solution exists when M≥2 and the stations are not collinear, ensuring HTWH is non-singular. 2.3. Dynamic Selection of Optimal Station Subset Although station locations are fixed, target–station geometry varies with vessel motion. A subset with good angular diversity at one epoch may become ill-conditioned elsewhere; thus, fixed offline selection is suboptimal. The proposed method dynamically updates the active subset based on the predicted target position to maintain favorable geometry. GDOP is fundamentally defined as [30]:(9)GDOP=σpx2+σpy2 where σpx and σpy are the standard deviations of positioning errors, respectively. From the AOA model, error propagation is obtained by differentiating Equation (1):dαj=∂αj∂pxdpx+∂αj∂pydpy+∂αj∂sxjdsxj+∂αj∂syjdsyj=cos2αjpx−sxjdpy−sinαjcosαjpx−sxjdpx(10)−cos2αjpx−sxjdsyj+sinαjcosαjpx−sxjdsxj. Then, the error propagation mechanism can be arranged as:(11)dα=G⋅dp+ΔS where dα=[dα1,dα2,…,dαM]T is the angle deviation vector, dp is the position error vector, G is the observation matrix the j-th row of which is [−sinαjcosαjpx−sxj,cos2αjpx−sxj], and ΔS means additional terms arising from station position uncertainties. Given the WLS weight matrix W, the position error covariance matrix can be derived as:(12)Cp=(GTWG)−1GTWΣαeqWG(GTWG)−1 where Σαeq is the equivalent angle-domain covariance matrix that incorporates both angular measurement noise and station position uncertainty. To ensure consistent units, station position errors are projected into the angle domain. From Equation (10), the partial derivatives of the AOA with respect to station coordinates are(13)∂αj∂sxj=py−syjrj2, ∂αj∂syj=−px−sxjrj2 where rj=(px−sxj)2+(py−syj)2 is the target-to-station distance. Let Js,j=[∂αj∂sxj,∂αj∂syj] denote the corresponding Jacobian with respect to the j-th station’s position. Assuming the station position error is isotropic with covariance Σs,j=σsj2I2,(14)σα,sj2=Js,jΣs,jJs,jT=(py−syjrj2)2σsj2+(−px−sxjrj2)2σsj2=σsj2rj2 Notably, σα,sj2 has units of m2/m2=rad2 (under the small-angle approximation), which is dimensionally compatible with the angular measurement variance σαj2 (also in rad2). Therefore, the total equivalent angular variance for the j-th station is the sum of these two independent contributions:(15)σαj,eq2=σαj2+σsj2rj2 Consequently, the equivalent angle covariance matrix can be written as:(16)Σαeq=diag(σα1,eq2,σα2,eq2,…,σαM,eq2) Ultimately, the GDOP for 2D positioning is obtained from the diagonal elements of Cp:(17)GDOP=σpx2+σpy2=Cp(1,1)+Cp(2,2)=trace(Cp) If only angular noise is considered, the GDOP can be simplified as:(18)GDOP=trace((HTWH)−1). For a system consisting of n observation stations, the optimal subset S* of size k is(19)S*=argminS⊂1,2,…,n∥S∥=kGDOPS. Exhaustive search is prohibitive; thus, an incremental greedy strategy is adopted. Under approximate submodularity, let f(S)=−GDOPS denote the utility function. For any subsets A⊂B and an element e∉B, the utility gain is assumed to satisfy the relaxed inequality:(20)f(A∪e)−f(A)≥f(B∪e)−f(B)−ϵ where ϵ≥0 is a relaxation parameter representing the deviation from strict submodularity. The algorithm proceeds as follows:(i)initialize S0 via exhaustive tri-station search;(ii)iteratively add the station maximizing GDOP reduction;(iii)terminate when |S|=k or improvement <ε (i)initialize S0 via exhaustive tri-station search; initialize S0 via exhaustive tri-station search; (ii)iteratively add the station maximizing GDOP reduction; iteratively add the station maximizing GDOP reduction; (iii)terminate when |S|=k or improvement <ε terminate when |S|=k or improvement <ε Since the GDOP reduction approximately exhibits a diminishing return behavior, the greedy strategy is used as a computationally efficient heuristic rather than as