Target tracking and trend prediction method and system based on open and closed source information fusion

Abstract

A method and system for target tracking and motion prediction based on open-source and closed-source information fusion is disclosed. The method includes: acquiring closed-source observation data and open-source information of the target; generating multiple fuzzy observations based on the open-source information and filtering the fuzzy observations according to the target's maximum speed; calculating the initial weights of each fuzzy observation and fusing the initial weights using a conflict confidence allocation rule to obtain the final weights of each fuzzy observation; combining the closed-source observations and the filtered fuzzy observations in chronological order to generate at least one possible trajectory; predicting the target state based on a dynamic model including the state of the target's destination, and updating the weights of the possible trajectories based on the final weights and the corresponding covariance matrix; calculating the posterior probability of the target going to each destination based on the updated trajectory weights to obtain the motion prediction result, and outputting a confidence index to characterize the reliability of the prediction based on the uncertainty metric of the state estimation.

Description

Technical Field

[0001] This invention relates to the field of target tracking and information fusion technology, and in particular to a target tracking and movement prediction method and system based on open and closed source information fusion. Background Technology

[0002] In the field of intelligent shipping, target tracking and movement prediction are core technological supports for ensuring scientific decision-making and effective execution. Currently, information can be categorized into closed-source and open-source information: closed-source information mainly refers to data collected by physical sensors such as Automatic Identification System (AIS) data and radar detection data, characterized by high structure and controllable accuracy, but limited coverage and high deployment costs; open-source information mainly refers to unstructured data such as online text, social media posts, and public event reports, with advantages of low acquisition costs and wide coverage, but suffers from problems such as vague descriptions, fluctuating reliability, and difficulty in quantifying uncertainty. Integrating open-source and closed-source information has become an important direction for improving tracking and prediction performance.

[0003] Existing technologies have yielded a series of research results in information fusion, open-source information processing, and target state modeling. In information fusion theory, Dempster-Shafer theory (DST) handles incomplete information through basic confidence assignment, but it is prone to distortion of fusion results when faced with highly conflicting evidence. Dezert-Smarandache theory (DSmT), as an extension of DST, allows focal element overlap, enhancing conflict handling capabilities, but its computational complexity surges under high-dimensional discrimination frameworks. In open-source information processing, existing methods mostly extract key information through natural language processing techniques, but they struggle to transform ambiguous expressions (such as "activity in a certain area" or "recent arrival") into quantitative forms that can be directly fused with closed-source information, and lack systematic methods for characterizing the reliability of open-source information. In target tracking and state modeling, traditional methods such as Kalman filtering are suitable for handling the probabilistic uncertainty of closed-source information, but they are difficult to adapt to the non-probabilistic reliability characteristics of open-source information. While stochastic differential equation models based on equilibrium recovery velocity (ERV) can characterize the mean recovery characteristics of target motion, they do not fully incorporate prior information such as destination and arrival time, leading to insufficient long-term prediction accuracy.

[0004] Despite the progress made by existing technologies in their respective sub-fields, the following key shortcomings still exist in the open-source and closed-source information fusion scenarios for tracking and trend prediction: First, the quantitative representation of the reliability of open-source information is insufficient. Existing technologies mostly treat open-source information as deterministic or simple probabilistic observations, failing to effectively handle its ambiguous representations. Furthermore, there is a lack of dynamic reliability assessment methods that combine the consistency of historical facts with feedback on the fusion effect, making it difficult to accurately quantify the uncertainty of open-source information and affecting the accuracy of fusion.

[0005] Second, the heterogeneous information fusion mechanism is imperfect. The uncertainty of closed-source information manifests as a probability distribution, while the uncertainty of open-source information manifests as reliability fluctuations. Existing technologies often use a single fusion rule to handle the two types of heterogeneous uncertainty, failing to fully consider the essential differences between them. When faced with information conflicts, the fusion results are prone to bias, and there is a lack of optimization for conflict resolution in the fusion process.

[0006] Third, there is a lack of movement prediction and confidence assessment. Existing target tracking models mostly focus on state estimation, ignoring key movement information such as the target's destination and arrival time. Furthermore, the fusion results lack interpretable confidence metrics, failing to provide a reference for the reliability of the fusion for decision-making, making it difficult to assess decision risks.

[0007] Fourth, there is a lack of interpretability and engineering adaptability. Existing fusion algorithms mostly rely on black-box models or complex mathematical derivations, making it difficult to trace the calculation process from information input to state output. Furthermore, they lack sufficient engineering processing for generating fuzzy observations and supplementing covariance of open-source information, making it difficult to directly adapt to the data formats and processing requirements in real-world scenarios.

[0008] Therefore, there is an urgent need to propose a technical solution that can accurately quantify the reliability of open source information, efficiently integrate heterogeneous information from open and closed sources, and take into account both trend prediction and confidence assessment, so as to solve the problems of inconsistent uncertainty quantification, poor resolution of fusion conflicts, lack of key constraints in prediction, and poor interpretability of confidence in existing technologies. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a target tracking and movement prediction method and system based on open and closed source information fusion, which realizes the quantification of open source information reliability and confidence assessment in target tracking and movement prediction.

[0010] To achieve the above objectives, this invention provides a target tracking and movement prediction method based on open and closed source information fusion, comprising: Obtain closed-source observation data and open-source information of the target; Multiple fuzzy observations are generated based on the open-source information, and the fuzzy observations are filtered according to the target's maximum speed. The initial weights of each fuzzy observation are calculated based on multiple benchmarks, and the initial weights are fused using a conflict confidence allocation rule to obtain the final weights of each fuzzy observation. The closed-source observations in the closed-source observation data are combined with the filtered fuzzy observations in chronological order to generate at least one possible track. The target state is predicted based on a dynamic model that includes the destination state of the target, and the weights of the possible trajectories are updated by combining the final weights and the corresponding covariance matrix. Based on the updated track weights, the posterior probability of the target going to each destination is calculated to obtain the motion prediction result, and the confidence index is output based on the uncertainty measure of state estimation to characterize the reliability of the prediction.

[0011] Furthermore, the step of generating multiple fuzzy observations based on the open-source information and filtering the fuzzy observations according to the target's maximum speed further includes: Extract time information, location information, and speed and heading information from the open-source information; At least one observation time point is generated based on the time information, center coordinates are generated based on the location information, and multiple candidate points are generated based on the center coordinates. Based on the speed and heading information, the multiple candidate points are assigned speed and heading values ​​to form fuzzy observation values; Based on the most recent closed-source observation of the target prior to the open-source information, the reachable distance is calculated according to the target's maximum speed, and ambiguous observations whose distance from the most recent closed-source observation exceeds the reachable distance are filtered out.

[0012] Furthermore, the step of calculating the initial weights of each fuzzy observation based on multiple benchmarks and fusing the initial weights using a conflict confidence allocation rule further includes: Based on historical track benchmarks and heading consistency benchmarks, the initial weights of each fuzzy observation are calculated. By adopting the proportional conflict redistribution rule, the initial weights of the same fuzzy observation obtained from different benchmarks are used as evidence to fuse them and obtain the final weight of the fuzzy observation.

[0013] Furthermore, it also includes the steps of establishing a Beta prior distribution based on the historical factual consistency data of the open source information, performing a Bayesian update on the Beta prior distribution according to the reliability value obtained from this fusion evaluation, and using the expected value of the updated distribution as the dynamic reliability value of the open source information and outputting it.

