Method, system and medium for deduplication and fusion of multi-source target data in unmanned aerial vehicle detection process

Abstract

This invention relates to the field of low-altitude security monitoring technology, and discloses a method, system, and medium for deduplication and fusion of multi-source target data during UAV detection. The method designs a coarse-to-fine two-level association between the edge and cloud sides to achieve accurate screening and determination of homogeneous trajectories. Combined with adaptive weighted fusion, it generates a unique system track, solving the problems of data redundancy and high false alarm rates in existing UAV multi-source target tracking. This achieves efficient deduplication and consistent track generation of multi-source detection data. By combining Remote ID authentication with an edge-cloud collaborative architecture, it improves UAV target recognition capabilities and response speed in complex environments. Remote ID authentication distinguishes between cooperative and non-cooperative UAVs, dynamically optimizes tracking priorities, and rationally allocates system computing and detection resources.

Description

Technical Field

[0001] This invention relates to the field of low-altitude security monitoring technology, and in particular to the identification of multi-source targets monitored by unmanned aerial vehicles (UAVs). Specifically, it relates to a method, system, and medium for deduplication and fusion of multi-source target data during UAV detection. Background Technology

[0002] With the development of the low-altitude economy, a large number of drones operate in cities, industrial parks, airports, and key areas, leading to an increasing demand for airspace security monitoring. Existing monitoring systems generally employ a multi-sensor collaborative detection architecture, utilizing radar and computer vision, to detect and identify drones. This multi-source heterogeneous data complementarity enables comprehensive capture of low, slow, and small drones. However, due to the different working principles, dispersed deployment locations, and varying sampling frequencies of various sensors, the same drone target is often independently captured by multiple devices and reported as multiple "suspected targets," resulting in system-level track redundancy, increased false alarm rates, and chaotic command and decision-making.

[0003] Multi-sensor information fusion technology, as the core of this architecture, directly determines the accuracy, real-time performance, and reliability of target detection, and is currently a research hotspot in the field of UAV monitoring. Traditionally, for multi-source UAV target identification and fusion, trajectory association and deduplication are used to determine whether multiple trajectories originate from the same target. Existing methods are divided into two levels: coarse association and fine association. Coarse association uses spatial location as the core metric, employing a fixed distance threshold to achieve rapid trajectory matching, or using density clustering algorithms such as DBSCAN and OPTICS to cluster discrete trajectories for deduplication. A secondary screening is then performed based on motion model prediction matching on top of coarse association. However, traditional coarse association uses a fixed distance threshold, making it susceptible to sensor positioning errors, leading to incorrect merging or missed associations. Fine association relies heavily on single features such as position and velocity, and lacks globally unique matching constraints, easily resulting in redundant "one-to-many" matching, and has poor adaptability in low-altitude environments with multiple targets and high trajectory density. Moreover, the use of preset fixed similarity metric weights in trajectory association judgment does not achieve dynamic adjustment according to the detection environment, resulting in low reliability in complex scenarios with frequent UAV maneuvers. Summary of the Invention

[0004] The purpose of this invention is to provide a method based on multi-sensor spatiotemporal unification, achieving accurate screening and determination of homogeneous trajectories through coarse-fine two-level correlation via edge-cloud collaboration, and generating unique system tracks by adaptive weighted fusion. This addresses the problems of multi-source data redundancy and high false alarm rates in existing UAV multi-source target tracking, achieving efficient deduplication and consistent track generation of multi-source detection data. Furthermore, by combining Remote ID authentication and an edge-cloud collaborative architecture, it improves UAV target recognition capabilities and response speed in complex environments. Through Remote ID authentication, it effectively distinguishes between cooperative and non-cooperative UAVs, enhancing the targeting of supervision. It automatically adjusts tracking priorities according to target type, increasing the tracking priority for non-cooperative targets and continuously tracking cooperative targets with normal priority, thus rationally allocating the system's computing and detection resources.

[0005] According to a first aspect of the present invention, a method for deduplication and fusion of multi-source target data during UAV detection is proposed, comprising the following steps:

[0006] Using a unified time source as the time reference, time calibration is performed on multiple sensors to unify the spatiotemporal reference. The multiple sensors include at least two of the following: detection radar, radio direction finder, acoustic detection array, and visual detection module.

[0007] Acquire detection data of UAV targets in the detection area from each sensor, extract the core motion parameters of the target, generate raw trajectory data including timestamp, three-dimensional position, velocity, and heading, and upload it to the edge node in real time;

[0008] Edge nodes perform coarse correlation by quickly matching spatial locations based on the received raw trajectory data, filter candidate homogeneous trajectory pairs, and upload them to the cloud server;

[0009] The cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the basis for association, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they belong to the same target; and

[0010] Trajectories that are determined to be the same target after association are adaptively weighted and fused to generate a unique system track, which is then output externally. Trajectories that are not associated are retained as independent targets, enter the subsequent tracking process, and continue to participate in association determination within the subsequent preset period.

