A verifiable admission method and system for isomorphic-aware candidate association edges

Abstract

The present application relates to multi-sensor information fusion and target association technology, and discloses a verifiable admission method and system for heterogeneous sensing candidate association edges. Target observations of at least two heterogeneous physical sensors are obtained and modal original evidence records are formed; the structural state and proof qualification are determined according to the heterogeneous relationship family, the two-side target state consistency set is constructed according to time synchronization, calibration and pose, measurement error and target motion boundary at a unified judgment moment, and a joint feasible set is formed in combination with motion, field of view and semantic constraints. Only when the joint feasible set is proved to be empty and the separation margin exceeds the numerical tolerance, a non-admission certificate is generated and the corresponding candidate edge is prohibited from being generated; the candidate edge is conservatively retained when the proof cannot be completed. Evidence hashing, boundary parameter, infeasibility witness and margin verification are performed on the certificate, and the candidate edge is restored when the verification fails. The present application is suitable for visible light, infrared, radar, radio frequency and remote identification and other heterogeneous sensing target association.

Description

Technical Field

[0001] This invention relates to the fields of multi-sensor information fusion, target association, set membership estimation, and computer graph processing. Specifically, it relates to a method and system for constructing a consistent set of target states at a unified decision time, verifying joint physical feasibility, and generating independently verifiable certificates for candidate edge non-admission decisions for observations from heterogeneous physical sensors such as visible light, infrared, radar, radio frequency, and remote identification. Background Technology

[0002] In anti-drone, low-altitude surveillance, intelligent transportation, robotic collaborative sensing, maritime target surveillance, and other multi-sensor target monitoring systems, visible light cameras, infrared sensors, radar, radio frequency (RF) receivers, and remote identification receivers are typically deployed simultaneously. Different sensors exhibit significant differences in observation methods, sampling frequencies, coordinate representations, measurement errors, calibration methods, and semantic information. The same physical target may be represented as an image detection box, an infrared response, a radar range-azimuth-velocity point, RF direction finding or positioning results, or a remote identification report carrying identification fields.

[0003] Existing target association methods typically generate candidate relationships for observation pairs from different sensors based on nominal time difference, Euclidean distance, Mahalanobis distance, field of view threshold, appearance similarity, or empirical threshold. Observation pairs that fall within the association threshold are added to the candidate graph, while those that exceed the threshold are directly excluded. Subsequently, nearest neighbor, bipartite graph matching, minimum cost flow, joint probabilistic data association, or other target assignments are performed on the candidate graph.

[0004] The aforementioned candidate generation methods typically make judgments based on a nominal projection position, a statistical center, or a preset threshold, which is insufficient to cover all permissible physical realizations under the combined effects of time synchronization deviation, sensor extrinsic parameter errors, platform pose errors, measurement errors, and target motion uncertainties. The fact that an observation pair is geographically far apart in the nominal state does not mean that the observation pair cannot correspond to the same target under all permissible errors and motion realizations.

[0005] On the other hand, the evidentiary basis for different cross-sensor relationships is inconsistent. For example, a visual-infrared relationship may have complete time synchronization and finite extrinsic parameters; a visual-radar relationship may only have partial pose information; radio frequency positioning results may only provide statistical standard deviation or circular error; and although remote identification reports have identity fields, their temporal or spatial mapping with non-cooperative detection targets may not have been verified. If candidate relationships are still excluded using a uniform empirical threshold, it is easy to implicitly write assumptions lacking model support into the candidate graph.

[0006] Existing candidate relation elimination processes often lack verifiable records corresponding to each decision that does not generate a candidate edge. Once boundary parameters become invalid, input evidence is replaced, numerical solution results are incorrect, or threshold settings are inappropriate, the system struggles to independently verify why the observation pair did not enter the candidate graph, and it also struggles to recover candidate edges and conservatively preserve ambiguity when verification fails.

[0007] Therefore, a technical solution is needed that explicitly distinguishes relational support states during the candidate association graph generation stage, transforms heterogeneous native observations into a bounded, consistent set of unified target states, excludes candidate edges only when it can be proven that "the joint state of the same target does not exist," and enables each non-admission decision to be verifiable and recoverable. Summary of the Invention

[0008] The technical problem to be solved by this invention is to avoid unsafely prohibiting candidate edges from entering the association graph based on nominal distance, statistical center or unverified relationship when there are inconsistencies in time, coordinates, calibration, pose, measurement error and semantic support between observations of heterogeneous physical sensors; and to provide a verifiable admission mechanism for candidate edges that prove that the observation pair to be judged cannot originate from the same target within the allowable physical perturbation range, and to perform conservative admission when the proof or verification conditions are not met.