[0014] Furthermore, the reliability value obtained from this fusion assessment is generated through the following steps: Using the target location in the closed-source observation data as the true location reference, calculate the absolute error between the fused target location and the true location; A confidence interval with a pre-set confidence level is constructed based on the uncertainty measure of state estimation, and it is determined whether the true location falls within the confidence interval; Based on the statistical percentage of the absolute error and the statistical percentage of the true location falling within the confidence interval, the evaluation level of this fusion is determined by referring to the preset evaluation level classification standard. The reliability value is obtained based on the preset mapping relationship between the evaluation level and the reliability value.

[0015] Furthermore, the step of calculating the posterior probability of the target going to each destination based on the updated track weights to obtain the motion prediction result further includes: Define the prior distributions of the destination set and arrival time; The likelihood probability under a given destination is calculated using numerical integration. By combining the likelihood probability and the prior probability of the destination, the posterior probability of each destination is calculated according to Bayes' theorem, and the optimal destination and its corresponding arrival time are selected.

[0016] Furthermore, the step of using the uncertainty metric output based on state estimation to characterize the confidence index of prediction reliability further includes: The uncertainty measure of the state estimate is decomposed into eigenvalues, and a confidence ellipse is constructed based on the obtained eigenvalues. The statistical proportion of the target's true location falling into the confidence ellipse is used as the confidence index.

[0017] To achieve the above objectives, the present invention also provides a target tracking and movement prediction system based on open and closed source information fusion, comprising: An open-source information processing module is used to acquire and process open-source information, and generate and filter fuzzy observation values. An open-source information reliability quantification module is used to calculate the initial weights of each fuzzy observation based on multiple benchmarks, and to fuse the initial weights using a conflict confidence allocation rule to obtain the final weights. The open-source and closed-source information fusion module is used to combine closed-source observations and filtered fuzzy observations in chronological order to generate possible tracks, and to update the weights of the possible tracks based on the dynamic model containing the destination state of the target and the final weights. The motion prediction and confidence assessment module is used to calculate the posterior probability of the target going to each destination based on the updated track weights to obtain the motion prediction result, and output a confidence index to characterize the reliability of the prediction based on the uncertainty measure of the state prediction.

[0018] To achieve the above objectives, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor is configured to execute the computer program stored in the memory to implement the target tracking and movement prediction method based on open and closed source information fusion as described above.

[0019] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the target tracking and movement prediction method based on open and closed source information fusion as described above.

[0020] The target tracking and trajectory prediction method based on open and closed source information fusion provided by this invention generates and fuses the weights of open source fuzzy observations, and updates the track weights by combining a dynamic model that includes the destination state. While achieving target tracking and trajectory prediction, it also realizes the quantification of the reliability of open source information and the interpretable confidence assessment.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a target tracking and motion prediction method based on open and closed source information fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the target tracking and motion prediction system based on open and closed source information fusion according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0023] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0025] The term "comprising" and its variations as used in this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0026] It should be noted that the concepts of "first" and "second" may be mentioned in this invention only to distinguish different devices, components or parts, and are not used to limit the order of the functions performed by these devices, components or parts or their interdependence.

[0027] It should be noted that the terms "one" and "multiple" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless explicitly stated otherwise in the context, they should be understood as "one or more". "Multiple" should be understood as two or more.

[0028] In this invention, the numbering of each step is only used to distinguish different operational stages and is not a strict limitation on the execution order. Without departing from the technical concept of this invention, some steps can be executed in parallel, their order can be changed, or they can be implemented in other equivalent ways.

[0029] In this invention: Closed-source information refers to structured data collected by physical sensors such as AIS (Automatic Identification System) data and radar detection data, which is characterized by controllable accuracy and high sampling frequency.

[0030] Open source information refers to unstructured data such as online text, social media posts, and public event reports. It is characterized by low acquisition cost, wide coverage, but vague description.

[0031] Fuzzy observations: refers to structured observation data containing time, speed, heading, and position generated after quantizing open-source information.

[0032] Reliability quantification refers to the process of numerically representing the credibility of open-source information through multi-benchmark weight fusion and dynamic Beta distribution updates.

[0033] Dempster-Shafer theory (DST) is a mathematical framework for dealing with uncertain information, also known as evidence theory. Dezert-Smarandache theory (DSmT) is an extended version of Dempster-Shafer theory (DST).

[0034] PCR6: Proportional Conflict Reassignment Rule Version 6, for multi-source conflict reliability assignment within the DSmT framework.

[0035] Confidence Ellipse: An elliptical region constructed based on an uncertainty metric, used to characterize the range of uncertainty in the target location estimation.

[0036] Definitions for other terms will be provided in the following description.

[0037] Example 1 Figure 1 The flowchart below illustrates the target tracking and motion prediction method based on open and closed source information fusion according to an embodiment of the present invention. Figure 1 The embodiments of the present invention will be described in further detail.

[0038] First, in step 101, acquire closed-source and open-source data: acquire closed-source observation data (such as observation data from AIS and radar) and open-source information of the target to be processed (such as moving targets such as ships, aircraft, or mobile equipment), and determine calibration parameters (such as maximum speed). Time interval (e.g., fuzzy point radius r, DSmT fusion rules, evaluation level threshold, etc.)

[0039] In this embodiment, the target to be processed is a ship. Publicly available shipping information websites are used as the open-source information source. Open-source text data of ship events (including hull number, event time, geographical location, speed and heading description, etc.) are collected via web crawler. Closed-source information uses AIS (Automatic Identification System) data, with fields including identification number (mmsi), speed (sog), course (cog), location coordinates (latitude and longitude), and timestamp (lasttime). The data sampling frequency is once every 10 minutes. The preset calibration parameter is: maximum speed. =30km / h, time interval =3600s (1 hour), fuzzy point generation radius r=5km (kilometers), DSmT fusion rule adopts proportional conflict redistribution rule (PCR6). Here, r=5km is to match the ship's motion capability and accessibility filtering rules. This embodiment presets the ship's maximum speed to 30km / h, the time processing granularity of open source information to 1 hour, the ship's maximum sailing distance in 1 hour to 30km, the sampling frequency of closed source AIS data to approximately 10 minutes / time, and the ship's maximum sailing distance in 10 minutes to 5km. This radius is aligned with the spatiotemporal sampling granularity of AIS data, facilitating subsequent time synchronization, position matching, and fusion calculation of open and closed source information.

[0040] In addition, the hardware environment used in this embodiment is an Intel Core i9-13900K processor, 64GB of memory, and an NVIDIA RTX 4090 graphics card. The software environment is Python 3.9 and PyTorch 2.0. The large model used is Tongyi Qianwen Qwen-72B-Chat, and the coordinate parsing API is the Gaode Map Open Platform API.

[0041] In step 102, the open-source information is processed and fuzzy observations are generated: key fields of the open-source information are extracted through a large model and transformed into structured data; fuzzy observations (including time, speed, heading, and position) are generated and filtered by reachability to obtain a set of valid open-source fuzzy observations.

[0042] In this embodiment, open source information processing includes three core steps: open source information extraction, fuzzy observation generation, and reachability filtering. These three steps are described below.

[0043] First, open-source information extraction is performed. For the open-source information crawled in step 101, a Prompt instruction (a natural language instruction written to help large AI models understand the task) is constructed. For example, "Please determine if the text provided by the user contains any factual rather than hypothetical paragraphs regarding the location of cargo ship A at a certain time. If the text mentions relative times such as holidays, these can be converted to absolute times using reasoning. For example, if the time is mentioned as occurring in October 2003 without specifying the exact day, then event_start_time defaults to the first day of October, and event_end_time defaults to the last day of October. event_address should describe a location; if multiple locations are mentioned, only the first one is taken. If a relatively vague location is mentioned, it needs to be converted to an absolutely clear location; otherwise, leave it blank. If there is an extraction result, please return the extraction result in the specified JSON format; otherwise, return an empty array []." After extraction using an information extraction model (such as Qwen-72B-Chat), structured data is obtained: event_ship="Cargo Ship A", event_start_time="2003-10-01", event_end_time="2003-10-31", event_address="Pacific Ocean". Then, the Gaode Map API (Application Programming Interface) is called to convert the geographic location into latitude and longitude.