[0011] As an optional implementation, the cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the association basis, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they belong to the same target, specifically including:

[0012] Receive the full data of candidate homogeneous trajectory pairs, and extract trajectory feature vectors reflecting the motion stability of the UAV for each candidate homogeneous trajectory pair, including the trajectory angle change rate, velocity variance, and altitude layer consistency feature index.

[0013] The extracted track angle change rate, velocity variance, and altitude layer consistency index were normalized.

[0014] An improved weighted Hausdorff distance and cosine similarity fusion metric method is adopted to calculate the comprehensive similarity S of candidate trajectory pairs, and to perform multi-dimensional matching from the two dimensions of trajectory spatial morphology and feature vector.

[0015] The calculated comprehensive similarity S is compared with the preset similarity threshold Th. If the comprehensive similarity S is greater than or equal to the preset similarity threshold Th, the candidate trajectory pair is determined to be the same UAV target; otherwise, the candidate pair is removed.

[0016] As an optional implementation, trajectories that are determined to be the same target after association are adaptively weighted and fused to generate a unique system track, which is then output externally. This includes the following process:

[0017] For multi-source trajectories identified as the same target, an adaptive weighted fusion is performed using a hierarchical structure of the federated Kalman filter algorithm.

[0018] Each sensor involved in the fusion is treated as an independent sub-filter. Each sub-filter performs independent Kalman filtering on its own trajectory data to complete state prediction and measurement update, thereby obtaining local estimates of the trajectory parameters and local estimation error covariance of each sensor.

[0019] The real-time velocity measurement error covariance matrix Q of each sensor participating in the fusion i Based on this, the inverse variance method is used to adaptively calculate the fusion weight wi;

[0020] The main filter, based on the fusion weight wi, weights and fuses the local estimates of all sub-filters to obtain the globally optimal estimate. This value represents the fused core motion parameters of the UAV target, including its three-dimensional position, velocity, heading, and timestamp.

[0021] The fusion parameters obtained from the global optimal estimate are continuously spliced ​​together according to the time series to generate a unique system track.

[0022] As an optional implementation, the method further includes the following steps:

[0023] A Remote ID resolution module is integrated into the cloud server. This module receives Remote ID radio frequency signals actively broadcast by drones within the detection area in real time, and analyzes the signals to extract the drone's core digital identity information: UAS ID (unique drone identifier), MAC address, and flight reporting information (takeoff point, planned route, approved flight altitude, and flight time period).

[0024] The generated unique system track is spatiotemporally matched with the parsed Remote ID data to find Remote ID radio frequency signals with a spatial position deviation ≤ preset distance threshold and a timestamp deviation less than or equal to preset period T, which are used for the initial binding of track and digital identity.

[0025] Based on the spatiotemporal matching results and Remote ID reporting information, the target type is labeled:

[0026] If the spatiotemporal matching is successful, and the flight reporting information in the Remote ID is consistent with the actual flight parameters of the system track, then the target will be marked as a cooperative target.

[0027] If a spatiotemporal match is successful but there is no valid flight reporting information, or no Remote ID data is matched, the target will be marked as a non-cooperative target.

[0028] According to a second aspect of the present invention, a computer system is provided, comprising:

[0029] One or more processors;

[0030] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the method for deduplication and fusion of multi-source target data during UAV detection in the foregoing embodiments.

[0031] In a third aspect of the invention, a computer-readable medium for storing software is provided, the software including instructions executable by one or more computers, the instructions causing the one or more computers to perform operations including the process of the aforementioned method for deduplication and fusion of multi-source target data during UAV detection.

[0032] Compared with existing technologies, the significant advantages of the method for deduplication and fusion of multi-source target data in the UAV detection process of the present invention are as follows:

[0033] 1) This invention designs a coarse-fine two-level trajectory association mechanism for edge-cloud collaboration. On the edge side, based on the target's three-dimensional spatial position, a nearest neighbor-global uniqueness hypothesis algorithm is used for rapid matching of coarse screening. The association threshold (5m~50m) is dynamically adjusted by the sensor positioning error covariance to replace the traditional fixed threshold, which quickly eliminates non-similar trajectories and greatly reduces the computing pressure on the cloud. On the cloud side, stable features are extracted from candidate trajectory pairs and high-precision matching is performed to achieve accurate determination of homogeneous trajectories, taking into account both deduplication efficiency and accuracy.