[0009] To address the aforementioned technical problems, this invention provides a verifiable admission method for heterogeneous sensing candidate association edges, executed by a target association processing device, comprising the following steps:

[0010] Step S1: Acquire target observations with time information generated by at least two heterogeneous physical sensors, and construct modal native evidence records for each target observation. The records store modal identifiers, report timestamps, native measurements, sensor-side uncertainty information, quality or validity indication information, and audit metadata such as data source and processing version.

[0011] Step S2: Divide the heterogeneous relation family into units of reference side sensor and corresponding side sensor, select the observation pair to be judged, and determine the structural state and non-access certification qualification of the observation pair to be judged based on auditable evidence such as time synchronization, coordinate transformation or calibration and pose, measurement error, target motion boundary and semantic correspondence.

[0012] Step S3: For observation pairs that are not eligible for non-access certification, determine a unified judgment time. Based on the reporting time, time deviation boundary, coordinate or pose disturbance boundary, measurement error boundary, and target motion model of each side's observations, the reference side observations and corresponding side observations are respectively converted into a reference side target state consistency set under a unified judgment time. Consistency set with the corresponding target state .

[0013] Step S4: Construct a set of physical consistency constraints corresponding to the observation pair to be determined. The physical consistency constraints may include at least one of maximum speed, maximum acceleration, turning rate, sensor working distance, field of view, category, time validity, and identity semantics, forming a joint feasible set. .

[0014] Step S5, for the joint feasible set Perform a feasibility assessment. When When proven to be empty, calculate the infeasibility witness and separation margin that can be verified by an independent verification procedure; only when the separation margin is greater than the preset numerical tolerance, generate a non-admission certificate corresponding to the observation pair to be judged, and prohibit the generation of the corresponding candidate edge in the candidate association graph.

[0015] Step S6, when If an edge is not empty, cannot be proven to be empty, the finite boundary upon which the proof is based is missing, or the separation margin does not meet the requirements, no non-admission certificate is generated, and the corresponding candidate edge is generated or retained. Thus, proof failure, insufficient evidence, or numerical uncertainty are all interpreted as conservative admission, rather than as the observation pair being impossible to correlate.

[0016] Step S7: Perform input evidence and version consistency checks, structural state and boundary validity checks, infeasibility witness validity checks, and positive separation margin checks on the non-permitted certificates. If the verification passes, accept the non-permitted result and output the candidate association graph and audit log; if the verification fails, restore the corresponding candidate edges, or restore all affected non-permitted edges within the current relation family and regenerate the candidate association graph.

[0017] The present invention also provides a verifiable admission system for heterogeneous sensing candidate association edges, including a modal native evidence record construction unit, a heterogeneous relation family and observation pair determination unit, a reference-side state consistency set construction unit, a corresponding-side state consistency set construction unit, a joint physical constraint construction unit, a joint feasibility determination unit, an infeasibility witness and non-admission certificate generation unit, a certificate independent verification unit, a candidate edge generation and recovery control unit, and a candidate association graph and audit record output unit.

[0018] The present invention also provides an electronic device, a computer-readable storage medium, and a computer program product, wherein the processor of the electronic device implements the above-described method when executing a corresponding computer program.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects:

[0020] 1. The admission of candidate edges is described as a joint physical feasibility problem under permissible time, calibration, pose, measurement and motion perturbation, rather than directly determining that the observation pair is impossible to be associated based on a single nominal position or empirical distance threshold.

[0021] 2. Preserve the native observation semantics of different sensors and construct a consistent set that can be compared in the target state space, without requiring images, radar, radio frequency and identity reports to be forcibly fused into the same raw data representation.

[0022] 3. Non-admission proofs are performed only for relationships with auditable finite boundaries; observed but unverified or unsupported relationships are conservatively admitted, thus distinguishing between uncertainties within the measurement model and structural uncertainties regarding the validity of the cross-sensor model.

[0023] 4. Each decision that does not generate a candidate edge carries an infeasibility witness, boundary parameters, separation margin, and input evidence hash, which can be verified by an independent verification procedure without relying on the target's real identity or simulated truth value.