[0044] Regarding time processing: if the open-source information only provides "a certain day", then the starting point is 0:00 of that day, and so on. =Generate time points at equal intervals of 3600s (1 hour), the formula is: ( (i represents the start time of the day, where i is the time index). If a "time period" is given, the center time of that time period is used. Based on, according to Generate time points, cover interval, Indicates from the center time The length of time traced back, Indicates from the center time The length of time traced backwards. For speed and heading processing: speed is described as "low speed / normal speed / high speed," with ranges divided according to the target's maximum speed (low speed ≤ 1 / 3). Constant speed ∈ (1 / 3) 2 / 3 High speed > 2 / 3 The values ​​within the range are randomly selected; the heading is described as "North / Northeast" and other 8 directions, which are converted into specific angles such as 0° and 45° (North=0°, Northeast=45°, East=90°, etc.). For position processing: based on the coordinates of the region center ( Using as a reference, eight fuzzy observation points are generated with a radius r. The coordinate calculation formula is as follows: East:( +r, );South:( , -r); West: ( -r, );north:( , +r); Southeast: ( + r, - r); Southwest: ( - r, - r); Northeast: ( + r, + r); Northwest: ( - r, + r); Iterative diffusion can generate more points, and a deduplication operation is performed after generation.

[0045] In the fuzzy observation generation step: time processing is adopted =3600s, with 0:00 on October 1, 2003 as the starting point. Generation time point This covers all hourly times throughout October; in speed processing, the ship's maximum speed is known. =30km / h. Since the open-source information does not specify the speed, 15km / h was randomly selected from the normal speed range (e.g., 10km / h~20km / h) as the speed value. The heading description is not specified, so 90° (eastward) was randomly selected as the heading. Position processing uses latitude and longitude (0.0°, -160.0°) as the center, and r=5km to generate 8 fuzzy observation points in each direction. For example, the coordinates of the eastward point are (0.0° + 5km corresponding latitude increment, -160.0°), and the coordinates of the northeastward point are (0.0° + 22×5km corresponding latitude increment, -16 ... (×5km corresponds to longitude increment), and after generation, the duplicate coordinate points are removed by performing a deduplication operation using Python's pandas library.

[0046] Reachability filtering is based on the most recent AIS closed-source observation (specific observation data tied to the closed-source observation time, such as position coordinates) prior to this open-source information, and is calculated according to the target's maximum speed. The formula for determining reachability is as follows: ( For fuzzy observation time, (the most recent closed-source observation time), only retaining d≤ (d is the distance between two points) fuzzy observation values ​​(fuzzy observation data bound to the fuzzy observation time). Assume the most recent closed-source observation time... =2003-09-30-18:00 (meaning 18:00 on September 30, 2003), fuzzy observation time =2003-10-01-00:00, their time difference is 6 hours, based on this, the maximum reachable distance can be calculated. The distance between the closed-source observation point (referring to the location information in the observation data, assumed to be 30.0°N, -150.0°W) and each fuzzy observation point was calculated using the Haversine formula (a mathematical formula used to calculate the shortest distance between two points on the Earth's surface). Only points with a distance less than or equal to 180 km were retained, and finally, 6 valid fuzzy observation values ​​were obtained.

[0047] In step 103, the reliability of open source information is quantified: the weights of open source observations are initialized, and the conflict weights are fused using DSmT; the covariance matrix is ​​calculated; and the reliability values ​​are dynamically updated by combining historical data with the fusion evaluation results.

[0048] This step, based on open-source fuzzy observations, historical factual data, and closed-source observation data, quantifies the reliability and uncertainty of open-source information, obtaining a weighted and covariance-weighted quantitative characterization of open-source information and a dynamic reliability value for the open-source information source. Specific steps include fuzzy observation weight initialization, random set theory, DSmT fusion, and covariance calculation. Random set theory studies finite sets with randomly occurring elements and an indefinite number of elements (such as multiple fuzzy locations or multiple candidate tracks), using a unified description of such fuzzy, conflicting, and multi-possible uncertainties using sets and probability. This invention employs random set theory to transform fuzzy open-source information, multiple candidate tracks, and uncertain destinations into computable, fusionable, and predictable random sets, thereby solving the fuzziness and conflict problems that traditional methods cannot handle.

[0049] The initialization of fuzzy observation weights employs a multi-reference point weighting strategy to ensure comprehensive reliability quantification. In this embodiment, historical track references and heading consistency references are used. The historical track reference uses fuzzy observations as the destination, and the probability of destination is calculated using Bayesian inference as the initial weight. The formula for calculating the probability of destination is as follows: The heading consistency benchmark is calculated by taking the azimuth angle from the nearest closed-source observation point to the fuzzy observation point. Difference between the open source heading transformation angle and the open source heading transformation angle The weight formula is u=+1, which is normalized and used as the initial weight.

[0050] To address the conflict of multiple benchmark weights, the DSmT weight fusion process employs the proportional conflict redistribution rule (PCR6) for fusion, as shown in the formula: in, It represents the initial confidence level of closed-source evidence sources regarding the target state X. It represents the initial level of trust that the open-source evidence source has in the conflicting state Y. It represents the initial level of confidence of closed-source evidence sources in the conflicting state Y. Y represents the initial trust level of the open-source evidence source towards the target state X, where X is the target proposition or state in the identification framework, i.e., the object whose final trust level we are currently calculating, and Y is a conflicting proposition or state in the identification framework that is mutually exclusive with X. For identification framework The superpower set contains all possible combinations of propositions (intersection, union).

[0051] The results are classical DSmT combination results, and the final fuzzy observation weights are obtained after fusion. .

[0052] The steps for calculating covariance are based on weights Classify into levels and assign polar coordinate covariance matrices: in, The distance between the closed source and the fuzzy observation point. The azimuth angle is transformed into a Cartesian coordinate system covariance matrix through eigenvalue decomposition. ,in , j represents the index of the j-th open-source fuzzy observation (or the j-th fuzzy observation point) (used to distinguish different individual fuzzy observations).

[0053] The dynamic reliability assessment process combines historical factual consistency with fusion assessment results, modeling reliability based on the Beta distribution. Its core logic lies in the fact that the Beta distribution is the conjugate prior distribution of the binomial distribution, enabling convenient updates of the post-hoc probability based on newly added fusion assessment results (reliability value c), thus achieving dynamic iteration of open-source information source reliability. First, the prior distribution parameters are calculated based on historical factual consistency data of the open-source information: the total number n of target-related reports historically published by the open-source information source is counted, and the number of reports consistent with historical facts is x, thus calculating the historical consistency rate. This refers to the prior probability that the open-source information description matches the facts; combined with the expectation and variance formulas of the Beta distribution, (p represents the historical consistency rate) The shape parameters of the prior distribution are obtained by solving: , Subsequently, the reliability value 'c' obtained from this fusion evaluation is used as a newly added valid observation sample, and the prior distribution parameters are updated using Bayesian methods: Since 'c' represents the credibility of this fusion result, it is equivalent to adding a new valid experiment with a "consistency level of c". Therefore, the shape parameter of the posterior distribution is updated as follows: , The final reliability value is taken as the expectation of the posterior Beta distribution. (in = +c, = +1-c, where c is the reliability value of the fusion assessment).