a strictly optimal solver. The relaxation parameter ϵ characterizes the deviation from exact submodularity; therefore, the greedy solution may deviate from the ideal approximation bound by a residual term that increases with ϵ. In this study, the greedy result is further verified by numerical GDOP comparison. To quantify the contribution of an individual station to positioning accuracy, a Station Impact Factor (SIF) is defined:(21)SIFi=GDOPS∖i−GDOPSGDOPS×100%. For any station i∈S, its SIF value satisfies(22)SIFi≈λi∑j∈Sλj×100% where λi is the contribution of station i to the Fisher information matrix. Therefore, SIF reflects the importance of a certain station to the overall geometric configuration. To address practical complexities, an adaptive weighting mechanism is proposed as:(23)wi=α⋅wrel,i+β⋅wacc,i+γ⋅wgeo,i where wrel,i, wacc,i, and wgeo,i represent station reliability, measurement accuracy, and geometric contribution, respectively, and α, β, γ are weight coefficients satisfying α+β+γ=1. Then, the modified GDOP calculation formula is(24)GDOPw=tr((GTW−1G)−1) where W=diag(w1−1,w2−1,...,wn−1) is the weight matrix. For a system with n observation stations, complexity analysis shows that: (1) the initial selection phase requires evaluating GDOP values for all three-station combinations, with complexity O((n3)); (2) the greedy iteration phase involves at most (k−3) iterations, each computing GDOP for (n−|S|) candidates, yielding O((k−3)(n−k+22)); and (3) the dynamic update phase recalculates GDOP for k stations after substitution, with complexity O(k(n−k)). 3. Improved RMPF Considering Physical Constraints Regarding Vessel Motion 3.1. PF and Systematic Resampling In PF [23], the posterior density is approximated by weighted particles {xki,wki}i=1N, where N is the number of sampling particles, xki is the state sample and wki its weight:(25)p(xk|z1:k)≈∑i=1Nw∼kiδ(xk−xki) where ∑i=1Nw∼ki=1, normalized as w∼ki; δ(⋅) is the Dirac function. The particles are typically drawn from a selected importance density function q(xk|xk−1i,zk), with weights:(26)wki∝wk−1ip(zk|xki)p(xki|xk−1i)q(xki|xk−1i,zk). PFs suffer from particle degeneracy, in which weight variance increases over time. This is mitigated using systematic resampling, triggered by the effective sample size [24].(27)Neff,k=1∑i=1N(w∼k(i))2 where w∼k(i) is the normalized particle weight, and the range of Neff,k is [1,N]. When Neff,k0 controlling constraint strictness. Unlike a hard indicator 𝟙1F, the soft specification assigns low but nonzero probability to near-infeasible states, preserving filter adaptability under extreme maritime conditions while maintaining physical consistency. The parameters in the motion feasibility factor are not treated as universal constants. Instead, they are specified according to the vessel class, navigation scenario, and available prior information. In practical deployment, the limits vmax, amax, and rmin can be obtained from vessel type-dependent maneuvering characteristics and historical AIS statistics. The penalty weights wv, wa, and wr determine the relative importance of speed, acceleration, and turning-radius violations, respectively, while λ controls the overall softness of the constraint. In this study, these parameters were fixed before testing and were selected using independent validation trajectories rather than the evaluation trajectories. Therefore, the proposed soft constraint should be understood as a scenario dependent feasibility model rather than a universal vessel motion law. Given the target πk in Equation (33), the Metropolis–Hastings acceptance ratio is(38)α(xk(i)→xk′)=min{1,πk(xk′∣xk−1(ak(i)),z1:k)q(xk(i)∣xk′)πk(xk(i)∣xk−1(ak(i)),z1:k)q(xk′∣xk(i))} Substituting Equation (34) into the above expression