[0024] 5. When certificate verification fails, candidate edges are restored to prevent input evidence replacement, boundary invalidation, program version inconsistency, and numerical errors from being silently propagated to subsequent correlation solutions.

[0025] 6. In scenarios with high-density targets or a large number of cross-modal observations, observation pairs that have been proven physically infeasible can be excluded before candidate graph construction, reducing the burden of candidate edge storage, cost calculation, and subsequent assignment solving. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of a heterogeneous sensing candidate associated edge verifiable admission system.

[0027] Figure 2 A schematic diagram illustrating the generation of spatiotemporal consistent sets for constructing heterogeneous observation relationships;

[0028] Figure 3 A schematic diagram illustrating the generation of non-permitted certificates for candidate edges and the admission control of candidate graphs.

[0029] Figure 4 This is a diagram illustrating the verification and conservative recovery process for non-access certificates.

[0030] Figure labeling: 101, Multi-source heterogeneous sensor; 102, Modal native evidence recording construction unit; 103, Heterogeneous relation family and observation pair determination unit; 104, Reference side state consistency set construction unit; 105, Corresponding side state consistency set construction unit; 106, Joint physical constraint construction unit; 107, Joint feasibility determination unit; 108, Infeasibility witnessing and non-access certificate generation unit; 109, Certificate independent verification unit; 110, Candidate edge generation and recovery control unit; 111, Candidate association graph and audit record output unit; 201, Multi-source observation input; 202. Heterogeneous relation family partitioning; 203. Construction of state set at unified decision moment; 204. Joint feasible set determination and candidate relation output; 301. Construction of consistency set and joint constraints; 302. Non-admission certificate generation; 303. Candidate graph admission control; 401. Receiving candidate relations and non-admission certificates; 402. Evidence hashing and version consistency check; 403. Boundary parameter and structural state validity check; 404. Infeasibility witness and separation margin check; 405. Verification decision; 406. Verification pass processing; 407. Verification fail processing; 408. Final output. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the technical solutions of the present invention, but are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0032] I. Terminology and Data Representation

[0033] The "modal native evidence record" referred to in this invention refers to a target-level record formed while preserving the original semantics of sensor observations. For the j-th target observation, the modal native evidence record can be represented as:

[0034] .

[0035] in, Indicates modal identifier; Indicates the report timestamp; Represents the original observation; This indicates uncertainty information on the sensor side; Information indicating quality, confidence level, or validity; It represents audit metadata such as sensor number, data source, processing version, fault flag, and input hash.

[0036] In this invention, "structural state" refers to whether the current system has sufficient evidence to establish a bounded physical compatibility model for the observation pair to be determined. The states can be verified (V), observed but unverified (O), or unsupported (U). A V state indicates that the time, coordinates or calibrations and poses, measurement errors, and motion parameters involved in the construction of the consensus set have well-defined and auditable finite boundaries; an O state indicates that the relationship has some support in a statistical or empirical sense, but at least one necessary finite boundary has not been verified; and a U state indicates that there is a lack of sufficient evidence to establish a cross-sensor physical compatibility model.

[0037] The "target state consistency set" referred to in this invention refers to the set of all target states at a unified decision moment that could still produce the observation, given sensor observations, allowable errors, and target motion boundaries. It represents all physical states allowed by the observation, rather than a single nominal estimate.

[0038] The "non-admission certificate" referred to in this invention is a proof record used to demonstrate that a pair of observations to be judged does not have a common target state within the permitted physical boundaries of the record, and can be verified by an independent verification procedure. The non-admission certificate proves that the corresponding candidate edge should not enter the admission candidate graph, rather than proving the true target identity of any observation.

[0039] Let the current verifiable evidence state be ξ, which includes at least the clock deviation boundary, coordinate system definition, sensor calibration and platform pose boundary, measurement error boundary, sensor operating range, target motion boundary, mass state, and necessary semantic information. Let the unified decision time be... The target state x can include three-dimensional position, velocity, acceleration, target category state, or a subset thereof.

[0040] II. System Overall Structure

[0041] See Figure 1 The multi-source heterogeneous sensor 101 may include at least two of the following: an RGB sensor, an infrared sensor, a radar sensor, an RF sensor, and a Remote ID sensor. The sensors may be installed on the same platform or distributed across a network and transmit target-level observations to the target association processing device via wired or wireless networks.