[0054] In this embodiment of the invention, the generation of the fusion evaluation reliability value c is the core link connecting the fusion results of open and closed source information with the dynamic reliability evaluation of open source information. Its calculation process is embedded in the fusion process, realizing a closed-loop feedback from the fusion effect to the reliability of the information source. First, the construction of the fusion evaluation benchmark data is completed. The calculation of c depends on the quantitative evaluation of the fusion results, and the benchmark for evaluation is the target's true location data. In this embodiment, the closed source information is based on AIS (Automatic Identification System) data, whose sampling frequency and positioning accuracy are much higher than open source text information. Therefore, the target's latitude and longitude coordinates provided by AIS data are used as the benchmark value for the target's true location. For any target state estimation or prediction time t that needs to be evaluated, if there is direct AIS sampling data at that time, the latitude and longitude of the sampling point are directly taken as the true location; if there is no direct AIS sampling data at that time, the true location at that time is calculated by linear interpolation. Specifically, the target's true latitude and longitude coordinates at time t are obtained by linear interpolation according to the time proportion of the timestamps and latitude and longitude coordinates of the two AIS sampling points closest to time t. This is used as a unified benchmark for the evaluation of the fusion results, ensuring the objectivity and repeatability of the evaluation results.

[0055] Based on this, the core evaluation indicators of the fusion output are calculated. After the open-source and closed-source information fusion module completes weight updates and state estimation, it outputs two core results: the mean position of the target state estimation or prediction, and the covariance matrix of the corresponding state. These two results are the core inputs for calculating the evaluation indicators. First, the first type of indicator is calculated: the percentage of absolute error. For the state estimation or prediction results at all times in this fusion process, the spherical distance between the mean position of the fusion output and the actual position at the corresponding time is calculated point by point using the Haversine formula, i.e., the absolute error. Then, the proportion of results with an absolute error of less than 5km and the proportion of results with an absolute error of less than 10km are statistically analyzed to form a quantitative indicator for the absolute error dimension. Next, the second type of indicator is calculated: the percentage of confidence interval coverage. This invention uses a 95% confidence level by default. Based on the covariance matrix of the fusion output, eigenvalue decomposition is performed to obtain the eigenvalues ​​of the longitude and latitude dimensions. , Based on this, a 95% confidence ellipse is constructed for the target location. The equation of the ellipse is: This critical value is derived from the 95th quantile of a chi-square distribution with 2 degrees of freedom. Point-by-point verification is performed to determine whether the target's true location falls within the confidence ellipse at the corresponding time. The proportion of results where the true location falls within the confidence ellipse is calculated to form a quantitative indicator of the confidence level. Here, x represents the deviation of the target location from the mean of the estimated location in the longitude dimension, and y represents the deviation of the target location from the mean of the estimated location in the latitude dimension.

[0056] After calculating the indicators, the reliability value 'c' of the fusion assessment is obtained through level mapping. Based on the above two core indicators, and in accordance with the five-level assessment level classification standard preset in this invention, the level of this fusion result is determined. Specifically, a Level 1 assessment level requires: ≥80% of results with an absolute error less than 5km, and ≥60% of results whose true location falls within the confidence interval; a Level 2 assessment level requires: ≥60% of results with an absolute error less than 5km and a confidence interval coverage rate of ≥60%, or ≥80% of results with an absolute error less than 10km and a confidence interval coverage rate of ≥80%; a Level 3 assessment level requires: ≥80% of results with an absolute error less than 10km and a confidence interval coverage rate of ≥60%; a Level 4 assessment level requires: ≥60% of results with an absolute error less than 10km and a confidence interval coverage rate of ≥80%; and a Level 5 assessment level requires: ≥60% of results with an absolute error less than 10km and a confidence interval coverage rate of ≥60%. After the level determination is completed, the reliability value c corresponding to this fusion assessment is obtained according to the preset mapping relationship between the assessment level and the reliability value: Level 1 corresponds to c=0.95, Level 2 to c=0.85, Level 3 to c=0.75, Level 4 to c=0.40, and Level 5 to c=0.10. This value c fully quantifies the validity of the result of this open-source and closed-source information fusion process. It reflects both the accuracy of the target state estimation after the open-source information is fused and, conversely, characterizes the credibility level of the open-source information itself, becoming the core bridge connecting the fusion process and the dynamic reliability assessment of open-source information.

[0057] For example, this embodiment focuses on four steps: weight initialization, DSmT fusion, covariance calculation, and dynamic reliability assessment. Each step is described in detail with specific numerical calculations and rule applications.

[0058] The open-source observation weight initialization adopts a dual-benchmark strategy: in the historical track benchmark, six valid fuzzy observations are used as possible destinations D={d1,d2,d3,d4,d5,d6}. The probability of destination is calculated through Bayesian inference. It is assumed that the prior probability of each destination in the historical data is p(D=d)=0.1667, and the likelihood probability p( |D=d) is obtained by matching the ship's historical tracks. For example, the likelihood probability of d1 is 0.25, and the others are 0.15. Substituting these values ​​into the formula... The initial weight of d1 was calculated to be 0.227, and the weights of the other points ranged from 0.157 to 0.168; in the heading consistency datum, the azimuth angle from the nearest closed-source observation point to d1 was... =85°, Open source heading conversion angle =90°, difference =5°, substitute into the weight formula We get u = 0.996 + 1 = 1.996. After normalization, the baseline weight of d1 is 0.182, and the weights of the other points are calculated based on the angle difference and are between 0.15 and 0.17.

[0059] The DSmT weighted fusion adopts the PCR6 rule, using the dual-benchmark weights as two evidence sources m1 and m2. For example, m1=0.227, m2=0.182 for d1, and m1=0.168, m2=0.170 for d2. First, the classic DSm combination result m12(X) is calculated, and then conflict confidence is assigned to the conflicting focal elements. For the conflicting parts of d1 and d2, the calculation... The conflict confidence was assigned to d1 and d2, and the final fusion weights of d1, o1, and d2 were 0.215 and 0.175, respectively. The weights of the remaining points were between 0.15 and 0.16 after fusion.

[0060] Covariance calculation: Assign covariance to the top 33% of points (o≥0.18) using the following formula: var( )=5、var( )=10; 33%~66% of the points are allocated var( =10、var( =15; Point allocation after 66% var( =15、var( =20. Taking d1 as an example, the weight 0.215 belongs to the top 33%, and the polar coordinate covariance is... azimuth =85°, orthogonal matrix The eigenvalue decomposition transforms the matrix into a Cartesian coordinate system covariance matrix. Calculations yielded .

[0061] Here, var( ) represents the variance of the distance error between the target location described in the open-source information and the actual location. The larger the value, the more ambiguous the description of the target distance in the open-source information; var( This represents the angular error variance (in degrees) between the target orientation described in the open-source information and the actual orientation. 2 The larger the value, the more ambiguous the description of the target's heading and / or bearing in the open-source information. Taking "Level 1 Assessment (Absolute error < 5km percentage ≥ 80%, confidence interval coverage percentage ≥ 60%)" as an example, when the var(...) of the high-weight position... )=5、var( When var( )=10, the confidence ellipse axis of the fused output is approximately 10 to 12 kilometers, which can cover most of the real locations without causing the confidence level to fail due to excessively large intervals; if var( Adjusting the value from 5 to 3 will result in an excessively small confidence ellipse and a true location coverage of less than 60%. If the value is adjusted to 8, the confidence ellipse will be too large, and the percentage of locations with an absolute error of less than 5 km will be difficult to meet the target. The same applies to the other values. Finally, the optimal value was determined.