gives(39)α(xk(i)→xk′)=min{1,p(zk∣xk′)p(xk′∣xk−1(ak(i)))ϕ(xk′,xk−1(ak(i)))q(xk(i)∣xk′)p(zk∣xk(i))p(xk(i)∣xk−1(ak(i)))ϕ(xk(i),xk−1(ak(i)))q(xk′∣xk(i))} For the symmetric Gaussian random-walk proposal used in this work, namely,(40)q(xk′|xk(i))=q(xk(i)|xk′) the proposal–density ratio cancels, and the acceptance probability reduces to(41)α(xk(i)→xk′)=min{1,p(zk∣xk′)p(xk′∣xk−1(ak(i)))ϕ(xk′,xk−1(ak(i)))p(zk∣xk(i))p(xk(i)∣xk−1(ak(i)))ϕ(xk(i),xk−1(ak(i)))} Thus, kinematically infeasible candidates are not rejected by an additional constrained proposal. Instead, they receive a lower acceptance probability through the feasibility ratio. This avoids the introduction of an intractable proposal normalizing constant and preserves a clear Metropolis–Hastings construction with respect to the constraint-modified target distribution. At each MCMC step, the candidate xk′ is accepted with probability α; otherwise, the current particle xk(i) is retained. 3.3.3. Invariance and Convergence The MH transition kernel K defined by the proposal q and the acceptance probability α satisfies detailed balance with respect to πk:(42)πk(xk)q(xk′∣xk)α(xk→xk′)=πk(xk′)q(xk∣xk′)α(xk′→xk) which is verified by direct substitution of Equations (33), (35) and (37). Detailed balance implies πk-invariance, and irreducibility and aperiodicity follow from the strictly positive support of the Gaussian base proposal q, combined with ϕ>0 almost everywhere on the feasible region. Therefore, the constrained MCMC kernel converges to the constraint-modified conditional posterior πk, and the overall algorithm converges to the constraint-modified filtering posterior p∼ defined in Equation (34). This integration yields a physically informed Bayesian framework that preserves probabilistic consistency while incorporating motion constraints, improving robustness and preventing implausible state transitions. 4. Implementing the MC-RMPF for Vessel Trajectory Tracking The implementation of the physically constrained MC-RMPF for vessel tracking is illustrated in Figure 2. By integrating MCMC sampling with motion constraints, the method mitigates particle degeneracy and impoverishment, improving robustness and accuracy. The recursive procedure is as follows. 4.1. Initialization If the prior distribution of the vessel target’s state at time k=0 is obtained as p(x0), sample N particles from it:(43)x0(i)∼p(x0),i=1,2,…,N Initial weights are uniformly assigned as:(44)w0(i)=1N,i=1,2,…,N Subsequently, the filtering process is executed recursively through the following “prediction—update” steps. 4.2. Prediction Step and Update Step Prediction Step (Time Update) Prediction Step (Time Update) Given the posterior particle set {xk−1(i),wk−1(i)}i=1N at time k−1, the filtering proceeds recursively.(45)xki=fk−1(xk−1(i))+vk−1(i) where i=1,2,…,N, fk−1(⋅) is the system state transition function; vk−1(i) is the process noise, typically assumed to be Gaussian white noise vk−1(i)∼N(0,Qk−1). The weights remain unchanged in this step, resulting in the predicted particle set {xk(i),wk−1(i)}. Update Step (Measurement Update) Update Step (Measurement Update) Given measurement zk, the likelihood is(46)p(zk|xk(i))=1(2π)nz|Rk|exp(−12[zk−hk(xk(i))]TRk−1[zk−hk(xk(i))]) where hk(⋅) denotes the measurement equation, Rk is the measurement noise covariance matrix, and nz is the dimension of the measurement. Then, update the weights:(47)wki=wk−1i⋅p(zk|xki),i=1,2,…,N Normalize the weights:(48)wki=wki∑j=1Nwkj,i=1,2,…,N Thus, the posterior approximation {xki,wki} is obtained. 4.3. Degeneracy Detection and Systematic Resampling After weight normalization, particle degeneracy may occur, where a few particles dominate. This is quantified by the effective sample size:(49)Neff,k=1∑i=1N(w∼k(i))2. When Neff,k