{"id":"186d9f7a-fe1d-4cba-8113-f498b438e7f8","arxiv_id":"2504.15040","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The proposed GM-TPHD-E and GM-TCPHD-E filters produce closed-form recursions for multiple extended target tracking with explicit ellipse shape estimates and trajectory set outputs.","lead":"Two new filters combine trajectory set theory with PHD/CPHD multi-target tracking and a decoupled ellipse shape model, so each output track carries the target's orientation and axis lengths. They beat an older random-matrix filter and a label-based filter on shape and trajectory error in simulations and on real drone traffic data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Printed TCPHD-E cardinality recursion is invalid as written: Eqs. (96)-(98) miscompute PGF derivatives, so the claimed closed-form CPHD update is not implemented by the published formulas.","rationale":"The reader's REJECT verdict is well supported, but my most load-bearing concern is not exactly the reader's stated weakest assumption about the Yang shape linearization. The PGF derivative errors in Eqs. (96)-(98) are directly checkable, are internal to the paper's own Poisson and cardinality PGF definitions, and affect the core TCPHD-E update: the posterior cardinality distribution, which is the distinguishing feature of the CPHD variant, is computed with non-derivatives. A second instance of the same failure mode is visible in the TPHD-E weight update Eq. (55), which appears to multiply by rho+zeta where the pseudo-likelihood Eq. (14) requires division by rho+zeta. These are not ambiguous modeling choices; they are algebraic slips in the closed-form recursion that the paper claims as its main contribution. The core idea may be salvageable with corrections, and the paper does acknowledge one trajectory-covariance approximation around Eqs. (56)-(57), but as published the equations do not support the claimed closed-form Bayesian recursions. This reinforces the reader's REJECT without moving the verdict, because the reader already reached that conclusion.","tokens_in":27011,"tokens_out":16852,"duration_ms":163202,"concrete_test":"Run a one-step sanity check of the GM-TCPHD-E update with a known predicted cardinality distribution P_omega (e.g., P_omega = {0.1, 0.4, 0.4, 0.1} for n=0,...,3), one singleton measurement, lambda=10 and gamma=20. Evaluate Eq. (22) twice: once with Eqs. (96)-(98) as printed, and once with the standard derivatives G_FA^{(j)}(0)=lambda^j e^{-lambda}, G_z^{(j)}(0)=gamma^j e^{-gamma}, and G_{k|k-1}^{(|P|)}(upsilon)=sum_{m>=|P|} P_omega(m) m!/(m-|P|)! upsilon^{m-|P|}. Check whether each posterior cardinality distribution sums to 1 over n and compare the two resulting P_{k|k}(n). If the printed formulas do not preserve normalization or change the posterior materially, the published TCPHD-E recursion is not the CPHD Bayes update it claims to be.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B implements the TCPHD-E cardinality update using three incorrect PGF formulas. For Poisson clutter, G_FA(z)=e^{lambda(z-1)}, so the n-th derivative at 0 is lambda^n e^{-lambda}, not n! e^{-lambda} as printed in Eq. (96). Eq. (97) has the same defect for the target measurement PGF, omitting gamma^n. Eq. (98) evaluates G_{k|k-1}^{(|P|)}(upsilon) as a single cardinality term (|P|)! upsilon^{|P|} P_omega(|P|), whereas the true derivative at upsilon is a sum over all cardinalities m >= |P| of P_omega(m) m!/(m-|P|)! upsilon^{m-|P|}. These derivatives enter the posterior cardinality distribution through epsilon_{P,C}, mu_{P,C}, nu_{P,C} and Eq. (22), so the printed recursion is not the extended-target CPHD correction. This is an internal inconsistency with the paper's own Poisson PGF definitions, not a disagreement with a modeling choice, and it directly undercuts the central claim of a correct closed-form TCPHD-E recursion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes two multi-extended-target tracking filters, the TPHD-E and TCPHD-E, obtained by embedding Trajectory Set Theory (TST) into the extended-target PHD and CPHD recursions, and by combining these with a decoupled elliptical-shape estimation model borrowed from Yang's work. Gaussian mixture implementations (GM-TPHD-E and GM-TCPHD-E) are given, including closed-form recursions, pruning/merging, and state extraction. The methods are evaluated in two scenarios, one simulated and one based on the SIND traffic dataset, using the Gaussian Wasserstein distance and the trajectory metric. The central claim is that these filters provide explicit, accurate ellipse shape estimates and complete, stable trajectory generation, outperforming an RMM-based trajectory PHD filter and a label-based ET-PHD filter.","tokens_in":27335,"tokens_out":5987,"duration_ms":56475,"significance":"If the recursions were correct, the paper would offer a practically useful combination of track-oriented TST filtering with explicit shape estimation for extended targets. A strength of the paper is