Multi-source heterogeneous data-based continuous tracking and track association method for marine target

Abstract

The invention discloses a maritime target continuous tracking and track association method based on multi-source heterogeneous data, and relates to the technical field of maritime monitoring and intelligent perception, and the method comprises the steps: constructing a multi-source observation model, obtaining multi-source observation data, carrying out the space-time registration of the multi-source observation data through linear interpolation and geographic coordinate transformation, obtaining observation track data, and carrying out the tracking of a target. Feature extraction and fusion are performed on the observation trajectory data to obtain fusion features, a nonlinear regression function is adopted to predict and update the track of the marine target based on the fusion features to obtain a multi-source tracking track, and a multi-detection-source space-time joint association module is constructed to obtain a tracking association track. And introducing an information comparison loss function based on the tracking association track to obtain a target joint optimization track. According to the method, the time evolution and space topology characteristics of the track are fully mined, the multi-target tracking robustness is improved, high-precision track association and continuous tracking capability can still be kept in a complex sea area environment, and the method has good engineering adaptability and expansibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of maritime surveillance and intelligent perception, and particularly to a maritime target continuous tracking and track association method based on multi-source heterogeneous data. BACKGROUND

[0002] In the existing maritime target monitoring and tracking technology, it mainly relies on a single sensor data source, such as sea surface radar monitoring, photoelectric imaging recognition or automatic identification system (AIS) signal analysis. Although this kind of single-source observation method has certain monitoring ability in a local range, it is easily affected by factors such as weather, sea conditions, electromagnetic interference and sensor viewing angle limitations in a complex marine environment, resulting in incomplete observation data, inconsistent spatio-temporal resolution, and decreased target recognition rate, thereby causing track breakage, target loss and association errors. In addition, due to the differences in detection mechanisms, different sensors have significant cross-source observation bias and heterogeneous feature differences. For example, radar sensors can provide high-precision distance and speed information, but false alarms are easily generated in target dense or strong sea clutter areas; photoelectric sensors can obtain target appearance features, but are limited by light and weather conditions; electronic reconnaissance sensors can achieve passive monitoring at a long distance, but have low positioning accuracy and are difficult to complete high-precision tracking tasks alone.

[0003] With the popularity of multi-satellite remote sensing, unmanned aerial vehicle monitoring and shore-based sensor networks, maritime observation data presents the characteristics of multi-source, heterogeneity and dynamic change. However, most of the existing multi-target tracking and track association algorithms are based on single-sensor or homogeneous data assumptions, lack unified modeling and fusion mechanisms for cross-sensor, cross-temporal and spatial, and multi-modal information, and are difficult to effectively deal with the observation differences and spatio-temporal inconsistency problems among multi-source data. This leads to the problems of track drift, repeated association or missed detection when traditional algorithms perform track fusion and target continuous tracking in a multi-source monitoring environment.

[0004] Some improved methods proposed by the academic and engineering fields, such as multi-target tracking algorithms based on Kalman filtering and particle filtering, target detection and re-identification models based on deep learning, and track association methods using graph neural networks have two main limitations: one is the inability to model the geometric and temporal bias of multi-source heterogeneous observations from a global perspective; the other is the lack of a unified optimization framework that can handle spatio-temporal misalignment and observation uncertainty simultaneously in a multi-source asynchronous observation environment.

[0005] In order to solve these problems, there is an urgent need for a maritime target continuous tracking and track association method that can fuse multi-source heterogeneous data, has spatio-temporal adaptive capability, and can realize cross-sensor continuous target tracking and track association. SUMMARY