[0034] 2) This invention proposes a judgment logic based on stability indicators, extracting three stable features that are less affected by UAV maneuvers: track angle change rate, velocity variance, and altitude layer consistency, to construct a trajectory feature vector, breaking through the limitations of traditional single position feature correlation; combined with an improved weighted Hausdorff distance and cosine similarity fusion measurement method, it calculates comprehensive similarity from two dimensions: trajectory spatial morphology and motion features, to determine the source of the sameness, improving the reliability of the sameness determination; at the same time, the initial similarity threshold is set to 0.85 and supports online learning and updating, which can adapt to the complex and ever-changing low-altitude detection environment;

[0035] 3) This invention proposes an adaptive weighted fusion algorithm based on a federated Kalman filter. Through a hierarchical fusion strategy of sub-filters and main filters, and based on the covariance of real-time velocity measurement errors of each sensor, the fusion weights are adaptively allocated by the inverse variance method (the higher the accuracy, the greater the weight), replacing the traditional static weights. This ensures the accuracy and stability of the fused trajectory, realizes fault-tolerant fusion of multi-source heterogeneous trajectory data, and improves the robustness and accuracy of the system.

[0036] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0037] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0038] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings.

[0039] Figure 1 This is a flowchart illustrating a method for deduplication and fusion of multi-source target data during UAV detection according to an embodiment of the present invention.

[0040] Figure 2 This is a flowchart illustrating a method for deduplication and fusion of multi-source target data during UAV detection according to another embodiment of the present invention.

[0041] Figure 3 It is based on Figure 1 The flowchart of the embodiment shows the process of using spatial location fast matching to perform coarse association and filter candidate homogeneous trajectory pairs.

[0042] Figure 4 It is based on Figure 1 The embodiment illustrates the process of a cloud server performing multi-dimensional feature measurement on candidate homologous trajectory pairs to achieve precise correlation of homologous trajectories.

[0043] Figure 5 Yes, according to Figure 1 A schematic diagram illustrating the process of adaptive weighted fusion of multi-source trajectories of the same target using the federated Kalman filter algorithm in an embodiment. Detailed Implementation

[0044] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0045] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0046] Combined with appendix Figure 1 The example shown illustrates a method for deduplication and fusion of multi-source target data during UAV detection according to an embodiment of the present invention, comprising the following steps:

[0047] Step 1: Using a unified time source as the time reference, perform time calibration on multiple sensors to unify the spatiotemporal reference. The multiple sensors include at least two of the following: detection radar, radio direction finder, acoustic detection array, and visual detection module.

[0048] Step 2: Acquire the detection data of each sensor on the UAV target in the detection area, extract the core motion parameters of the target, generate the raw trajectory data including timestamp, three-dimensional position, speed and heading, and upload it to the edge node in real time;

[0049] Step 3: The edge nodes perform coarse correlation by quickly matching spatial locations based on the received raw trajectory data, filter candidate homogeneous trajectory pairs, and upload them to the cloud server;

[0050] Step 4: The cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the association basis, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they belong to the same target; and

[0051] Step 5: Adaptively weightedly fuse the trajectories that are determined to be the same target after association to generate a unique system track and output it externally; the unassociated trajectories are retained as independent targets, enter the subsequent tracking process, and continue to participate in association determination within the subsequent preset period.

[0052] As an optional implementation, in step 1, a unified time source can be deployed for sensors located in different locations. In this example, a combination of BeiDou satellite time synchronization and PTP precision time protocol is used to perform real-time time calibration on four types of sensors: radar, photoelectric, radio direction finding, and acoustic array. This ensures that the time synchronization accuracy of the output data of all sensors is ≤1ms, so that the local timestamp of each sensor is consistent with the unified time reference of the system, and clock drift deviation is eliminated.

[0053] At the same time, the sampling frequency of all sensors is unified to a common time base of 10Hz (i.e., sampling period Δt=0.1s) to achieve sampling frequency normalization, ensure that the sampling of multi-source data in the time dimension remains synchronized, and achieve alignment of multi-source detection data on the time axis.

[0054] In an embodiment of the present invention, in step 2, multiple sensors independently detect the UAV target, acquire the raw detection data of the UAV target and standardize the format to eliminate the differences in the output data structure of heterogeneous sensors, and upload the raw trajectory detection data to the edge node.

[0055] Among them, four types of sensors, based on a calibrated unified time reference, independently detect and continuously track UAV targets within the detection area in real time. Each sensor extracts the core motion parameters of the target according to its own working principle, and finally generates raw trajectory data containing timestamps, three-dimensional positions, velocities, and headings.

[0056] Detection radar: The core function is to acquire the target's three-dimensional spatial position (x, y, z), radial velocity, and heading angle;

[0057] Visual detection module (photoelectric visual image sensing): supplements the continuity of target position through visual positioning and corrects local deviations in radar positioning;

[0058] Radio direction finder: Combines radio frequency signals to obtain the azimuth and distance of the drone, assisting in positioning;

[0059] Acoustic detection array: can supplement position and speed information for drones with low speed and low radio frequency characteristics.

[0060] It should be understood that each sensor uses its own logic to calculate and obtain the drone's motion data, forming initial trajectory data.