that it provides full Gaussian-mixture implementations and evaluates on both simulated and real-data scenarios, which goes beyond purely conceptual derivations. However, the published recursions contain several specific mathematical errors that affect the core update equations. The most serious are the PGF-derivative formulas in the GM-TCPHD-E update, which are internally inconsistent with the paper's own Poisson assumptions. Because these formulas feed directly into the posterior cardinality distribution, the printed TCPHD-E recursion does not implement the claimed extended-target CPHD correction. The experimental results therefore cannot be used as evidence for the algorithms as stated, although the errors appear to be correctable.","major_comments":[{"comment":"The PGF derivative formulas used in the GM-TCPHD-E cardinality update are incorrect. For the clutter PGF G_FA(z)=e^{λ(z−1)}, the n-th derivative at 0 is λ^n e^{−λ}, not n! e^{−λ} as stated in Eq. (96). Similarly, for a target measurement PGF with parameter γ, G_z^{(n)}(0)=γ^n e^{−γ}, not n! e^{−γ} as in Eq. (97). Eq. (98) evaluates G_{k|k−1}^{(|P|)}(υ) as a single term (|P|)! υ^{|P|} P_{ωk}(|P|), whereas the true derivative of the PGF is a sum over all cardinalities m≥|P| of P_{ωk}(m) m!/(m−|P|)! υ^{m−|P|}. These errors propagate through ε_{P,C}, μ_{P,C}, ν_{P,C} and Eq. (22), so the printed cardinality update is not the extended-target CPHD recursion. This is an internal inconsistency, not a modeling choice, and it directly undercuts the central claim of a correct closed-form TCPHD-E recursion.","section":"Section IV-B, Eqs. (96)-(98)"},{"comment":"The component weight update for the GM-TPHD-E filter appears to have the wrong dependence on the measurement-cell weight. According to the pseudo-likelihood in Eq. (14), the contribution of a measurement cell C is divided by (ϱ(P,C)+ς(P,C)); specifically, L_Z contains φ((P,C)|ξ)/(ϱ(P,C)+ς(P,C)). However, Eq. (55) multiplies the component weight by (ϱ(P,C)+ς(P,C)) instead of dividing by it. The resulting weight w^{(j,P,C)}_{πk} is therefore dimensionally and functionally inconsistent with the continuous-time recursion it is meant to implement. If this is a typographical error, it must be corrected; as printed, the GM-TPHD-E update does not match Eqs. (13)–(17).","section":"Section IV-A, Eq. (55)"},{"comment":"The GM-TCPHD-E cardinality prediction formula contains two errors. The sum over l is written as 'Σ_{l−j}^{∞}' rather than 'Σ_{l=j}^{∞}', and the argument of the posterior cardinality distribution inside the sum is P_{πk−1}(n) rather than P_{πk−1}(l). With these errors, the formula does not sum over the correct index and does not use the prior cardinality distribution correctly. The intended expression is presumably the constant-pS specialization of Eq. (20), which uses P_{πk−1}(l) and a sum from l=j to infinity. As printed, this formula would produce an incorrect predicted cardinality distribution and would invalidate the subsequent update.","section":"Section IV-B, Eq. (87)"},{"comment":"The trajectory covariance update in Eqs. (56)–(57) sets the posterior trajectory covariance to a block-diagonal matrix via the construction with I(n^{(j)}_{ωk}), discarding cross-time correlations. The text acknowledges this as a 'slight error' and compares it to a '1-Scan' implementation. However, the paper claims closed-form Bayesian recursions and complete trajectory generation. A block-diagonal covariance after update is an approximation, not an exact Bayesian update for the trajectory state, and the paper does not quantify the resulting error. This assumption should be stated as an approximation in the main derivation and its effect on trajectory accuracy should be assessed, particularly since the experiments report trajectory metrics.","section":"Section IV-A, Eqs. (56)-(58)"}],"minor_comments":[{"comment":"The phrase 'closed Bayesian recursive' in the abstract should be 'closed-form Bayesian recursions'; the same wording issue appears in the introduction and conclusion.","section":"Abstract and Introduction"},{"comment":"The terms 'Semi Positive Definite Matrix' should be 'positive semidefinite matrix'; the standard terminology is used inconsistently.","section":"Introduction, Section I"},{"comment":"The shape matrix S^{(j)}_{ωk} is defined with rotation and diagonal scaling but the associated measurement covariance computation in Eq. (64) would benefit from a brief justification of why the multiplicative noise term S Q_h S^T is inserted in the measurement covariance; the paper currently refers only to [40] without a self-contained explanation.","section":"Section IV-A, Eq. (65)"},{"comment":"The table header uses 'Qe = 1\n4\n˙Qe' and similar notation which is ambiguous; the intended meaning should be stated explicitly, e.g., 'Qe = (1/4) times the nominal value Qe'.","section":"Section V, Table IV"}],"recommendation":"major_revision","confidential_remarks":"The mathematical errors in Eqs. (55), (87), and especially (96)-(98) are substantive and affect the main algorithms. They are, however, local and correctable: replacing the PGF derivatives with the correct ones and fixing the weight update would yield the intended recursions. Given that the paper claims a rigorous closed-form derivation, the authors should be asked to correct these formulas, re-run the simulations with the corrected implementation, and report whether the qualitative conclusions still hold. The contribution is incremental in combining existing components (trajectory PHD/CPHD, extended-target PHD/CPHD, and Yang's shape model), but the combination is a reasonable engineering contribution if the formulas are fixed. I would not recommend rejection solely because the errors are fixable, but the current version is not publishable as the recursions stand."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper combines trajectory PHD/CPHD filtering with Yang's decoupled ellipse-shape model, which is a genuinely new combination and a plausible direction for extended-target tracking. But the printed recursions have load-bearing math errors, so the closed-form GM-TPHD-E and GM-TCPHD-E filters are not correctly specified as written.\n\nCredit where due: the specific combination isn't in the cited literature—trajectory PHD work so far uses random-matrix shapes, and Yang's model has been used with network-flow labeling rather than trajectory set theory. The derivation follows the standard extended-target PHD/CPHD machinery, and the GM implementation is a natural extension. The experiments are reasonably scoped: two scenarios, 100 MC runs, GWD and trajectory metric, plus real SIND data in Scenario 2. The paper also openly flags the block-diagonal trajectory covariance approximation in Eqs. (56)-(57), which is good practice.\n\nThe soft spots are real. Eq. (55) multiplies the GM-TPHD-E component weight by (rho+zeta), whereas the pseudo-likelihood in Eq. (14) divides by that term; as printed the update is mis-normalized. More seriously, the TCPHD-E cardinality update in Section IV-B uses incorrect PGF derivatives. For Poisson clutter, G_FA(z)=e^{lambda(z-1)}, so G_FA^(n)(0)=lambda^n e^{-lambda}, not n! e^{-lambda} as in Eq. (96). Eq. (97) omits gamma^n for the target-measurement PGF. Eq. (98) evaluates the predicted cardinality PGF derivative as a single term, but the n-th derivative at upsilon is a sum over all cardinalities m>=n of P_omega(m) m!/(m-n)! upsilon^{m-n}. These derivatives feed into the posterior cardinality through Eqs. (22)-(29), so the printed recursion is not the extended-target CPHD correction it claims to be. Eq. (87) also has an index inconsistency: P_{pi k-1}(n) appears inside a sum over l that should be P_{pi k-1}(l).\n\nAre these fatal? For the paper's central claim, yes as printed. The idea may be salvageable, but no code or corrected derivation is provided, so there is no way to verify that the implemented filters repair the formulas. The experiments also lack error bars, and the Scenario 1 birth model uses true initial positions with small noise—a mild bias, not a fatal flaw. The citation pattern is fine; I see no self-citation loop.\n\nWho is it for? Researchers in extended-target tracking who want to see TST and decoupled shape estimation combined. The paper deserves a serious referee because the combination is sensible and the errors are likely correctable. I would send it to review with the expectation of major revision, and I would not build on the formulas as printed.","headline":"A worthwhile combination of trajectory PHD/CPHD filtering with decoupled shape estimation, but the printed recursions contain load-bearing math errors that invalidate the closed-form claims as written.","tokens_in":27795,"tokens_out":5469,"would_cite":false,"duration_ms":45072,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper derives closed-form recursions that let PHD/CPHD filters output whole trajectories and explicit ellipse shapes.","keywords":["multi-target tracking","extended target tracking","unresolvable group targets","trajectory set theory","PHD filter","CPHD filter","elliptical extent estimation","Gaussian mixture implementation"],"falsifier":"Apply the GM-TCPHD-E filter to a single, highly elongated target executing a sharp turn with low detection probability, and compare the estimated $\\theta,l_1,l_2$ with ground truth over repeated Monte Carlo runs: if the pseudo-measurement linearization is inadequate, orientation error or Gaussian Wasserstein distance will fail to converge even when cardinality is correct. A second check compares the block-diagonal trajectory covariance against a full-scan implementation on two closely spaced, parallel-moving targets; if cross-time correlations matter, the block-diagonal version should show track fragments or elevated switch cost.","tokens_in":26796,"feed_emoji":"📡","tokens_out":7506,"duration_ms":64393,"temperature":0.7,"pith_summary":"This paper proposes two closed-form Bayesian filters, the TPHD-E