[0062] In the dynamic reliability assessment step, based on historical factual consistency data, the open-source information source had a total of 8 historical reports (n=8), of which 5 were consistent with the facts, with a historical consistency rate of p=0.625. Substituting this into the Beta distribution formula... Solving for r1 and s1, we get r1 = 5.25 and s1 = 3.15. Combining this with the reliability value c = 0.8 (corresponding to a level 2 assessment), we update r2 = r1 + c = 6.05 and s2 = s1 + 1 - c = 3.35, resulting in the final reliability... Wherein, μ is the unknown parameter to be modeled in this invention, representing the inherent probability that the target information described by a certain open source (such as a specific website or a specific publishing entity) is consistent with the facts, and its value range is [0,1]; E(μ) is the mathematical expectation of μ, representing the best single-point guess of the reliability of the open source based on all current information; Var(μ) is the variance of μ, which measures the confidence level of the guess of the true value of μ. The smaller its value, the more certain and confident the estimate of reliability is.

[0063] In step 104, possible track generation: combine open and closed source observations in chronological order to generate all possible tracks, and assign weights and covariances to the observations associated with each track.

[0064] In this embodiment, the possible track generation may be a set of closed-source observations arranged in chronological order. (where n is the total number of observations in the closed-source observation set) and the open-source fuzzy observation set (k is the total number of observations in the open-source fuzzy observation set), by arranging and combining the elements in the open-source set, generate... Possible flight path ( Let be the number of observations in the j-th open source set (j=1,...,k) and fused open and closed source observations in chronological order for each track.

[0065] In step 105, open-source and closed-source fusion calculation: based on the dynamic model and the observation model, the target state is predicted by Kalman filtering; the track weights are updated by combining the observation weights, and high-weight tracks are retained.

[0066] In embodiments of the present invention, the equilibrium recovery velocity (ERV) model is used as the dynamic model, and the state equation is: ,in , This represents the state vector at time t (including the current state and the destination state). , Let be the position components of the target in a two-dimensional plane (such as latitude and longitude) at the current time t. , Let be the velocity component of the target at the current time t. , For the location component of the destination, , Let F be the velocity component from the target to the location, F be the state transition matrix (the transition coefficient from velocity to position in the off-diagonal elements is k (time step)), and M be the offset vector. For process noise; the observation model is , where G is the observation matrix (G=[ ,0], (For closed-source observation matrix) Indicates time The observation value at the specific time point corresponding to the nth observation value. To observe noise.

[0067] The steps for predicting the target state using Kalman filtering include: Possible trajectory weight prediction: Based on the state equation, the target state focus set weights ,in , (This formula is the standard form for further prediction of covariance using the state equation in Kalman filtering), Q is the process noise covariance matrix, F (state transition matrix, Its transpose matrix comes from the dynamic model. Let represent the state covariance matrix of the i-th focus set at the current time (t). Let represent the prediction covariance matrix of the i-th focus set after state prediction (i.e., at time t+1). The unnormalized weights of the focus set of the i-th target state at the predicted next time step (t+1), This represents the prediction weight factor for the i-th focus set. denoted by , det represents the weight of the focus set of the i-th target state at time t, and det represents the determinant of the matrix. Meaning and They are the same; the different symbols used are only for differentiation. and Similarly.

[0068] Possible track weight updates: combining open-source fuzzy observation weights Update the weights of possible tracks as follows: in, , ( (for Kalman filter gain) The weight of the i-th possible track (or the i-th target state focus set) before fusing the current observations (i.e., the weight after the previous update or prediction). The weight of the j-th open-source fuzzy observation (derived from the above open-source information reliability quantification step, characterizing the reliability of the observation). Indicates a Gaussian distribution. This represents the updated weight (i.e., the new confidence level of the track after combining the new observation) obtained by fusing the i-th track with the j-th observation. Let be the covariance matrix of the Cartesian coordinate system. In the formula... , , respectively with , , They have the same meaning, and are only used as a distinguishing feature in summation.

[0069] In some exemplary embodiments, steps 104 to 105 are performed in the order of possible trajectory generation, dynamic model construction, Kalman filter optimization, weight update, and prediction, combining the ERV model and random set theory to complete the fusion calculation. For example: In the possible trajectory generation process, the set of closed-source observations Arranged chronologically as {dc1(2003-09-30-18:00), dc2(2003-10-01-06:00), dc3(2003-10-01-12:00)}, this is an open-source fuzzy observation set. ={do1(d1~d6, 2003-10-01-00:00)}, generates 6 possible tracks in chronological order. For example, track 1 is [dc1, do1(d1), dc2, dc3]. Each track is associated with a corresponding weight and covariance.

[0070] The dynamic model adopts the ERV model, with state vectors. ,in The current latitude and longitude. For the current velocity component, The destination's latitude and longitude. This represents the velocity component heading towards the ground. State equations In this matrix, F is the state transition matrix, set as an 8×8 matrix according to the ship's motion characteristics. The velocity-to-position transition coefficient in the off-diagonal elements is k (time step). Other correlation coefficients (such as position autoregression, velocity autoregression, coupling between current state and destination state, etc.) are set to 0.95 based on the mean recovery characteristic. M is the offset vector, with a value of... ; For process noise, the process noise covariance matrix is ​​set as a diagonal matrix with all diagonal elements being 0.01. Observation model In the middle, G=[ , 0], It is an identity matrix (only observing the current position and velocity). To observe the noise, the observation noise covariance matrix is ​​set as a diagonal matrix according to the AIS data precision, with diagonal elements of [0.0001, 0.0001, 0.1, 0.1].

[0071] Kalman filter optimization: Taking DC1 of track 1 as an example, Let be the state covariance matrix of the i-th focus set at the current time. In the closed-source observation processing example, this is specifically represented as the closed-source observation covariance matrix. =0.00000001, =0.01+F×0.00000001× Calculated ≈0.010002, , combined =0.8 (closed-source observation weight), thus obtaining ≈0.00253. Weight update combines fuzzy observation weights. Taking track 1's do1(d1) as an example, =0.215, Calculated ≈0.0042, Gaussian distribution The final updated weight ≈0.00007, retain the top 30% of the tracks (2 tracks in total) for subsequent calculations.

[0072] In step 106, the movement prediction and confidence assessment are performed: the posterior probability of each destination is calculated, the optimal destination and arrival time are selected, confidence intervals are constructed, confidence indices are calculated, levels are divided according to the assessment criteria, and the final results are output.

[0073] In embodiments of the present invention, based on the updated track weights, the posterior probability of the target going to each destination is calculated to obtain the motion prediction result, and a confidence index is output based on the uncertainty metric of the state estimation to characterize the reliability of the prediction. This step specifically includes: based on the fused state estimation result and the set of possible destinations of the target (the possible destinations of the target can be automatically generated according to the destination state components in the ERV dynamics model, or can be preset by external prior knowledge (such as historical route databases, port lists), after steps such as destination probability calculation, confidence interval construction, and evaluation level classification, the motion prediction result, confidence index, and evaluation level are obtained.

[0074] In this invention, the uncertainty measure of state estimation refers to a mathematical index used to quantify the reliability or dispersion of the target state estimation result. This measure can take various specific forms, including but not limited to: the state covariance matrix, the standard deviation of each dimension of the state estimation, the parameters of the confidence ellipse constructed based on the covariance matrix (such as axis length and orientation angle), information entropy, or the directly output confidence probability value. For ease of engineering implementation, this embodiment uses the covariance matrix as the main form of uncertainty measure, constructs a confidence ellipse based on the target's state covariance matrix updated by Kalman filtering, and then obtains the confidence index. Those skilled in the art will understand that other equivalent uncertainty measure forms can also achieve the technical objectives of this invention, and all fall within the protection scope of this application.