[0042] Modal native evidence record construction unit 102 receives target observations from each sensor, parses timestamp, native measurement, measurement validity, and uncertainty fields, and generates modal native evidence records. Heterogeneous relation family and observation pair determination unit 103 divides relation families according to the reference side and the corresponding side, and determines the observation pairs that need to perform candidate edge admission judgment.

[0043] Reference-side state consistency set construction unit 104 and corresponding-side state consistency set construction unit 105 are generated at the unified decision time, respectively. and Joint physical constraint building block 106 is generated. The joint feasibility assessment unit 107 was constructed and assessed. .

[0044] when When proven to be empty and the separation margin meets the requirements, the infeasibility witness and non-access certificate generation unit 108 generate the infeasibility witness. Separation margin Non-access certificates The certificate independent verification unit 109 verifies the certificate. For For observation pairs that are not empty, cannot complete the proof, or fail the verification, the candidate edge generation and recovery control unit 110 generates, retains, or recovers candidate edges. The candidate association graph and audit record output unit 111 outputs the admission candidate graph, non-admission certificates, and audit records.

[0045] III. Heterogeneous Relational Families and Proof Qualifications

[0046] See Figure 2 The multi-source observation input 201 includes RGB, infrared, radar, RF, and Remote ID observations. The heterogeneous relation family division 202 can use RGB observations as the reference side to form RGB-infrared, RGB-radar, RGB-RF, and RGB-Remote ID relation families respectively; or it can use infrared, radar, persistent tracks, or other observations with object-level indexes as the reference side.

[0047] Each relation family independently stores the time base, coordinate transformation, sensor extrinsic parameters, platform pose, measurement error, and semantic fields used. Different relation families are not required to use the same state set construction method. An observation pair is only qualified for non-admission proof if all the necessary parameters of a relation family used for the current observation pair have finite and auditable boundaries.

[0048] For example, if there are verified upper bounds for time deviation, camera intrinsic and extrinsic parameter errors, mounting attitude errors, and detection center positioning errors between the RGB and infrared sensors, the corresponding observations can be set to the V state. If the RF positioning only provides an untruncated statistical distribution and does not cover a finite area that can cover all permissible cases, the corresponding observations are set to the O state, and non-admission verification is not performed.

[0049] Structural states are used to control eligibility for proof, not to directly determine whether two observations come from the same real target. O-state and U-state observation pairs can enter subsequent candidate graphs and be treated with conservative costs, or they can be recorded as having unknown relationships; however, they must not be excluded solely based on unrecorded projections, identities, or empirical assumptions.

[0050] IV. Unified Judgment Time and Consistency Set of Target State

[0051] The unified decision time for constructing the state set 203 is determined first. . This can be the current fusion cycle time, the later of the two observation report times, the world track prediction time, or a time grid point predefined by the system. Different observations are represented to the same time grid through time propagation relationships. This avoids directly comparing the nominal positions of asynchronous observations.

[0052] For observation Let the set of allowed parameters be... Including time deviation Calibration or pose disturbance Measurement error and motion model perturbation The target state consistency set can be represented as:

[0053] This enables x to generate observations under the corresponding time propagation, coordinate transformation, and measurement model. }

[0054] Corresponding side set Constructed according to the set of allowed parameters corresponding to observation j. The above set can be represented by interval boxes, ellipsoids, polyhedra, cones, second-order cones, zonotopes, support functions, finite sampling with continuous expansion, union of convex sets, or other set representations that can verify the inclusion relationship.

[0055] When using a bounded velocity and acceleration model, if a certain observation is at time... Give location area ,and Then the position at the unified determination time can be reached by the outer envelope. The Minkowski sum is formed with respect to the velocity reachable set and the acceleration reachable set. The velocity and acceleration boundaries used should be derived from the aircraft's operating range, tracking status, or audited scenario configuration, rather than post-hoc truth values.

[0056] For visual or infrared observation, the detection frame, detection center, or target mask forms a line of sight or cone with thickness through the camera's intrinsic and extrinsic parameters and the platform's pose boundary. If depth measurement is lacking, a bounded outer envelope can be formed by combining the sensor's effective working distance, the ground or height range, and the target's reachable area; if a finite set still cannot be obtained, non-access certification is not granted.

[0057] For radar observations, spatial shells and velocity stripes can be formed based on range, azimuth, elevation, and radial velocity ranges. Radar observations can be maintained in a polar coordinate measurement model, or a world coordinate state envelope can be formed after considering coordinate transformation errors.