and TCPHD-E filters, for tracking multiple extended or group targets whose sensor footprint is an ellipse. It takes the PHD and CPHD filters — random-finite-set trackers that propagate an intensity over targets and, for CPHD, a full cardinality distribution — and adds Trajectory Set Theory so each Gaussian component represents an entire target history. A decoupled shape state carries the ellipse's orientation and two semi-axis lengths, updated through a linearized pseudo-measurement, so shape estimates emerge explicitly rather than inside a random matrix. In simulations and a real traffic-intersection dataset, the paper reports better shape accuracy than a random-matrix baseline and more complete, stable trajectories than a label-based baseline, with the CPHD variant most accurate in target number.","feed_headline":"Two new filters track whole trajectories and ellipse shapes at once","feed_subtitle":"Closed-form recursions add ellipse orientation and axes to trajectory PHD/CPHD filters, beating both baselines","key_machinery":"The load-bearing mechanism is the trajectory Gaussian component $N(\\chi;t,\\dot{\\xi},\\dot{\\Xi})$, whose block-diagonal mean and covariance store the target state at every time step of the trajectory, propagated by Trajectory Set Theory. The decoupled shape update uses the multiplicative-noise measurement model of [40]: each measurement contributes a pseudo-measurement $Z=\\breve{F}\\big((z-\\bar{z})\\otimes(z-\\bar{z})\\big)$ whose expectation is a linear function of the shape parameters, keeping the ellipse orientation and semi-axis estimates inside the Gaussian-mixture recursion. Sequential updates and a merging scheme with separate kinematic and shape thresholds keep the mixture manageable. The paper notes that building the trajectory covariance by block-diagonal concatenation in Eqs. (56)-(57) is a slight approximation, since it ignores correlations between states at different times.","core_discovery":"The central claim is that trajectory PHD/CPHD filtering and a decoupled elliptical shape model can be merged into closed recursions for multiple extended targets. Kinematic state $r$ and shape state $s=[\\theta,l_1,l_2]^T$ are propagated independently, and shape information enters through the pseudo-measurement update of [40], so the posterior remains a Gaussian mixture over trajectories. The paper asserts that the resulting GM-TPHD-E and GM-TCPHD-E filters are strictly defined trackers that avoid trajectory switching in close parallel motion, converge to accurate ellipse orientation and axis lengths, and outperform the GM-TPHD-GIW filter in shape error and the GM-LPHD-E filter in trajectory metric, with GM-TCPHD-E the best overall.","pith_inferences":["The same decoupled shape update could be combined with measurement-driven birth models, such as Poisson multi-Bernoulli filters, which the authors list as future work, to remove the dependence on a hand-placed birth density.","Because the shape merging threshold is set smaller than the kinematic one, two targets with similar kinematics but distinct orientations will be kept separate; a testable corollary is that performance degrades when shapes are genuinely similar and close.","If the pseudo-measurement approximation holds, the same recursion should work for other parametric extents, such as rectangles or star-convex shapes, by swapping the shape model and its pseudo-measurement.","The claim that Trajectory Set Theory adds no component growth assumes the Gaussian-mixture representation; an extension to non-Gaussian or unknown measurement-rate settings would need to check whether trajectory dimension growth changes complexity."],"forward_implications":["With these recursions, an extended-target PHD/CPHD tracker can output complete trajectories, not just current states, without a separate labeling or network-flow track-construction stage.","The ellipse orientation and semi-axis lengths are estimated explicitly in closed form, which is what obstacle avoidance and formation-flight applications need.","The CPHD version, GM-TCPHD-E, should estimate the number of targets more accurately at births and deaths, and the paper's experiments show it is least affected by parameter changes.","The separate kinematic and shape merging thresholds let the filter keep shape-distinct components alive while pruning kinematically redundant ones."],"supporting_citations":[{"why":"Supplies the linearized pseudo-measurement shape update that keeps the recursions closed.","marker":"[40]"},{"why":"Establishes Trajectory Set Theory and the trajectory PHD/CPHD recursions on which TPHD-E and TCPHD-E are built.","marker":"[46]"},{"why":"Combines Trajectory Set Theory with a GIW-PHD filter for extended targets; supplies the baseline and the trajectory-extension idea.","marker":"[33]"},{"why":"Provides the extended-target PHD filter with random matrices and the Gaussian-mixture update and pruning pattern.","marker":"[21]"},{"why":"Provides the extended-target CPHD