[0075] The specific process is as follows: Destinations and arrival time prediction: Define the set of destinations D = { ,..., Arrival time T~U( , The likelihood probability is calculated using numerical integration of Simpson's rule. Where q represents the number of integration points. The time node corresponding to the endpoint of the integration interval. The time node corresponding to the starting and ending points of the integration interval. Let be the probability value corresponding to a single integration point. Let be the probability value corresponding to the q-th integration point (the last endpoint integration point). Let be the probability value corresponding to the even-numbered integration point within the integration interval.

[0076] The probability of going to / from Bayes' theorem is obtained by combining the two. .

[0077] Confidence calculation: Based on the posterior covariance matrix of the target state in the final fused output, construct a 95% confidence ellipse. ( , (where covariance eigenvalues ​​are used), and the axis diameter is... , The confidence level is measured by the proportion of the true location within the ellipse. The fused target state covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The two eigenvalues ​​correspond to the estimated variances of the horizontal and vertical directions of the target position plane, respectively. The eigenvectors are used to determine the deflection angle of the confidence ellipse. The chi-square distribution quantile of 5.991, corresponding to the 95% confidence level in the two-dimensional plane, is selected. Using this quantile, the eigenvalues ​​are square-rooted and scaled to calculate the major and minor axes of the confidence ellipse. Then, with the mean of the fused target estimated position as the center, the major and minor axes, combined with the rotation angle, can be used to construct a 95% confidence ellipse, representing a 95% probability that the true target position falls within this elliptical region.

[0078] State estimation and prediction: Current state estimation The Gaussian mixture distribution is used as an approximation. The pure position state of the target in the two-dimensional plane at time tn only represents the location of the target and is derived from the augmented state vector. The extracted position sub-state component does not include other state information such as velocity and destination; future state prediction. Iterative calculations based on the dynamic model using Kalman filtering > , The starting node for integration in the time dimension when performing numerical integration on the probability distribution of the target state.

[0079] The assessment level is divided into five levels based on the absolute error (distance from the true AIS location) and the coverage ratio of the confidence interval. For example, Level 1: the proportion of absolute error <5km is ≥80% and the coverage ratio of the confidence interval is ≥60%; Level 2: the proportion of absolute error <5km is ≥60% or <10km is ≥80% and the coverage ratio is ≥60%, etc.

[0080] For example, in predicting the destination and arrival time, the destination set D = {d1~d6} is defined, and the arrival time T ~ U(2003-10-01, 2003-10-07). The likelihood probability p is calculated using Simpson's rule numerical integration. |D=d), select 7 integration points (q=7), integration interval [T1=2003-10-01, T7=2003-10-07], step size =1 day. Calculate the likelihood probability p( ) at each integration point. |D=d,T= ), Let d1 be the arrival time corresponding to the i-th integration point: for example, d1 at... The likelihood probability of 2003-10-03 is 0.32. Substituting this into the formula... (This formula is a 7-point Simpson numerical integral scheme, where pn represents the conditional likelihood probability value corresponding to the nth discrete integration time step), and the p( of d1 is calculated.) Given |D=d1)≈0.28, and combining this with the prior probability p(D=d1)=0.1667, we obtain the posterior probability p(D=d1| )∝0.28×0.1667≈0.0467. Similarly, calculate the posterior probabilities of other destinations and select d3 as the optimal destination (posterior probability 0.052), with an arrival time of 2003-10-04.

[0081] Confidence calculation is based on the fused covariance matrix, extracting covariance eigenvalues. =2.32、 =3.10, construct a 95% confidence ellipse The axis diameters are 11.4 km and 15.2 km. Through AIS real-world location verification, the proportion of real locations within the ellipse in 100 fusion results was 68%, which serves as the confidence index.

[0082] In state estimation and prediction, current state estimation A mixture distribution approximation using three Gaussian components is employed, with the component mean representing the state estimate of each high-weighted track and the weights representing track weights; future state prediction is then performed. Based on the dynamic model, the calculation is performed iteratively using Kalman filtering. =2003-10-07, the predicted position is (0.5°N, -159.8°W) and the speed is 14km / h.

[0083] The assessment level is determined by the weighting of absolute error and confidence coverage. In the calculation of absolute error... The estimated location on October 1, 2003, is 3.2 km from the actual AIS location. In 100 fusion results, 75% have an absolute error <5 km, and 68% cover the confidence interval. Based on the assessment standards, it is classified as a Level 2 assessment. In the open-source information reliability calculation, combining historical report data (consistent in 5 out of 8 reports) with the current fusion assessment level (Level 2 corresponding to c=0.85), the Beta distribution parameters are updated: r²=5.25+0.85=6.1, s²=3.15+0.15=3.3. The final reliability is 0.649, corresponding to a "very reliable" level.

[0084] Through the implementation of the above steps, this invention completes the entire process from open-source information collection, quantification, fusion to predictive assessment, and finally outputs target state estimation results, optimal destination and arrival time, confidence index, and assessment level, verifying the feasibility and effectiveness of the technical solution of this invention. All parameters, formulas, and steps in this implementation process are clearly quantified, and those skilled in the art can adapt them to different target tracking and prediction needs by adjusting parameters such as ship type and navigation scenario.

[0085] The target tracking and motion prediction method based on open and closed source information fusion provided in this embodiment has the following beneficial effects: (1) The reliability quantification of open source information is accurate and efficient: Through multi-benchmark weight initialization (historical track benchmark and heading consistency benchmark), PCR6 rule conflict fusion and dynamic reliability assessment based on Beta distribution, the fuzzy representation of open source information is transformed into a quantitative form with weights and covariance. This effectively solves the problems of insufficient processing of fuzzy information and static reliability assessment in traditional methods, making the uncertainty quantification of open source information more in line with the actual scenario and laying a solid foundation for fusion accuracy.

[0086] (2) The heterogeneous information fusion mechanism has strong adaptability: In view of the probability uncertainty of closed source information and the reliability fluctuation of open source information, the ERV dynamic model and Kalman filter are combined to achieve deep fusion of heterogeneous information through possible trajectory generation and dynamic weight update; the PCR6 rule effectively resolves information conflict, avoids the result bias caused by a single fusion rule, and significantly improves the fusion accuracy and stability.

[0087] (3) Comprehensive trend prediction and confidence assessment: It innovatively integrates the prediction function of destination and arrival time, calculates the posterior probability through Simpson integral, and constructs a 95% confidence ellipse and a five-level evaluation level, filling the gap of existing technologies that lack key trend prediction and interpretable confidence indicators, providing a quantitative reliability reference for decision-making and effectively reducing decision-making risks.

[0088] (4) Excellent interpretability and engineering adaptability: The technical solution is traceable throughout the process and has no black box model dependence; the design of fuzzy observation generation, accessibility filtering, covariance supplementation and other features are fully adapted to the actual data format. Ordinary technicians can fully implement it according to the description of this embodiment, which solves the problem of the difficulty of engineering implementation of existing algorithms and can be widely adapted to the needs of multiple scenarios such as intelligent shipping.

[0089] Example 2 In embodiments of the present invention, a target tracking and movement prediction system based on open and closed source information fusion is also provided, which is used to implement the steps of the target tracking and movement prediction method based on open and closed source information fusion described in Embodiment 1.

[0090] Figure 2This is a schematic diagram of the target tracking and motion prediction system based on open and closed source information fusion according to an embodiment of the present invention, as shown below. Figure 2 As shown, the target tracking and movement prediction system based on open and closed source information fusion provided by the present invention includes an open source information processing module 201, an open source information reliability quantification module 202, an open and closed source information fusion module 203, and a movement prediction and confidence assessment module 204.