[0058] For RF observations, direction finding results with finite angular errors can form direction bands; TDOA results with finite errors can form hyperbolic bands; and positioning results with finite positional error regions can form spatial outer envelopes. Non-access certificates will not be generated based solely on unbounded probability distributions or unverified statistical parameters.

[0059] For Remote ID observations, after verifying the report's authenticity, time accuracy, coordinate definition, and location error boundaries, a spatiotemporal state management system can be formed based on the broadcast location, velocity, and report time. The identity field can serve as a semantic constraint or audit source, but it cannot be used to directly determine whether a non-cooperative observation is the same as or different from a given broadcast identity without verification.

[0060] V. Combined Physical Constraints and Feasibility Assessment

[0061] Joint physical constraint construction unit 106 is generated based on relation families and the current evidence state. . It can limit the target's maximum displacement, maximum velocity, maximum acceleration, or turning rate between two reporting times. It can also limit the target to be located in the common observable area of ​​both sensors, meet category compatibility, or conform to verified identity time validity.

[0062] Joint feasible set determination and candidate relation output 204 construction:

[0063] .

[0064] like Non-empty indicates that there exists at least one target state that simultaneously satisfies both side observations and all recorded physical constraints, thus generating or retaining candidate edges. . A non-empty condition does not confirm that two observations definitely belong to the same target; it only indicates that the possibility cannot be ruled out at present.

[0065] like An empty value indicates that there is no common target state that can simultaneously explain both observations within all recorded perturbations. To avoid numerical solution failures being incorrectly interpreted as an empty set, it is also necessary to generate verifiable infeasibility witnesses and calculate the positive separation margin.

[0066] VI. Infeasibility Witnessing and Non-Access Certificates

[0067] See Figure 3 Consistent sets and joint constraints construct 301 sets , and Combined into The non-permissioned certificate generation 302 error selects the appropriate proof method based on the set representation.

[0068] In the interval representation embodiment, if two permissible intervals of a necessary state component remain strictly separated after considering the motion boundary, the corresponding interval endpoints, separation direction, and minimum gap can be recorded as interval contradiction witnesses. This witness is only valid if the interval separation of the necessary component holds true for all permissible parameters.

[0069] In the convex set embodiment, the separation vector a and the separation margin can be obtained. This ensures that the support function intervals of the two convex sets involved in the decision are strictly separated along the direction 'a'. The independent verification procedure verifies the results by recalculating or estimating the upper bound of the two support functions. .

[0070] In the linear constraint embodiment, if Represented as It can output λ≥0. and The Farkas-type witnesses have λ≥0. The verification procedure only needs to check the nonnegativity of λ, the equality residuals, and the negative margin; there is no need to re-execute the full feasibility search.

[0071] In second-order cone, convex quadratic constraint, or other convex optimization embodiments, a dual infeasibility proof generated by the solver can be output, along with the required matrices, vectors, scaling rules, and numerical tolerances written to the certificate. If the solver returns "infeasible" but cannot provide a verifiable witness, a non-admission certificate is not generated.

[0072] When the consensus set is the union of multiple convex subsets, a feasibility determination is performed for each possible combination of subsets. Only when all combinations obtain valid infeasibility witnesses can the original joint feasible set be considered empty. If any combination is unknown, times out, or has an uncertain value, candidate edges are conservatively generated or retained.

[0073] Non-access certificate This may include: observation pair identifiers, relation family identifiers, hashes of evidence records from both sides, structural state, unified decision time, allowed parameter sets or their boundaries, set representation parameters, physical consistency constraints, and infeasibility witnesses. Separation margin Numerical tolerance The generation program version, solver version, and generation time.

[0074] Candidate graph admission control 303 prevents the generation of candidate edges for observation pairs with valid non-admission certificates; for other observation pairs, candidate edges are generated or retained. This control occurs during the candidate graph construction phase and is independent of edge pruning based on association cost or global assignment optimality after candidate graph generation.

[0075] VII. Certificate Verification and Conservative Restoration

[0076] See Figure 4 The certificate independent verification unit 109 receives candidate relationships and non-admission certificates 401, performs evidence hash and version consistency checks 402, boundary parameter and structural state validity checks 403, infeasibility witness and separation margin checks 404, and forms a verification decision 405.

[0077] Evidence hashing and version consistency checks are used to verify that the observation records, calibration files, pose files, motion boundary configurations, and program versions used when generating the certificate are consistent with the current candidate map generation process. Certificate verification fails if any input is replaced and the certificate is not regenerated.