[0006] To address the aforementioned issues, this application proposes a method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data, addressing the shortcomings of existing technologies in multi-source fusion, spatiotemporal calibration, and correlation modeling. It comprehensively utilizes computer algorithms such as multi-source remote sensing information fusion, spatiotemporal registration, deep learning, target association, and trajectory prediction to achieve continuous identification and dynamic tracking of moving maritime targets under multi-sensor observations. This effectively improves the accuracy of target identification and the stability of trajectory association in complex maritime environments. The specific steps are as follows: S1. Construct a multi-source observation model and use the multi-source observation model to observe maritime targets and obtain multi-source observation data; S2. Based on the differences in time synchronization, resolution and coordinate reference system of the multi-source observation model, the observation trajectory data is obtained by performing spatiotemporal registration of the multi-source observation data through linear interpolation and geographic coordinate transformation. S3. Extract and fuse features from the observed trajectory data to obtain fused features; S4. Based on the fusion features, a nonlinear regression function is used to predict and update the trajectory of the maritime target to obtain a multi-source tracking trajectory. S5. Construct a multi-source spatiotemporal joint correlation module, and use the multi-source spatiotemporal joint correlation module to correlate multi-source tracking tracks to obtain tracking correlation tracks; The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; S6. Based on the tracking and associated trajectory, an information comparison loss function is introduced to obtain the target joint optimized trajectory.

[0007] Preferably, the multi-source observation model described in S1 consists of multiple observation sensors; The observation sensors include active radar sensors, photoelectric imaging sensors, and passive electronic reconnaissance sensors. The multi-source observation model uses a two-layer Walker constellation structure to geometrically configure multiple detection platforms. The dual-layer Walker constellation structure includes a first layer Walker-1 group and a second layer Walker-2 group; The first layer, Walker-1, consists of 128 medium-high orbit detection units at an altitude of 1000 km and an inclination of 76°. The second layer, Walker-2, consists of 72 low-Earth orbit detection units at an altitude of 500 km and an inclination of 32°.

[0008] Preferably, the dual-layer Walker constellation structure uses the Geometric Diluted Precision (GDOP) index for geometric target layout; The expression for the geometric dilution precision (GDOP) index is: ; in, , , This is used to detect the components of geometric error in three dimensions. GDOP reflects the impact of the observed geometric distribution on positioning accuracy; the smaller the value, the better the positioning geometry.

[0009] To satisfy the target GDOP constraint, we have: ; in, For the duration of observation, Let be the angle between the stations. This formula shows that, under a certain observation time, by rationally designing the geometric distribution between the stations (i.e., the angle between them)... This can keep GDOP at an optimal or better level.

[0010] When extending the above GDOP constraints to multi-level orbital systems, the combined effect of the overall constellation geometry on GDOP must be considered.

[0011] For a two-layer Walker constellation structure, the joint constraints satisfying the GDOP optimality condition can be expressed as: ; in, Indicates the first Number of probe units in each orbital layer This indicates the orbital inclination angle of that layer. This condition shows that when the detection units of each orbital layer satisfy a certain equilibrium relationship in spatial angular distribution, the GDOP of the entire constellation system can be globally minimized, thereby obtaining the optimal positioning geometric accuracy.

[0012] Preferably, the specific content of obtaining observation trajectory data by performing spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation in S2 includes mapping multi-source observation data to observation trajectory data, and the mapping relationship is as follows: ; in, This represents the original observation coordinates and timestamps, i.e., multi-source observation data. To unify the spatiotemporal alignment results, i.e., the observation trajectory data, under a unified coordinate system, express.

[0013] Preferably, the specific content of feature extraction and fusion of the observed trajectory data in S3 to obtain fused features includes: For each observation trajectory segment in the observation trajectory data A multimodal feature encoder is used to extract its temporal, velocity, and heading information to obtain multimodal embedding features; The expression for multimodal embedding features is: ; in, For trajectory time series encoder, For the speed (SOG) feature extractor, For heading (COG) feature encoder; The multimodal embedding features are jointly obtained by using a temporal fusion transformer and a spatial fusion transformer: ; The TFT module captures the temporal dependencies between multiple frames through a self-attention mechanism, while the SFT module is based on the spatial proximity graph between targets. Aggregate local spatial context to achieve multi-objective interactive modeling.