[0091] In this embodiment, the data flow between modules is as follows: Open-source text data is first input to the open-source information processing module 201. After information extraction, fuzzy observation generation, and accessibility filtering, an open-source fuzzy observation set is output. This set, along with historical fact data and closed-source observation data, is input to the open-source information reliability quantification module 202, which outputs weighted and covariant-weighted open-source fuzzy observations (i.e., open-source fuzzy observations). , The open-source and closed-source dynamic reliability values ​​are input together with the closed-source observation data into the open-closed-source information fusion module 203. After possible trajectory generation, weight prediction and updating, the target state estimation result and trajectory weight are output. Finally, the fusion result and the set of possible destinations of the target are input into the movement prediction and confidence assessment module 204, which outputs the movement prediction result, confidence index and assessment level.

[0092] The open-source information processing module 201 is used to transform unstructured open-source information into a quantitative form that can be fused with closed-source information. The input to this module is open-source information, such as open-source text data crawled from the web (originating from an external data acquisition layer), and the output is a standardized set of open-source fuzzy observations. In this embodiment, the open-source information processing module 201 includes an information extraction unit, a fuzzy observation generation unit, and an accessibility filtering unit.

[0093] The information extraction unit uses a large model (such as Qwen-72B-Chat) to extract key information from open-source text data. By using the preset Prompt command, it extracts fields such as target name, event time, geographical location, speed and heading description from the open-source text data, transforming unstructured text into structured data.

[0094] The fuzzy observation generation unit addresses the fuzziness of open-source information by performing time, speed, and position processing, supplementing quantified data based on closed-source information observation elements (including time, speed, heading, and position). Specifically, time processing generates time points at equal intervals; speed and heading processing divides the range according to the target's maximum speed and randomly selects values ​​or converts them into specific angles based on eight bearings; position processing uses the area center coordinates as a reference, generating fuzzy points in eight bearings with a preset radius r, and can iteratively expand to generate more points, performing deduplication after generation.

[0095] The reachability filtering unit uses the most recent closed-source observation value before the open-source information as a benchmark, judges the reachability according to the target's maximum speed, and retains only the fuzzy observation values ​​whose distance is less than or equal to the maximum reachable distance, thus obtaining a set of effective open-source observation values ​​(open-source fuzzy observation values ​​that meet the target's motion capability constraints and are retained after the open-source information is generated by fuzzy observation values ​​and reachedability filtering).

[0096] The open-source information reliability quantification module 202 is used to quantify the reliability and uncertainty of open-source information. Its inputs are open-source fuzzy observations, historical factual data, and closed-source observation data. The output is a weighted and covariant-weighted quantitative representation of the open-source information (i.e., weighted data). Cartesian coordinate system covariance ), and the dynamic reliability value of open source information.

[0097] In this embodiment, the open-source information reliability quantification module 202 includes the following sub-units: Weight initialization unit: A multi-reference weight initialization strategy is adopted, including at least historical track reference and heading consistency reference. The historical track reference uses fuzzy observations as the destination and calculates the destination probability through Bayesian inference as the initial weight; the heading consistency reference calculates the difference between the azimuth angle from the nearest closed-source observation point to the fuzzy observation point and the open-source heading transformation angle, calculates the weight through a cosine function and normalizes it.

[0098] DSmT Weight Fusion Unit: To address the conflict among multiple benchmark weights, the proportional conflict redistribution rule PCR6 is used for fusion to obtain the final fuzzy observation weights. .

[0099] Covariance calculation unit: based on weights Classify into levels and allocate polar coordinate covariance. The diagonal elements are the distance error variance and the azimuth error variance, respectively, which are transformed into the Cartesian coordinate system covariance through eigenvalue decomposition. .

[0100] Dynamic Reliability Assessment Unit: This unit combines historical factual consistency with fusion assessment results to model reliability based on Beta distribution. It statistically analyzes the total number of historical release reports from open-source sources and the number consistent with the facts, calculates prior distribution parameters, and then applies the reliability values ​​obtained from this fusion assessment. As a new observation sample, the posterior distribution parameters are updated, and the expectation of the posterior Beta distribution is taken as the final dynamic reliability value.

[0101] The open-source / closed-source information fusion module 203 is used to fuse the probabilistic uncertainty of closed-source information with the reliability quantification results of open-source information. Its inputs are closed-source observation data (such as AIS data and radar data) and the weights output by the open-source information reliability quantification module. Covariance The output is the target state estimation result and track weights.

[0102] In this embodiment, the open / closed source information fusion module 203 includes a possible trajectory generation unit, a dynamics and observation model construction unit, and a weight prediction and update unit. Specifically: the possible trajectory generation unit arranges the closed-source observation set and the open-source fuzzy observation set in chronological order, and combines elements within the open-source set to generate all possible trajectories. Each trajectory fuses open and closed-source observations in chronological order and associates the corresponding weights and covariance. The dynamics and observation model construction unit uses the equilibrium recovery velocity (ERV) model as the dynamics model, with the state equation being... The state vector contains the current state and the destination state; the observation model is... The weight prediction and update unit is used to predict weights based on the state equation and calculate the predicted weight factors. This yields the unnormalized weights for the next time step; combined with the weights from open-source fuzzy observations. The fused track weights are calculated using the Kalman filter update formula. And retain high-weight tracks.

[0103] The motion prediction and confidence assessment module 204 is used to predict the future destination and arrival time of the target and quantify the confidence of the fusion results. Its inputs are the fused state estimation results (i.e., the state estimation results and track weights output by the open and closed source information fusion module 203) and the set of possible destinations of the target. The outputs are the motion prediction results, confidence index, and assessment level.

[0104] In this embodiment, the movement prediction and confidence assessment module 204 includes the following sub-units: Destination and Arrival Time Prediction Unit: Define the set of destinations The arrival times follow a uniform distribution, and the likelihood probability is calculated using numerical integration based on Simpson's rule. By combining Bayes' theorem to obtain the probability of destination, the optimal destination and arrival time are selected.

[0105] Confidence Calculation Unit: Based on the covariance matrix output by the open / closed source information fusion module 203, eigenvalue decomposition is performed to construct a 95% confidence ellipse with axis diameter of... , The confidence level is measured by the proportion of the actual position within the ellipse.

[0106] State estimation and prediction unit: The posterior probability distribution of the current state is approximated by a Gaussian mixture distribution, and the predicted value of the future state is calculated iteratively by Kalman filtering based on the dynamic model.

[0107] Assessment level classification unit: Based on the absolute error (distance from the true location of AIS) and the coverage ratio of the confidence interval, the assessment level is classified according to the preset five-level assessment standard, and the assessment level and corresponding reliability value are output.

[0108] In this embodiment, the specific parameter settings, calculation formulas, and processing procedures of each module completely correspond to the method steps described in Embodiment 1. Those skilled in the art, based on the description of this embodiment and in conjunction with… Figure 2 The system architecture shown can be used to build a complete software and hardware system, enabling the quantification of open source information, the fusion of open and closed source information, target tracking, and trend prediction.

[0109] In this embodiment, the system also includes an output module 205, which is used to format and visualize the output results of the movement prediction and confidence assessment module to facilitate user understanding and decision-making. The input of this module is the movement prediction results (optimal destination and arrival time), confidence index (confidence ellipse parameters, confidence percentage), and assessment level (levels one to five and corresponding reliability values) output by the movement prediction and confidence assessment module 204; the output is a formatted output result, including but not limited to: text reports, visual charts (such as target track charts, confidence ellipse diagrams), structured data (such as JSON format, for third-party system calls), etc., which are output to the user terminal, storage device, or external decision-making system.