[0061] Each sensor encapsulates the raw trajectory data it generates according to the unified message format (Protobuf or JSON format) preset by the edge nodes. The encapsulated data must include the following core fields: timestamp t (calibration), three-dimensional position coordinates (x,y,z), flight speed v, and heading angle θ. The standardized data is then uploaded to the edge nodes in real time. After receiving and storing the data, the edge nodes perform coarse correlation on the edge side.

[0062] As an optional implementation, in step 3, the edge node performs coarse correlation based on the received raw trajectory data through rapid spatial location matching, filters candidate homologous trajectory pairs, and uploads them to the cloud server, such as... Figure 3 The examples shown include:

[0063] The edge node extracts the three-dimensional spatial coordinates of each trajectory from all received raw trajectory data, denoted as the coordinates (x, y) of trajectory i. i ,y i ,z i The coordinates of trajectory j (x) j ,y j ,z j ), i≠j, represent the trajectories generated by different sensors;

[0064] Calculate the spatial distance d between any two trajectories i and j to determine whether they are spatially adjacent trajectories. As an example, we use three-dimensional Euclidean distance as a metric for spatial similarity, as follows:

[0065] ;

[0066] In the formula, d is the three-dimensional Euclidean distance between the two trajectories (in meters); x i / y i / z i Let x be the three-dimensional spatial coordinates of the i-th trajectory; j / y j / z j Let be the three-dimensional spatial coordinates of the j-th trajectory;

[0067] The correlation threshold D is dynamically calculated based on the covariance of the sensor positioning error. The range of the correlation threshold D is 5m–50m, which can be adapted to the positioning accuracy of different sensors. The positioning error of photoelectric sensors is small, so the threshold is 5-10m; the positioning error of radar is large in complex environments, so the threshold is 30-50m.

[0068] Based on the nearest neighbor-global uniqueness hypothesis algorithm, the trajectories are coarsely correlated to find the spatially nearest suspected homogeneous trajectory for each trajectory as an initial matching pair; if no trajectory meets the conditions, it is marked as an independent trajectory.

[0069] Successfully matched trajectory pairs are marked as candidate homologous trajectory pairs, and the full data of the candidate homologous trajectory pairs is uploaded to the cloud server for subsequent fine correlation; for independent trajectories that do not find a matching object, they are temporarily stored in the edge node and enter the subsequent continuous tracking process.

[0070] As an optional implementation, the step of dynamically calculating the correlation threshold D based on the sensor positioning error covariance includes:

[0071] The sensor's correlation threshold D is calculated as follows:

[0072] ;

[0073] Where k is the confidence coefficient, tr is the trace operation of the matrix, that is, the sum of the diagonal elements of the sensor positioning error covariance matrix; P i P j Let be the positioning error covariance matrix of the i-th and j-th sensors. It is determined by the sensor's own performance and real-time operating status. When disturbed, the trace value of the matrix increases and the threshold is expanded accordingly.

[0074] As an example, the confidence coefficient k takes values ​​of 2-3, corresponding to a confidence interval of 95%-99.7%.

[0075] As an optional embodiment, the aforementioned algorithm based on the nearest neighbor-global uniqueness assumption performs coarse correlation on the trajectory, specifically including the following operations:

[0076] For each trajectory, iterate through all other trajectories and find the other trajectory with the smallest spatial distance d that satisfies d < D. This is used as the initial matching pair. Each pair of trajectories can only be matched once globally. If a trajectory matches multiple objects, only the pair with the smallest distance is kept, and all other matches are discarded.

[0077] Therefore, by ensuring that each pair of trajectories can only be matched once globally, redundant matching is eliminated by avoiding the repeated association of the same trajectory with multiple trajectories.

[0078] Therefore, successfully matched trajectory pairs can be marked as candidate homologous trajectory pairs, and the full data of the candidate pair can be uploaded to the cloud for subsequent fine association. Independent trajectories that do not find a matching object are temporarily stored on the edge side and enter the subsequent continuous tracking process. If the subsequent sampling points meet the association conditions, they can participate in the association again.

[0079] As an optional implementation, in step 4, the cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the basis for association, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they are the same target, thereby achieving accurate determination of homogeneous trajectories, filtering out mismatched pairs in coarse association, improving the reliability of homogeneous determination, and adapting to the complex flight environment of UAVs with frequent maneuvers.

[0080] As an optional implementation, in step 4, for the candidate homologous trajectory pairs uploaded from the edge side, complex feature extraction and matching processes are performed on the cloud server, such as... Figure 4 As shown, the specific steps include:

[0081] Receive the full data of candidate homogeneous trajectory pairs, and extract trajectory feature vectors reflecting the motion stability of the UAV for each candidate homogeneous trajectory pair, including the trajectory angle change rate, velocity variance, and altitude layer consistency feature index.