[0078] The validity check of boundary parameters and structural state is used to confirm that the observation pair to be judged is still in the V state, the time synchronization, extrinsic parameters, pose and measurement error boundaries are still within the validity period, and the outer envelope of the entire set does indeed cover the recorded allowable parameters.

[0079] Infeasibility witnessing and separation margin checks are performed based on the certificate type, including interval endpoint checks, support function separation checks, linear duality witnessing checks, or other corresponding verifications. Verification uses a numerical tolerance τ not less than the cumulative rounding error, data quantization error, and the upper bound of the solver residual.

[0080] If verification passes, execute Pass Processing 406, maintaining the non-admission status of the corresponding observation pair and recording an audit log. If verification fails, execute Fail Processing 407, transferring the corresponding observation pair to the Candidate Edge Generation and Recovery Control Unit 110, recovering the candidate edges, and performing conservative admission. If the affected single observation pair cannot be determined, all non-admission edges in the relevant relation family, time window, or certificate batch are recovered.

[0081] The final output 408 includes an admission candidate association graph, non-admission certificates, certificate verification results, recovery records, and audit logs. Audit logs can be saved by relation family and fusion period and written to local storage, a remote server, a message queue, or a tamper-proof log device.

[0082] VIII. Five-Source Anti-UAV Implementation Examples

[0083] In one anti-drone embodiment, the RGB and infrared sensors output object-level detection boxes and confidence levels; the radar outputs range, azimuth, pitch, and radial velocity points; the RF sensor outputs direction-finding angles or location regions; and the Remote ID sensor outputs broadcast identity, location, velocity, and time. The system forms RGB-IR, RGB-radar, RGB-RF, and RGB-Remote ID relationship families, respectively.

[0084] For the RGB-infrared relation family, the reference-side and corresponding-side detection boxes form sets of view frustums with pixel positioning errors, camera intrinsic and extrinsic parameter errors, and platform attitude errors, respectively. Combining the target distance range, flight altitude range, and the motion reachability range between the two reporting times, two sets of spatial states at a unified judgment time are constructed. If the two sets are still strictly separated by a separating plane under all permissible errors, a non-admission certificate is generated.

[0085] For the RGB-radar relation family, RGB detection forms the outer envelope of the view frustum, while radar measurements form the range-angle shell and radial velocity stripes. Joint physical constraints can also require that the target velocity direction be compatible with the radar radial velocity. If the reachable set of the view frustum and the radar shell do not intersect under all calibrations and pose perturbations, then generating corresponding candidate edges is prohibited.

[0086] For RGB-RF relation families, if the RF device only provides the statistical standard deviation and not a finite error region that covers all permissible cases, the corresponding observation pair for this relation family is set to 0, and non-admission proof is not performed. Even if the nominal RF location is far from the RGB line of sight, candidate edges are conservatively generated or preserved, or the relation is recorded as unknown.

[0087] For the RGB-Remote ID relation family, the Remote ID spatiotemporal state tube is constructed only when the broadcast report authenticity, time base, coordinate definition, and error boundary all meet the requirements. The identity field is used to record the source and semantic compatibility, and is not used directly to exclude non-cooperative target observations when a verification mapping is lacking.

[0088] In a faulty implementation, after a certificate is generated, the camera extrinsic file is updated, and the evidence hash and version consistency check fails. Instead of using the original non-admission conclusion, the system restores the corresponding candidate edges and reconstructs the consistency set and generates a candidate graph using the updated extrinsic boundary.

[0089] IX. Electronic Equipment and Program Implementation

[0090] The target association processing device can be deployed in sensor edge computing devices, fusion servers, anti-drone command terminals, vehicle computing platforms, or cloud-edge collaborative nodes. Electronic devices may include processors, memory, communication interfaces, and buses. The memory stores computer programs, relational family configurations, sensor calibration parameters, pose and error boundaries, motion boundaries, certificates, and audit logs.

[0091] The processor executes computer programs to perform tasks such as evidence record generation, relation family partitioning, structural state determination, construction of target state consistency sets, construction of joint physical constraints, joint feasibility determination, generation of infeasibility witnesses, certificate verification, candidate edge recovery, and candidate graph output. These units can be implemented using a central processing unit, a graphics processing unit, a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a combination thereof.