[0014] Preferably, the spatial fusion transformer constructs an adjacency matrix based on a k-nearest neighbor graph. And a self-attention mechanism is used to achieve spatial feature aggregation: ; in, For learnable matrices, For feature dimensions; The time fusion transformer can employ a long sequence memory mechanism to fuse and encode the trajectory history of multiple time steps, thereby enhancing the dynamic prediction capability for complex maritime maneuvering targets.

[0015] Preferably, the specific content of S4, which uses a nonlinear regression function based on fused features to predict and update the trajectory of a maritime target to obtain a multi-source tracking trajectory, includes: By fusing features, a nonlinear regression function is used. Predicting future trajectory points; The prediction error is optimized by minimizing the L2 norm loss function to obtain the multi-source tracking track; The expression for minimizing the L2 norm loss function is: ; in, For the number of observed trajectories, The time window length, These are the actual trajectory points.

[0016] Preferably, the time matching submodule in S5 defines a matching score based on the Euclidean distance between the predicted trajectory and the current observation point, expressed as: ; in, For the first The predicted trajectory point and the first Matching score between observation points For the target predicted by the model at time... The position vector, For the sensor at any time The obtained target observation position vector.

[0017] Define time-related thresholds ,when Time, trajectory and Continuous in time; The spatial matching submodule is used to calculate the similarity between trajectory embedding features and normalize it into an association probability, expressed as follows: ; in, Temperature is a parameter used to adjust the smoothness of similarity. At that time, determine the trajectory and Belonging to the same goal; The joint optimization submodule combines the temporal and spatial matching matrices to form a global trajectory association map across sensors.

[0018] Preferably, the expression for the information contrast loss function is: ; The expression for the joint optimization trajectory is: ; in These are weighting parameters used to balance trajectory prediction accuracy and correlation accuracy; The information comparison loss This is achieved through the embedding and comparison of positive and negative sample trajectories.

[0019] A continuous maritime target tracking and trajectory correlation system based on multi-source heterogeneous data includes: Multi-source observation unit: Construct a multi-source observation model and use the multi-source observation model to observe maritime targets to obtain multi-source observation data; Spatiotemporal registration unit: Based on the differences in time synchronization, resolution and coordinate reference system of multi-source observation models, observation trajectory data is obtained by spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation; Feature fusion unit: Extracts and fuses features from observed trajectory data to obtain fused features; Track prediction unit: Based on fused features, a nonlinear regression function is used to predict and update the track of maritime targets to obtain multi-source tracking tracks; Spatiotemporal Joint Association Unit: Construct a multi-source spatiotemporal joint association module, and use the multi-source spatiotemporal joint association module to associate multi-source tracking tracks to obtain tracking associated tracks; The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; Track optimization unit: Based on the tracking and associated track, information comparison loss function is introduced to obtain the target joint optimized track.

[0020] In summary, the maritime target continuous tracking and track association method based on multi-source heterogeneous data of the present invention has the following advantages compared with traditional technologies: 1. This application constructs a dynamic observation constellation simulation system to model the observation characteristics of various sensors (including radar, photoelectric, electronic reconnaissance, etc.), and realizes unified spatiotemporal alignment of multi-source observation data based on an adaptive spatiotemporal registration module; 2. This application designs a multi-line fusion trajectory prediction module and a spatiotemporal joint correlation module. By fusing multimodal features, time series evolution features and spatial topological constraints, it realizes dynamic matching of multi-source observation trajectories and continuous target tracking. 3. This application can effectively eliminate multi-sensor observation bias, improve the accuracy and continuity of track correlation, and is suitable for ship surveillance, traffic control and situational awareness tasks in complex maritime environments.

[0021] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the steps of the maritime target continuous tracking and track association method based on multi-source heterogeneous data according to the present invention. Detailed Implementation

[0023] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0024] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0025] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0026] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0028] This invention addresses the issues of track breakage and unstable correlation of maritime targets under observation by multiple heterogeneous sensors due to cross-sensor observation bias, temporal asynchrony, and spatial registration errors. It provides a method for continuous tracking and track correlation of maritime targets based on multi-source heterogeneous data. By performing unified modeling, spatiotemporal registration, and multimodal feature fusion of data from multiple types of observation sensors, a continuous track tracking framework with global consistency and dynamic adaptability is established, thereby achieving high-precision, long-term continuous tracking and automatic track correlation of multiple targets in complex maritime environments.