[0082] The extracted track angle change rate, velocity variance, and altitude layer consistency index are normalized. In this embodiment, min-max normalization is used to eliminate the dimensional differences of each feature and ensure the fairness of subsequent similarity calculation.

[0083] An improved weighted Hausdorff distance and cosine similarity fusion metric method is adopted to calculate the comprehensive similarity S of candidate trajectory pairs, and to perform multi-dimensional matching from the two dimensions of trajectory spatial morphology and feature vector.

[0084] The calculated comprehensive similarity S is compared with the preset similarity threshold Th. If the comprehensive similarity S is greater than or equal to the preset similarity threshold Th, the candidate trajectory pair is determined to be the same UAV target; otherwise, the candidate pair is removed.

[0085] As an optional embodiment, the extraction process of the aforementioned track angle change rate, velocity variance, and altitude layer consistency characteristic indicators is as follows:

[0086] 1) The rate of change of track angle α is used to characterize the trend of change in the UAV's heading:

[0087] α = Δθ / Δt;

[0088] In the formula, Δθ is the change in heading angle between adjacent sampling points of the trajectory, and Δt is the sampling time interval;

[0089] 2) Velocity variance σ v 2 This is used to characterize the stability of the drone's flight speed.

[0090] σ v 2 =(∑(vi-vˉ) 2 ) / n; i=1,2,3,…,n;

[0091] In the formula, n is the number of sampling points of the trajectory, vi is the flight speed of a single sampling point, and vˉ is the average flight speed of the trajectory;

[0092] 3) Height layer consistency γ is used to characterize the degree of high-level matching between two candidate trajectories:

[0093] γ = |zi-zj| / max (zi,zj);

[0094] In the formula, zi and zj are the average flight altitudes of the two candidate trajectories, respectively.

[0095] Among them, the smaller the height layer consistency γ value, the more consistent the height layer is. As an optional example, γ < 0.1 is judged as height consistency.

[0096] As an optional implementation, an improved weighted Hausdorff distance and cosine similarity fusion metric method is used to calculate the comprehensive similarity S of candidate trajectory pairs, performing multi-dimensional matching from two dimensions: trajectory spatial morphology and feature vectors, including:

[0097] 1) Based on the weighted Hausdorff distance Hw, measure the spatial morphological similarity between any two trajectories A and B:

[0098] Hw(A,B)=max{max a∈A min b∈B w(a,b)d(a,b),max b∈B min a∈A w(b,a)d(b,a)};

[0099] In the formula, A and B are the sampling point sets of the two trajectories, a is any point in sampling point set A, b is any point in sampling point set B, w(a,b) is the weight of points a and b, which can be allocated according to the real-time accuracy of the sensor; the higher the accuracy, the greater the weight. d(a,b) is the Euclidean distance between the two points. The smaller Hw is, the more similar the spatial shape of the trajectories.

[0100] 2) Based on cosine similarity (cosθ), measure the similarity of the feature vectors of two trajectories:

[0101] cosθ=X·Y / (∣X∣∣Y∣);

[0102] In the formula, X and Y are the standardized feature vectors of the two trajectories, X·Y is the vector dot product, and |X| and |Y| are the magnitudes of the vectors, respectively; the larger the cosθ, the better the feature match.

[0103] 3) Calculate the comprehensive similarity S of the candidate trajectory pairs based on the weighted Hausdorff distance Hw:

[0104] S=λ×(1-Hw / maxHw)+(1-λ)×cosθ;

[0105] Where maxHw is the maximum value of the weighted Hausdorff distance Hw, used to map Hw to the interval [0,1], and the final comprehensive similarity S takes values ​​in the range [0,1].

[0106] Wherein, λ is the comprehensive similarity fusion weight, and its value ranges from [0,1]. In a particularly preferred embodiment, λ is set to 0.5-0.7, emphasizing trajectory spatial morphology matching.

[0107] Therefore, the calculated comprehensive similarity S is compared with the preset similarity threshold Th, which is initially set to 0.85. If S ≥ 0.85, the candidate trajectory pair is determined to be the same UAV target; if S < 0.85, they are determined to be different targets, and the candidate pair is removed.

[0108] In an optional embodiment, the threshold Th supports online learning and updating, and can be dynamically adjusted according to the actual detection accuracy of the same source determination by the system (the threshold is increased when the false positive rate is high and decreased when the false negative rate is high), thereby adapting to the complex and ever-changing low-altitude detection environment.

[0109] As an optional implementation, in step 5, trajectories determined to be the same target after association are adaptively weighted and fused to generate a unique system track, completing the final deduplication and fusion of multi-source duplicate trajectories, ensuring the accuracy and stability of the fused track. Figure 5 As shown, it specifically includes the following processes:

[0110] For multi-source trajectories identified as the same target, an adaptive weighted fusion is performed using a hierarchical structure of the federated Kalman filter algorithm.