[0092] Computer programs may be provided as program instructions on a computer-readable storage medium or as computer program products transmitted via wired or wireless networks. Set construction and feasibility verification may be performed by interval computation libraries, linear programming solvers, second-order cone programming solvers, satisfiability modular theory solvers, or dedicated hardware modules.

[0093] Those skilled in the art will understand that, without departing from the core concept of this invention, the reference-side sensor, target state dimension, motion model, set representation, physical constraint form, infeasibility witness type, candidate graph type, audit record format, or deployment location can be changed. As long as the consistent set of target states on both sides is still constructed based on the auditable finite boundary, a verifiable non-admission certificate is generated after proving that the joint feasible set is empty, and the candidate edge is restored when the verification fails, all of these are equivalent implementations of the technical solution of this invention.

Claims

1. A verifiable admission method for heterogeneous sensing candidate association edges, characterized in that, Performed by a target association processing device, the process includes: acquiring target observations with time information generated by at least two heterogeneous physical sensors; generating modal native evidence records that retain the native measurement semantics of the corresponding sensors based on each target observation; determining the structural state and non-admission proof qualifications of the observation pair to be judged based on the time relationship, coordinate transformation or calibration and pose relationship, measurement error boundary, and semantic relationship for the heterogeneous relation family composed of the reference-side sensor and the corresponding-side sensor; determining a unified judgment time for the observation pair to be judged that possesses the non-admission proof qualifications, and constructing a reference-side target state consistency set and a corresponding-side target state consistency set based on the modal native evidence records, time propagation relationship, calibration and pose perturbation boundary, measurement error boundary, and target motion boundary; and setting the reference-side target state consistency set... A joint feasible set is constructed by combining the set of the corresponding side target state consistency set with at least one physical consistency constraint, representing that the observation pair to be judged can originate from the same target; a feasibility judgment is performed on the joint feasible set; when the joint feasible set is proven to be empty and the separation margin of the corresponding infeasibility witness is greater than a preset numerical tolerance, a non-admission certificate corresponding to the observation pair to be judged is generated, and the generation of the corresponding candidate edge in the candidate association graph is prohibited; when the joint feasible set is not empty, cannot be proven to be empty, or the separation margin is not greater than the preset numerical tolerance, the corresponding candidate edge is generated or retained; the non-admission certificate is verified; when the verification passes, a candidate association graph and audit record that do not contain the corresponding candidate edge are output; when the verification fails, the corresponding candidate edge is restored and conservative admission is performed.

2. The method according to claim 1, characterized in that, The modal native evidence record includes at least modal identifier, report timestamp, native observation, sensor-side uncertainty information, quality or validity indication information, and audit metadata; the native observation includes at least one of image target detection results, infrared response, radar spot or track, radio frequency direction finding or positioning results, and remote identification reports carrying identity information, and does not require that the native observations of different sensors be pre-converted to the same image plane or the same feature representation.

3. The method according to claim 1, characterized in that, The structural states include verified state, observed but unverified state, and unsupported state. When the time relationship, coordinate transformation or calibration and pose relationship, measurement error boundary, and target motion boundary required to establish the observation pair to be judged all have auditable finite boundaries, the observation pair to be judged is set to verified state and granted non-access certification qualification. When at least one of the finite boundaries is missing, the observation pair to be judged is set to observed but unverified state or unsupported state, and the generation of non-access certificates for the observation pair to be judged is prohibited. Corresponding candidate edges are conservatively generated or retained.

4. The method according to claim 1, characterized in that, Let the current verifiable evidence state be... For observation The defined set of allowed parameters is The unified determination time is The target state is ; The reference-side target state consistency set It consists of a target state that satisfies the following condition: there exists a state belonging to Time deviations, coordinate or pose disturbances, and measurement errors cause the target state x to propagate from the unified determination time to the reporting time of observation i via the target motion model, resulting in a difference between the observed and the target state x. The original observations are consistent; the corresponding target state consistency set According to observation The corresponding set of allowed parameters The target state consistency set is constructed as follows: For visible light or infrared observation, the target state consistency set includes a line-of-sight set, a set of view cones, or a spatial position envelope determined by the detection area, imaging model, camera intrinsic parameters, extrinsic parameters, platform pose, and their error boundaries; For radar observation, the target state consistency set includes a spatial shell, angular sector, velocity strip, or state envelope determined by range, azimuth, elevation, radial velocity, and their error boundaries; For radio frequency observation, the target state consistency set includes a direction band, hyperbolic region, or spatial envelope determined by direction finding angle, time difference of arrival, position estimation, or its finite error region; For long-range identification observation, the target state consistency set includes a spatiotemporal state tube formed by verified position, velocity, time accuracy, and their boundaries.