[0029] Example 1 like Figure 1 As shown, the specific steps of the method for continuous tracking and track association of maritime targets based on multi-source heterogeneous data are as follows: S1. Constructing a multi-source observation model: Using this model to observe maritime targets and obtain multi-source observation data. It can be seen that the core idea of ​​this invention lies in constructing a dynamic information processing system that integrates heterogeneous multi-source data. Through multi-satellite, multi-sensor observation modeling, the observation parameters, spatial layout, and error models of multi-source sensors such as radar, electro-optical, and electronic reconnaissance are uniformly described. Based on the Geometric Diluted Precision (GDOP) criterion, an optimal constellation structure is designed, thereby ensuring the geometrical balance of multi-source detection in the spatial dimension.

[0030] Preferably, the multi-source observation model described in S1 consists of multiple observation sensors; The observation sensors include active radar sensors, photoelectric imaging sensors, and passive electronic reconnaissance sensors. To ensure the rationality of the geometric relationship of the multi-source observation, the multi-source observation model adopts a two-layer Walker constellation structure to geometrically configure the multiple detection platforms; The dual-layer Walker constellation structure comprises a first layer, Walker-1 (a high-inclination, high-orbit observation group), and a second layer, Walker-2 (a low-inclination, low-orbit observation group), to ensure global coverage and high observational overlap. Its orbital parameter set is defined as follows: ; in, For the orbital height, For eccentricity, For the track inclination angle, Right ascension of the ascending node, The perigee argument, It is the true near point angle.

[0031] The first layer, Walker-1, consists of 128 medium-high orbit detection units at an altitude of 1000 km and an inclination of 76°. The second layer, Walker-2, consists of 72 low-Earth orbit detection units at an altitude of 500 km and an inclination of 32°.

[0032] Preferably, the dual-layer Walker constellation structure uses the Geometric Diluted Precision (GDOP) index for geometric target layout; The expression for the geometric dilution precision (GDOP) index is: ; in, , , This is used to detect the components of geometric error in three dimensions (positioning uncertainty in three spatial directions). GDOP reflects the impact of the observed geometric distribution on positioning accuracy; the smaller the value, the better the positioning geometry.

[0033] To satisfy the target GDOP constraint, we have: ; in, For the duration of observation, Let be the angle between the stations. This formula shows that, within a certain observation time, by rationally designing the geometric distribution between stations (i.e., the angle between them)... This can keep GDOP at an optimal or better level.

[0034] When extending the above GDOP constraints to multi-level orbital systems, the joint influence of the overall constellation geometry on GDOP must be considered. For a two-level Walker constellation structure, the joint constraints satisfying the GDOP optimality condition can be expressed as: ; in, Indicates the first Number of probe units in each orbital layer This indicates the orbital inclination angle of that layer. This condition shows that when the detection units of each orbital layer satisfy a certain equilibrium relationship in spatial angular distribution, the GDOP of the entire constellation system can be globally minimized, thereby obtaining the optimal positioning geometric accuracy.

[0035] S2. Based on the differences in time synchronization, resolution, and coordinate reference system of the multi-source observation model, observation trajectory data is obtained by performing spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation. It can be seen that this application addresses the problem of time asynchrony and spatial deviation among multiple sensors by establishing an adaptive spatiotemporal alignment module (SAM). It utilizes geographic coordinate transformation and time interpolation mechanisms to achieve unified alignment of cross-source observations, ensuring that data from different observation platforms can be fused and calculated under the same reference system and time scale.

[0036] Because multi-source sensors differ in sampling frequency, coordinate reference system, and time delay, this module uses an adaptive spatiotemporal registration algorithm to align the observation data within a unified reference frame. Define the observation point set: ; in, Geographic coordinates Sampling time.