[0111] Each sensor involved in the fusion is treated as an independent sub-filter. Each sub-filter performs independent Kalman filtering on its own trajectory data to complete state prediction and measurement update, thereby obtaining local estimates of the trajectory parameters and local estimation error covariance of each sensor.

[0112] The real-time velocity measurement error covariance matrix Q of each sensor participating in the fusion i Based on this, the inverse variance method is used to adaptively calculate the fusion weight wi;

[0113] The main filter, based on the fusion weight wi, weights and fuses the local estimates of all sub-filters to obtain the globally optimal estimate. This value represents the fused core motion parameters of the UAV target, including its three-dimensional position, velocity, heading, and timestamp.

[0114] The fusion parameters obtained from the global optimal estimate are continuously spliced ​​together according to the time series to generate a unique system track.

[0115] As an optional implementation, the real-time velocity measurement error covariance matrix Q of each sensor participating in the fusion is used. i Based on this, the inverse variance method is used to adaptively calculate the fusion weights wi, including:

[0116] Obtain the real-time speed measurement error covariance matrix Qi for each sensor. The diagonal elements of the matrix represent the speed measurement variance of the sensor in the x, y, and z directions. The trace tr(Qi) of the matrix represents the sum of the diagonal elements. The smaller Qi is, the higher the real-time speed measurement accuracy of the sensor.

[0117] Calculate the fusion weight wi based on Qi and tr(Qi):

[0118] ∑wi=1, i=1,2,3,…,m;

[0119] In the formula, m is the number of sensors participating in the fusion;

[0120] wi=(1 / tr(Qi)) / (∑1 / tr(Qj));

[0121] In the formula, j = 1, 2, 3, ..., m.

[0122] Therefore, by ensuring that the highest accuracy sensor dominates the fusion process, adaptive fusion is achieved based on the real-time speed measurement error accuracy of each sensor.

[0123] Therefore, the fusion parameters obtained from the global optimal estimation are continuously spliced ​​together according to the time series to generate a unique system trajectory, which is then output to the low-altitude security command system or data fusion center for target tracking and decision-making. Unassociated independent trajectories are temporarily stored on the cloud side, and can be processed again if subsequent sampling points meet the association and fusion conditions.

[0124] Combined with appendix Figure 2 As shown, according to another embodiment of the present invention, a method for deduplication and fusion of multi-source target data during UAV detection is described above. Figure 1 In addition to the above embodiments, the following steps are also included:

[0125] A Remote ID resolution module is integrated into the cloud server. This module receives Remote ID radio frequency signals actively broadcast by drones within the detection area in real time, and analyzes the signals to extract the drone's core digital identity information: UAS ID (unique drone identifier), MAC address, and flight reporting information (takeoff point, planned route, approved flight altitude, and flight time period).

[0126] The generated unique system track is spatiotemporally matched with the parsed Remote ID data to find Remote ID radio frequency signals whose spatial position deviation from the system track is ≤ a preset distance threshold (less than or equal to 10m) and whose timestamp deviation is less than or equal to a preset period T (less than 1s), which are used to perform the initial binding of track and digital identity.

[0127] Based on the spatiotemporal matching results and Remote ID reporting information, the target type is labeled:

[0128] If the spatiotemporal matching is successful, and the flight reporting information in the Remote ID is consistent with the actual flight parameters of the system track, then the target will be marked as a cooperative target.

[0129] If a spatiotemporal match is successful but there is no valid flight reporting information, or no Remote ID data is matched, the target will be marked as a non-cooperative target.

[0130] Therefore, tracking priorities can be dynamically optimized based on the target type label. For example, tracking priorities can be automatically adjusted according to the target type: the tracking priority of non-cooperative targets can be increased (such as increasing the sampling frequency, increasing the number of fusion calculations, pushing the track to the security command terminal in real time, and triggering early warning prompts); cooperative targets can be continuously tracked according to the normal priority, and the system's computing and detection resources can be reasonably allocated.

[0131] Based on the method for deduplication and fusion of multi-source target data during UAV detection in the above embodiments, the present invention also proposes a computer system, including: one or more processors, and a memory.

[0132] The aforementioned memory is configured to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the method for deduplication and fusion of multi-source target data during UAV detection in the foregoing embodiments.

[0133] In conjunction with the method for deduplication and fusion of multi-source target data during UAV detection in the above embodiments, the present invention also proposes a computer-readable medium for storing software, the software including instructions executable by one or more computers, the instructions causing the one or more computers to perform an operation including the flow of the method for deduplication and fusion of multi-source target data during UAV detection in the aforementioned embodiments.