5. The method according to claim 4, characterized in that, The physical consistency constraints include at least one of the following: maximum speed constraint, maximum acceleration constraint, turning rate constraint, sensor working distance constraint, sensor field of view constraint, target category compatibility constraint, time validity constraint, and identity semantic constraint; assuming that the physical consistency constraints form a constraint set. Then the joint feasible set Represented as: .

6. The method according to claim 5, characterized in that, An infeasibility witness proving that the joint feasible set is empty includes at least one of the following: Witnessing of interval contradictions based on the fact that the allowed intervals of one or more target state components are mutually non-overlapping; A separating hyperplane and its positive separating margin are used to separate two convex state sets involved in the joint feasibility determination; when the joint feasible set is represented as a system of linear inequalities. When, satisfy , and Witness to the duality infeasibility , where τ is the preset numerical tolerance; The infeasibility proof record is output by a second-order cone programming, convex optimization, or interval constraint solver and can be verified by an independent verification program; when any target state consistency set or the physical consistency constraint is a non-convex set, the non-convex set is represented as the union of a finite number of verifiable subsets, and the non-admission certificate is generated only when the joint feasible set corresponding to each possible subset combination is proven to be empty; otherwise, the corresponding candidate edge is generated or retained.

7. The method according to any one of claims 1 to 6, characterized in that, The non-access certificate includes at least the observation pair identifier, relation family identifier, input evidence hash, structural state, unified judgment time, time synchronization boundary, calibration and pose boundary, measurement error boundary, target motion boundary, representation parameters of the target state consistency set, infeasibility witness, separation margin, preset numerical tolerance, and program version. Verification of the non-access certificate includes input evidence and version consistency checks, structural state and boundary validity checks, infeasibility witness validity checks, and checks that the separation margin exceeds the preset numerical tolerance. The candidate association graph is constructed as a bipartite or multipartite graph according to heterogeneous relation families, with graph nodes corresponding to sensor-side target observations or continuous tracks, and graph edges corresponding to non-valid non-access certificates. The certificate is a pair of observations to be judged; the candidate edge admission, non-admission certificate generation and non-admission certificate verification only use sensor-side observations, time information, calibration and pose parameters, physical working range, semantic relationships and the uncertainty boundaries determined by them, without using the target's real identity, simulation truth projection or post-evaluation labels; the audit record includes at least a summary of modal native evidence records, relation family, structural state, target state consistency set parameters, physical consistency constraints, joint feasibility judgment results, infeasibility witnesses, separation margin, certificate verification results, candidate edge recovery reasons and final candidate graph state, and the candidate association graph is output to the target assignment, target tracking, situation display, alarm or anti-drone decision-making module.

8. A verifiable admission system for heterogeneous sensing candidate association edges, characterized in that, The system includes at least two heterogeneous physical sensors and a target association processing device communicatively connected to the heterogeneous physical sensors. The target association processing device includes: a modal native evidence record construction unit, used to generate modal native evidence records based on target observations generated by the heterogeneous physical sensors; a heterogeneous relation family and observation pair determination unit, used to determine reference-side observations, corresponding-side observations, their structural states, and non-access certification qualifications; a reference-side state consistency set construction unit and a corresponding-side state consistency set construction unit, used to construct a reference-side target state consistency set and a corresponding-side target state consistency set respectively at a unified determination time; and a joint physical constraint construction unit and a joint feasibility determination unit, used to combine... The system comprises: a joint feasible set constructed by combining motion constraints, field of view constraints, or semantic constraints, and a determination of whether the joint feasible set is empty; an infeasibility witness and non-admission certificate generation unit, used to generate a non-admission certificate when the joint feasible set is proven to be empty and the separation margin of the corresponding infeasibility witness is greater than a preset numerical tolerance; a certificate independent verification unit, used to verify the input evidence, boundary parameters, infeasibility witness, and separation margin; a candidate edge generation and recovery control unit, used to generate or retain candidate edges for observation pairs that have not obtained a valid non-admission certificate, and to recover the corresponding candidate edges when certificate verification fails; and a candidate association graph and audit record output unit, used to output the candidate association graph, non-admission certificate, and audit record.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable by the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images