[0037] Preferably, the specific content of obtaining observation trajectory data by spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation in S2 includes mapping multi-source observation data to observation trajectory data, and the mapping relationship (registration function) is as follows: ; in, This represents the original observation coordinates and timestamps, i.e., multi-source observation data. To unify the spatiotemporal alignment results, i.e., the observation trajectory data, under a unified coordinate system, This represents the spatiotemporal alignment mapping matrix, which achieves time axis synchronization through interpolation functions and uses a geographic coordinate transformation matrix. Achieve spatial registration: This step can spatially map the observation trajectories of different sensors to a unified geographic reference system (latitude and longitude coordinates) and achieve synchronization in time through a unified sampling period.

[0038] This mapping unifies data from different sources to the same geographic reference system and time sampling interval, ensuring consistency in subsequent track fusion.

[0039] S3. Multi-clue feature extraction and fusion are performed on the observation trajectory data to obtain fused features. Time series, velocity (SOG), heading (COG) and neighborhood spatial features are extracted from multi-source observation data, and spatiotemporal joint modeling is achieved.

[0040] Preferably, the specific content of feature extraction and fusion of the observed trajectory data in S3 to obtain fused features includes: For each observation trajectory segment in the observation trajectory data ,in, These are latitude and longitude coordinates, For speed, For the heading angle, a multimodal feature encoder is used to extract its temporal, velocity, and heading information to obtain multimodal embedding features; The expression for multimodal embedding features is: ; in, For trajectory time series encoder, For the speed (SOG) feature extractor, , is the heading (COG) feature encoder; To integrate spatiotemporal correlation characteristics, a temporal fusion transformer (TFT) and a spatial fusion transformer (SFT) are designed.

[0041] A fusion feature is obtained by jointly applying a temporal fusion transformer and a spatial fusion transformer to the multimodal embedded features to enhance multidimensional information. ; The TFT module captures the temporal dependencies between multiple frames through a self-attention mechanism, while the SFT module is based on the spatial proximity graph between targets. By aggregating local spatial context and enabling multi-objective interactive modeling, the aforementioned dual transformers, combined with a jump connection structure, can significantly enhance the model's robustness and spatiotemporal awareness under multi-source uncertain observations.

[0042] Preferably, the spatial fusion transformer constructs an adjacency matrix based on a k-nearest neighbor graph. And a self-attention mechanism is used to achieve spatial feature aggregation: ; in For learnable matrices, For feature dimensions; The time fusion transformer can employ a long sequence memory mechanism to fuse and encode the trajectory history of multiple time steps, thereby enhancing the dynamic prediction capability of complex maritime maneuvering targets. Through the above-mentioned dual transformation structure, spatiotemporal fusion of multiple cues is realized, thereby enhancing the model's ability to characterize target motion patterns in complex dynamic scenarios.

[0043] Understandably, in the multi-source data fusion layer, this application proposes a multi-threaded fusion mechanism based on a dual-transformer structure. A Temporal Fusion Transformer (TFT) captures the temporal dynamic features of the target trajectory, while a Spatial Fusion Transformer (SFT) extracts the neighborhood constraints of the target in its spatial distribution, thereby achieving spatiotemporal joint modeling of the trajectory. The fused multimodal feature vector can express the spatiotemporal consistency features of the target in the embedding space, providing semantic support for subsequent trajectory prediction and association.

[0044] S4. Based on the fusion features, a nonlinear regression function is used to predict and update the trajectory of the maritime target to obtain a multi-source tracking trajectory. The future position of the target is predicted based on historical trajectory fragments, and the observation uncertainty is corrected.

[0045] Preferably, the specific content of S4, which uses a nonlinear regression function based on fused features to predict and update the trajectory of a maritime target to obtain a multi-source tracking trajectory, includes: By fusing features, a nonlinear regression function is used. Predicting future trajectory points; Based on fusion features Design a nonlinear prediction function : ; in, Indicating a view on the future The predicted position of the frame.