recursion and probability generating function machinery reused here.","marker":"[22]"},{"why":"Provides the GIW-PHD implementation details used for the TPHD-E recursion derivation.","marker":"[48]"},{"why":"Provides the extended-target CPHD inference used for the TCPHD-E recursion derivation.","marker":"[49]"},{"why":"Defines the trajectory metric used to evaluate trajectory completeness, missed cost, and switch cost.","marker":"[54]"},{"why":"Defines the Gaussian Wasserstein distance used to evaluate shape accuracy.","marker":"[53]"},{"why":"Supplies the real traffic-intersection dataset used in the second experimental scenario.","marker":"[55]"}],"fun_headline_variants":["Trajectory PHD/CPHD filters that also output ellipse orientation and axes","Shape and trajectory captured in one closed-form filter update","New filters track extended targets' full paths and ellipse extent","GM-TCPHD-E leads in shape accuracy and trajectory completeness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the approximation that the ellipse's shape can be updated with a linearized pseudo-measurement and that each trajectory's state history can be stored with a block-diagonal covariance that ignores correlations between different times; the paper concedes this second part is slightly wrong, so everything rests on that error being negligible.","fun_headline_variants_meta":{"raw":{"variants":["Trajectory PHD/CPHD filters that also output ellipse orientation and axes","Shape and trajectory captured in one closed-form filter update","New filters track extended targets' full paths and ellipse extent","GM-TCPHD-E leads in shape accuracy and trajectory completeness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1731,"prompt_tokens":908,"completion_tokens":823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":751}},"tokens_in":524,"tokens_out":823,"duration_ms":7045,"temperature":1.0,"reasoning_tokens":751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:36:32.995348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the GM-TCPHD-E filter to a single, highly elongated target executing a sharp turn with low detection probability, and compare the estimated $\\theta,l_1,l_2$ with ground truth over repeated Monte Carlo runs: if the pseudo-measurement linearization is inadequate, orientation error or Gaussian Wasserstein distance will fail to converge even when cardinality is correct. A second check compares the block-diagonal trajectory covariance against a full-scan implementation on two closely spaced, parallel-moving targets; if cross-time correlations matter, the block-diagonal version should show track fragments or elevated switch cost.","supporting_citations":[{"cited_title":"Sind: A drone dataset at signalized intersection in china,","cited_arxiv_id":null,"evidence_quote":"Supplies the real traffic-intersection dataset used in the second experimental scenario."},{"cited_title":"Metrics for performance evaluation of elliptic extended object tracking methods,","cited_arxiv_id":null,"evidence_quote":"Defines the Gaussian Wasserstein distance used to evaluate shape accuracy."},{"cited_title":"Tracking the orientation and axes lengths of an elliptical extended object,","cited_arxiv_id":null,"evidence_quote":"Supplies the linearized pseudo-measurement shape update that keeps the recursions closed."},{"cited_title":"Trajectory phd and cphd filters,","cited_arxiv_id":null,"evidence_quote":"Establishes Trajectory Set Theory and the trajectory PHD/CPHD recursions on which TPHD-E and TCPHD-E are built."},{"cited_title":"Extended object tracking using sets of trajectories with a phd filter,","cited_arxiv_id":null,"evidence_quote":"Combines Trajectory Set Theory with a GIW-PHD filter for extended targets; supplies the baseline and the trajectory-extension idea."},{"cited_title":"A phd filter for tracking multiple extended targets using random matrices,","cited_arxiv_id":null,"evidence_quote":"Provides the extended-target PHD filter with random matrices and the Gaussian-mixture update and pruning pattern."},{"cited_title":"An extended target cphd filter and a gamma gaussian inverse wishart implementation,","cited_arxiv_id":null,"evidence_quote":"Provides the extended-target CPHD recursion and probability generating function machinery reused here."},{"cited_title":"Implementation of the giw-phd filter,","cited_arxiv_id":null,"evidence_quote":"Provides the GIW-PHD implementation details used for the TPHD-E recursion derivation."},{"cited_title":"Extended target tracking with a cardinalized probability hypothesis density filter,","cited_arxiv_id":null,"evidence_quote":"Provides the extended-target CPHD inference used for the TCPHD-E recursion derivation."},{"cited_title":"A metric on the space of finite sets of trajectories for evaluation of multi-target tracking algorithms,","cited_arxiv_id":null,"evidence_quote":"Defines the trajectory metric used to evaluate trajectory completeness, missed cost, and switch cost."}],"review_version":1}