[0134] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for deduplication and fusion of multi-source target data during UAV detection, characterized in that, Includes the following steps: Using a unified time source as the time reference, time calibration is performed on multiple sensors to unify the spatiotemporal reference. The multiple sensors include at least two of the following: detection radar, radio direction finder, acoustic detection array, and visual detection module. Acquire detection data of UAV targets in the detection area from each sensor, extract the core motion parameters of the target, generate raw trajectory data including timestamp, three-dimensional position, velocity, and heading, and upload it to the edge node in real time; Edge nodes perform coarse correlation by quickly matching spatial locations based on the received raw trajectory data, filter candidate homogeneous trajectory pairs, and upload them to the cloud server; The cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the basis for association, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they are the same target. as well as Trajectories that are determined to be the same target after association are adaptively weighted and fused to generate a unique system track, which is then output externally. Unassociated trajectories are retained as independent targets and entered into the subsequent tracking process, and will continue to participate in association determination within the subsequent preset period.

2. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 1, characterized in that, The edge-side nodes perform coarse correlation based on the received raw trajectory data through rapid spatial location matching, filter candidate homogeneous trajectory pairs, and upload them to the cloud server, including: The edge side node extracts three-dimensional spatial position coordinates of each trajectory from all the received original trajectory data, denoted as coordinates (x i ,y i ,z i ) of the trajectory i and coordinates ((x j ,y j ,z j ) of the trajectory j, i≠j, representing trajectories generated by different sensors; Calculate the spatial distance d between any two trajectories i and j to determine whether they are spatially adjacent trajectories; The correlation threshold D is dynamically calculated based on the sensor positioning error covariance, and the range of the correlation threshold D is 5m–50m. Based on the nearest neighbor-global uniqueness hypothesis algorithm, the trajectories are coarsely correlated to find the spatially nearest suspected homogeneous trajectory for each trajectory as an initial matching pair; if no trajectory meets the conditions, it is marked as an independent trajectory. Successfully matched trajectory pairs are marked as candidate homologous trajectory pairs, and the full data of the candidate homologous trajectory pairs is uploaded to the cloud server for subsequent fine correlation; for independent trajectories that do not find a matching object, they are temporarily stored in the edge node and enter the subsequent continuous tracking process.

3. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 2, characterized in that, The dynamic calculation of the correlation threshold D based on the sensor positioning error covariance includes: The sensor's correlation threshold D is calculated as follows: ； Where k is the confidence coefficient, tr is the trace operation of the matrix, that is, the sum of the diagonal elements of the sensor positioning error covariance matrix; P i P j Let be the positioning error covariance matrix of the i-th and j-th sensors.

4. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 2, characterized in that, The algorithm based on the nearest neighbor-global uniqueness assumption for coarse correlation of trajectories specifically includes the following operations: For each trajectory, iterate through all other trajectories and find the other trajectory with the smallest spatial distance d that satisfies d < D. This is used as the initial matching pair. Each pair of trajectories can only be matched once globally. If a trajectory matches multiple objects, only the pair with the smallest distance is kept, and all other matches are discarded.

5. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 1, characterized in that, The cloud server extracts trajectory feature vectors from the candidate homogeneous trajectory pairs, combines them with the stable motion features of the UAV as the basis for association, and performs fine association of homogeneous trajectories through multi-dimensional feature measurement to determine whether they belong to the same target, specifically including: Receive the full data of candidate homogeneous trajectory pairs, and extract trajectory feature vectors reflecting the motion stability of the UAV for each candidate homogeneous trajectory pair, including the trajectory angle change rate, velocity variance, and altitude layer consistency feature index. The extracted track angle change rate, velocity variance, and altitude layer consistency index were normalized. An improved weighted Hausdorff distance and cosine similarity fusion metric method is adopted to calculate the comprehensive similarity S of candidate trajectory pairs, and to perform multi-dimensional matching from the two dimensions of trajectory spatial morphology and feature vector. The calculated comprehensive similarity S is compared with the preset similarity threshold Th. If the comprehensive similarity S is greater than or equal to the preset similarity threshold Th, the candidate trajectory pair is determined to be the same UAV target; otherwise, the candidate pair is removed.

6. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 5, characterized in that, The extraction process for the track angle change rate, velocity variance, and altitude layer consistency feature indicators is as follows: 1) The rate of change of track angle α is used to characterize the trend of change in the UAV's heading: α = Δθ / Δt; In the formula, Δθ is the change in heading angle between adjacent sampling points of the trajectory, and Δt is the sampling time interval; 2) Velocity variance σ v 2 This is used to characterize the stability of the drone's flight speed. σ v 2 =(∑(vi-vˉ) 2 ) / n;i=1,2,3,…,n; In the formula, n is the number of sampling points of the trajectory, vi is the flight speed of a single sampling point, and vˉ is the average flight speed of the trajectory; 3) Height layer consistency γ is used to characterize the degree of high-level matching between two candidate trajectories: γ = |zi-zj| / max (zi,zj); In the formula, zi and zj are the average flight altitudes of the two candidate trajectories, respectively.

7. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 5, characterized in that, The improved weighted Hausdorff distance and cosine similarity fusion metric method is used to calculate the comprehensive similarity S of candidate trajectory pairs, performing multi-dimensional matching from two dimensions: trajectory spatial morphology and feature vectors, including: 1) Based on the weighted Hausdorff distance Hw, measure the spatial morphological similarity between any two trajectories A and B: Hw(A,B)=max{max a∈A min b∈B w(a,b)d(a,b),max b∈B min a∈A w(b,a)d(b,a)}; In the formula, A and B are the sampling point sets of the two trajectories, a is any point in sampling point set A, b is any point in sampling point set B, w (a,b) is the weight of points a and b, and d (a,b) is the Euclidean distance between the two points. 2) Based on cosine similarity (cosθ), measure the similarity of the feature vectors of two trajectories: cosθ=X·Y / (∣X∣∣Y∣); In the formula, X and Y are the standardized feature vectors of the two trajectories, X·Y is the vector dot product, and |X| and |Y| are the magnitudes of the vectors, respectively. 3) Calculate the comprehensive similarity S of the candidate trajectory pairs based on the weighted Hausdorff distance Hw and the cosine similarity cosθ: S=λ×(1-Hw / maxHw)+(1-λ)×cosθ; Where maxHw is the maximum value of the weighted Hausdorff distance Hw, and λ is the weight of the comprehensive similarity fusion.

8. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 1, characterized in that, The trajectories that are determined to be the same target after association are adaptively weighted and fused to generate a unique system track, which is then output. This process includes the following steps: For multi-source trajectories identified as the same target, an adaptive weighted fusion is performed using a hierarchical structure of the federated Kalman filter algorithm. Each sensor involved in the fusion is treated as an independent sub-filter. Each sub-filter performs independent Kalman filtering on its own trajectory data to complete state prediction and measurement update, thereby obtaining local estimates of the trajectory parameters and local estimation error covariance of each sensor. The real-time velocity measurement error covariance matrix Q of each sensor participating in the fusion i Based on this, the inverse variance method is used to adaptively calculate the fusion weight wi; The main filter is based on the fusion weight w i The local estimates of all sub-filters are weighted and fused to obtain the globally optimal estimate, which is the fused core motion parameters of the UAV target, including three-dimensional position, velocity, heading, and timestamp; and The fusion parameters obtained from the global optimal estimate are continuously spliced ​​together according to the time series to generate a unique system track.

9. The method for deduplication and fusion of multi-source target data during UAV detection according to claim 5, characterized in that, The real-time velocity measurement error covariance matrix Q of each sensor participating in the fusion is described. i Based on this, the inverse variance method is used to adaptively calculate the fusion weights wi, including: Obtain the real-time speed measurement error covariance matrix Q for each sensor. i The diagonal elements of the matrix represent the variance of the sensor's velocity measurements in the x, y, and z directions. The trace tr(Qi) of the matrix represents the sum of the diagonal elements. The smaller Qi is, the higher the real-time velocity measurement accuracy of the sensor. Calculate the fusion weight wi based on Qi and tr(Qi): ∑wi=1, i=1,2,3,…,m; In the formula, m is the number of sensors participating in the fusion; wi=(1 / tr(Qi)) / (∑1 / tr(Qj)); In the formula, j = 1, 2, 3, ..., m.

10. The method for deduplication and fusion of multi-source target data during UAV detection according to any one of claims 1-9, characterized in that, The method further includes the following steps: A Remote ID resolution module is integrated into the cloud server. This module receives Remote ID radio frequency signals actively broadcast by drones within the detection area in real time, and analyzes the signals to extract the drone's core digital identity information: UAS ID (unique drone identifier), MAC address, and flight reporting information (takeoff point, planned route, approved flight altitude, and flight time period). The generated unique system track is spatiotemporally matched with the parsed Remote ID data to find Remote ID radio frequency signals with a spatial position deviation ≤ preset distance threshold and a timestamp deviation less than or equal to preset period T, which are used for the initial binding of track and digital identity. Based on the spatiotemporal matching results and Remote ID reporting information, the target type is labeled: If the spatiotemporal matching is successful, and the flight reporting information in the Remote ID is consistent with the actual flight parameters of the system track, then the target will be marked as a cooperative target. If a spatiotemporal match is successful but there is no valid flight reporting information, or no Remote ID data is matched, the target will be marked as a non-cooperative target.

11. A computer system, characterized in that, include: One or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of a method for deduplication and fusion of multi-source target data during UAV detection as described in any one of claims 1-10.

12. A computer-readable medium for storing software, characterized in that, The software includes instructions executable by one or more computers, which, through execution, cause the one or more computers to perform operations, including the process of a method for deduplication and fusion of multi-source target data during UAV detection as described in any one of claims 1-10.

Images