[0046] The prediction error is optimized by minimizing the L2 norm loss function to obtain the multi-source tracking track; The expression for minimizing the L2 norm loss function is: ; in, For the number of observed trajectories, The time window length, Using the actual trajectory points, the model parameters are adjusted through error backpropagation to achieve adaptive prediction and continuous updating of the target's motion trend.

[0047] By adjusting the time span through a sliding time window mechanism, adaptive predictions for different motion modes can be achieved. This step can continuously output dynamic predicted trajectories, providing prior references for subsequent track matching and association.

[0048] Understandably, during the continuous target tracking phase, this invention designs a trajectory prediction module based on a depth regression model. This module achieves dynamic position prediction and motion trend estimation by minimizing the Euclidean distance loss function between the predicted point and the actual trajectory point. To enhance the robustness of tracking, the algorithm introduces a multi-frame sliding time window mechanism, which dynamically adjusts the prediction step size according to the observation frequency.

[0049] S5. To address the track matching problem under multi-source observation, a multi-source spatiotemporal joint association module is constructed. Considering both temporal continuity and spatial similarity, this module associates multi-source tracking tracks to obtain associated tracks. To achieve cross-sensor track stitching and globally consistent association, this module establishes matching relationships between target trajectories based on a spatiotemporal joint probability map. In essence, during the track association stage, this invention calculates the association probability matrix between trajectories using the spatiotemporal joint association module. The temporal matching part is constrained by the time difference of trajectory points and the similarity of motion trends, while the spatial matching part calculates the track association probability using the inner product similarity of the target embedded features.

[0050] The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; Preferably, the time matching submodule in S5 defines a matching score based on the Euclidean distance between the predicted trajectory and the current observation point, expressed as: ; in, For the first The predicted trajectory point and the first Matching score between observation points For the target predicted by the model at time... The position vector, For the sensor at any time The obtained target observation position vector.

[0051] Define time-related thresholds ,when Time, trajectory and Continuous in time; The spatial matching submodule is used to calculate the similarity between trajectory embedding features and normalize it into an association probability, expressed as follows: ; in, Temperature is a parameter used to adjust the smoothness of similarity. At that time, determine the trajectory and Belonging to the same goal; The joint optimization submodule combines time and space matching matrices to form a global trajectory association map across sensors, enabling dynamic stitching and continuous updating of the target trajectory.

[0052] Preferably, the expression for the information contrast loss function is: ; The expression for the joint optimization trajectory is: ; in These are weighting parameters used to balance trajectory prediction accuracy and correlation accuracy; The information comparison loss This is achieved by embedding and comparing positive and negative sample trajectories to maximize the similarity between correctly paired trajectories and minimize the similarity between incorrect matches, thereby enhancing the robustness of cross-sensor matching.

[0053] S6. Based on the tracking and association of tracks, in order to improve the overall matching consistency, an information comparison loss function is introduced to obtain the target joint optimized track. The spatiotemporal joint association model is optimized by comparing the learning loss function to maximize the similarity of correctly associated trajectories and minimize the similarity of incorrectly matched trajectories, thereby achieving global consistent association of cross-sensor tracks.

[0054] To improve the continuity and global stability of the tracks, this step performs clustering and consistency optimization on the spatiotemporal correlation results. High-similarity trajectory segments are merged using a connected component clustering algorithm, and information comparison constraints are used to maximize the similarity of correctly matched tracks and minimize the similarity between incorrectly correlated tracks, thereby obtaining the globally optimal track sequence. ; The final output track sequence is continuous in time, smooth in space, and consistent across sensors.

[0055] Example 2 A continuous maritime target tracking and trajectory correlation system based on multi-source heterogeneous data includes: Multi-source observation unit: Construct a multi-source observation model and use the multi-source observation model to observe maritime targets to obtain multi-source observation data; This unit is used to uniformly model and configure parameters for observation information from different types of sensors in the marine monitoring system. The multi-source observation model includes radar, photoelectric, electronic reconnaissance, AIS (Automatic Identification System), and satellite remote sensing equipment, forming a multi-source heterogeneous detection network.

[0056] Spatiotemporal registration unit: Based on the differences in time synchronization, resolution and coordinate reference system of multi-source observation models, observation trajectory data is obtained by spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation; Feature fusion unit: Extracts and fuses features from observed trajectory data to obtain fused features; Track prediction unit: Based on fused features, a nonlinear regression function is used to predict and update the track of maritime targets to obtain multi-source tracking tracks; Spatiotemporal Joint Association Unit: Construct a multi-source spatiotemporal joint association module, and use the multi-source spatiotemporal joint association module to associate multi-source tracking tracks to obtain tracking associated tracks; The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; Track optimization unit: Based on the tracking and associated track, information comparison loss function is introduced to obtain the target joint optimized track.

[0057] The various units are connected through data flow and feature vectors, forming an integrated dynamic information processing system.

[0058] In summary, this application achieves unified modeling and spatiotemporal fusion of multi-source heterogeneous observation data such as radar, optoelectronic, and electronic reconnaissance, solves the registration problem of multi-source asynchronous observation, adopts a multi-modal fusion mechanism based on a dual-transformer structure, can fully explore the temporal evolution and spatial topological features of the track, improve the robustness of multi-target tracking, and introduces a spatiotemporal joint correlation mechanism and joint loss optimization strategy, enabling the algorithm to maintain high-precision track correlation and continuous tracking capability in complex marine environments. It has good engineering adaptability and scalability, and can be widely applied to scenarios such as maritime traffic supervision, ship identification, abnormal behavior detection, and maritime defense monitoring.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data, characterized in that, Includes the following steps: S1. Construct a multi-source observation model and use the multi-source observation model to observe maritime targets and obtain multi-source observation data; S2. Based on the differences in time synchronization, resolution and coordinate reference system of the multi-source observation model, the observation trajectory data is obtained by performing spatiotemporal registration of the multi-source observation data through linear interpolation and geographic coordinate transformation. S3. Extract and fuse features from the observed trajectory data to obtain fused features; S4. Based on the fusion features, a nonlinear regression function is used to predict and update the trajectory of the maritime target to obtain a multi-source tracking trajectory. S5. Construct a multi-source spatiotemporal joint correlation module, and use the multi-source spatiotemporal joint correlation module to correlate multi-source tracking tracks to obtain tracking correlation tracks; The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; S6. Based on the tracking and associated trajectory, an information comparison loss function is introduced to obtain the target joint optimized trajectory.

2. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 1, characterized in that, The multi-source observation model described in S1 consists of a variety of observation sensors; The observation sensors include active radar sensors, photoelectric imaging sensors, and passive electronic reconnaissance sensors. The multi-source observation model uses a two-layer Walker constellation structure to geometrically configure multiple detection platforms. The dual-layer Walker constellation structure includes a first layer Walker-1 group and a second layer Walker-2 group; The first layer, Walker-1, consists of 128 medium-high orbit detection units at an altitude of 1000 km and an inclination of 76°. The second layer, Walker-2, consists of 72 low-Earth orbit detection units at an altitude of 500 km and an inclination of 32°.

3. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 2, characterized in that, The dual-layer Walker constellation structure uses the Geometric Diluted Precision (GDOP) index for geometric target layout. The expression for the geometric dilution precision (GDOP) index is: ； in, , , To detect the components of geometric errors in three dimensions; To satisfy the target GDOP constraint, we have: ； in, For the duration of observation, The angle between the stations; The above GDOP constraints are extended to multi-level orbital systems, taking into account the joint influence of the overall constellation geometry on GDOP; For a two-layer Walker constellation structure, the joint constraints satisfying the GDOP optimality condition are expressed as: ； in, Indicates the first Number of probe units in each orbital layer This indicates the inclination angle of the track at that level.

4. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 1, characterized in that, In S2, the spatiotemporal registration of multi-source observation data using linear interpolation and geographic coordinate transformation to obtain observation trajectory data specifically includes mapping multi-source observation data to observation trajectory data. The mapping relationship is as follows: ； in, This represents the original observation coordinates and timestamps, i.e., multi-source observation data. To unify the spatiotemporal alignment results, i.e., the observation trajectory data, under a unified coordinate system, express.

5. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 1, characterized in that, The specific content of feature extraction and fusion of observed trajectory data in S3 to obtain fused features includes: For each observation trajectory segment in the observation trajectory data A multimodal feature encoder is used to extract its temporal, velocity, and heading information to obtain multimodal embedding features; The expression for multimodal embedding features is: ； in, For trajectory time series encoder, For the speed (SOG) feature extractor, For heading COG feature encoder; The multimodal embedding features are jointly obtained by using a temporal fusion transformer and a spatial fusion transformer: ； The TFT module captures the temporal dependencies between multiple frames through a self-attention mechanism, while the SFT module is based on the spatial proximity graph between targets. Aggregate local spatial context to achieve multi-objective interactive modeling.

6. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 5, characterized in that, The spatial fusion transformer constructs an adjacency matrix based on a k-nearest neighbor graph. And a self-attention mechanism is used to achieve spatial feature aggregation: ； in, For learnable matrices, For feature dimensions; The time fusion transformer can employ a long sequence memory mechanism to fuse and encode the trajectory history of multiple time steps, thereby enhancing the dynamic prediction capability for complex maritime maneuvering targets.

7. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 1, characterized in that, The specific content of S4, which uses a nonlinear regression function based on fused features to predict and update the trajectory of maritime targets to obtain multi-source tracking trajectories, includes: By fusing features, a nonlinear regression function is used to predict future trajectory points; The prediction error is optimized by minimizing the L2 norm loss function to obtain the multi-source tracking track; The expression for minimizing the L2 norm loss function is: ； in, For the number of observed trajectories, The time window length, These are the actual trajectory points.

8. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 1, characterized in that, The time matching submodule described in S5 defines a matching score based on the Euclidean distance between the predicted trajectory and the current observation point, expressed as: ； in, For the first The predicted trajectory point and the first Matching score between observation points For the target predicted by the model at time... The position vector, For the sensor at any time The obtained target observation position vector; Define time-related thresholds ,when Time, trajectory and Continuous in time; The spatial matching submodule is used to calculate the similarity between trajectory embedding features and normalize it into an association probability, expressed as follows: ； in, Temperature is a parameter used to adjust the smoothness of similarity. At that time, determine the trajectory and Belonging to the same goal; The joint optimization submodule combines the temporal and spatial matching matrices to form a global trajectory association map across sensors.

9. The method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data according to claim 6, characterized in that, The expression for the information contrast loss function is: ； The expression for the joint optimization trajectory is: ； in These are weighting parameters used to balance trajectory prediction accuracy and correlation accuracy; The information comparison loss This is achieved through the embedding and comparison of positive and negative sample trajectories.

10. A system for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data, used to implement the method for continuous tracking and trajectory association of maritime targets based on multi-source heterogeneous data as described in any one of claims 1-9, characterized in that, include: Multi-source observation unit: Construct a multi-source observation model and use the multi-source observation model to observe maritime targets to obtain multi-source observation data; Spatiotemporal registration unit: Based on the differences in time synchronization, resolution and coordinate reference system of multi-source observation models, observation trajectory data is obtained by spatiotemporal registration of multi-source observation data through linear interpolation and geographic coordinate transformation; Feature fusion unit: Extracts and fuses features from observed trajectory data to obtain fused features; Track prediction unit: Based on fused features, a nonlinear regression function is used to predict and update the track of maritime targets to obtain multi-source tracking tracks; Spatiotemporal Joint Association Unit: Construct a multi-source spatiotemporal joint association module, and use the multi-source spatiotemporal joint association module to associate multi-source tracking tracks to obtain tracking associated tracks; The multi-source spatiotemporal joint correlation module includes a time matching submodule, a spatial matching submodule, and a joint optimization submodule; Track optimization unit: Based on the tracking and associated track, information comparison loss function is introduced to obtain the target joint optimized track.

Images