[0110] It is understandable that, without departing from the core idea of ​​this invention—"accurately quantifying the reliability of open-source information, efficiently integrating heterogeneous information from open and closed sources, and balancing trend prediction and confidence assessment"—equivalent modifications (related alternatives that achieve the same functionality and advantages) can be made to the partial implementation or specific algorithm of this solution. For example: Regarding alternative implementations for the open-source information processing module 201: Fuzzy observation generation can employ various point expansion strategies. Besides generating points through 8-directional iterative diffusion, alternative methods include generating uniformly distributed fuzzy points based on the target's historical track distribution characteristics using a grid partitioning method, or generating points through random sampling based on a Gaussian distribution. The key is to ensure the points cover the area described by the open-source information and to guarantee validity through deduplication. For open-source information extraction, in addition to using Qwen-72B-Chat, it can be replaced with large models with strong information extraction capabilities such as GPT-4 and DeepSeek, or a hybrid extraction scheme of "rule templates + machine learning models (such as BERT)". The core is to ensure accurate extraction of key fields such as target name, time, and location, achieving the transformation from unstructured text to structured data. For the accessibility judgment of fuzzy observation filtering, besides calculating the maximum distance based on the maximum speed, the target movement mode (such as constant speed or variable speed) can be introduced to dynamically adjust the reachable distance threshold, or the maximum travel distance can be corrected by combining environmental factors such as weather and sea conditions, further improving filtering accuracy.

[0111] Regarding alternative implementations for the open-source information reliability quantification module 202: In addition to using the PCR6 rule, weight fusion can employ the PCR5 rule or the weighted average fusion rule under the DSmT framework. PCR5 achieves similar accuracy to PCR6 in pairwise evidence fusion with lower computational complexity, effectively resolving weight conflicts. Weight initialization, besides using historical track and heading consistency as dual benchmarks, can add a source credibility benchmark (such as the historical reliability rate of the open-source information publisher), or calculate initial weights based on the semantic matching degree of the target activity area, ensuring the weights comprehensively reflect the reliability of the open-source information. Covariance calculation, besides allocating fixed values ​​according to weight levels, can dynamically adjust the covariance based on the historical matching error between open-source and closed-source information, or introduce information entropy to quantify uncertainty, allocating covariance according to entropy values ​​to achieve precise matching between covariance and information uncertainty.

[0112] Regarding alternative implementations for the open / closed source information fusion module 203: In addition to the ERV model, the dynamic model can utilize filtering models suitable for nonlinear systems, such as the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF), or a particle filter model, as long as it effectively characterizes the target's motion state and is adaptable to heterogeneous information fusion scenarios. For trajectory generation, besides using a full permutation combination, when the number of open-source observations is large, it can be replaced by selecting the Top-K most reliable fuzzy observations based on weights and combining them, or using a clustering algorithm to merge similar observations before generating the trajectory, thus reducing computational complexity while maintaining fusion accuracy. In weight updates and predictions, besides using a Gaussian mixture distribution approximation, a Dirichlet distribution or a multivariate t-distribution can be used, which can also accurately characterize the probability distribution features of the state variables.

[0113] Regarding alternative implementations for the motion prediction and confidence assessment modules: In addition to using Simpson's rule numerical integration for calculating the posterior probability of destinations, numerical integration methods such as trapezoidal integration and Monte Carlo integration can be employed. Alternatively, a prior probability model can be established based on the statistical patterns of historical destinations of the target, simplifying the integration calculation process. For confidence assessment, besides using a 95% confidence ellipse, the confidence level can be adjusted according to actual needs (e.g., 90%, 99%), or representation forms such as rectangular confidence intervals and spherical confidence regions can be used. The core is to reflect the reliability of the fusion results through quantitative indicators. In addition to being based on absolute error and confidence coverage ratio, the assessment level classification can introduce relative error (the ratio to the target's movement distance) or a dynamic error threshold (adjusted according to the target's speed) to make the assessment criteria more adaptable to targets in different motion states.

[0114] The aforementioned alternatives all follow the core design logic of this invention and do not change the core processes of this invention, such as structured processing of open-source information, quantification of reliability, fusion of heterogeneous information, prediction of trends, and assessment of confidence. Adjustments are made only in specific algorithm selection, parameter settings, or implementation details. These variations all achieve the same functional objectives as the original solution and fall within the protection scope of this invention. Those skilled in the art can flexibly choose the appropriate implementation method based on the computational resources, data characteristics, and other requirements of the actual scenario.

[0115] Example 3 In embodiments of the present invention, an electronic device is also provided. Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention, such as... Figure 3 As shown, the electronic device of the present invention includes a processor 301 and a memory 302, wherein, The memory 302 stores a computer program. When the computer program is read and executed by the processor 301, it performs the steps described above in the embodiment of the target tracking and movement prediction method based on open and closed source information fusion.

[0116] Example 4 In embodiments of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, wherein the computer program is configured to execute the steps described above in the embodiments of the target tracking and movement prediction method based on open and closed source information fusion.

[0117] In this embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0118] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A target tracking and movement prediction method based on open- and closed-source information fusion, characterized in that, include: Obtain closed-source observation data and open-source information of the target; Multiple fuzzy observations are generated based on the open-source information, and the fuzzy observations are filtered according to the target's maximum speed. The initial weights of each fuzzy observation are calculated based on multiple benchmarks, and the initial weights are fused using a conflict confidence allocation rule to obtain the final weights of each fuzzy observation. The closed-source observations in the closed-source observation data are combined with the filtered fuzzy observations in chronological order to generate at least one possible track. The target state is predicted based on a dynamic model that includes the destination state of the target, and the weights of the possible trajectories are updated by combining the final weights and the corresponding covariance matrix. Based on the updated track weights, the posterior probability of the target going to each destination is calculated to obtain the motion prediction result, and the confidence index is output based on the uncertainty measure of state estimation to characterize the reliability of the prediction.

2. The target tracking and movement prediction method based on open and closed source information fusion according to claim 1, characterized in that, The step of generating multiple fuzzy observations based on the open-source information and filtering the fuzzy observations according to the target's maximum speed further includes: Extract time information, location information, and speed and heading information from the open-source information; At least one observation time point is generated based on the time information, center coordinates are generated based on the location information, and multiple candidate points are generated based on the center coordinates. Based on the speed and heading information, the multiple candidate points are assigned speed and heading values ​​to form fuzzy observation values; Based on the most recent closed-source observation of the target prior to the open-source information, the reachable distance is calculated according to the target's maximum speed, and ambiguous observations whose distance from the most recent closed-source observation exceeds the reachable distance are filtered out.

3. The target tracking and movement prediction method based on open and closed source information fusion according to claim 1, characterized in that, The step of calculating the initial weights of each fuzzy observation based on multiple benchmarks and fusing the initial weights using a conflict confidence allocation rule further includes: Based on historical track benchmarks and heading consistency benchmarks, the initial weights of each fuzzy observation are calculated. By adopting the proportional conflict redistribution rule, the initial weights of the same fuzzy observation obtained from different benchmarks are used as evidence to fuse them and obtain the final weight of the fuzzy observation.

4. The target tracking and movement prediction method based on open and closed source information fusion according to claim 1, characterized in that, It also includes the steps of establishing a Beta prior distribution based on the historical factual consistency data of the open source information, performing a Bayesian update on the Beta prior distribution according to the reliability value obtained from this fusion evaluation, and using the expected value of the updated distribution as the dynamic reliability value of the open source information and outputting it.

5. The target tracking and movement prediction method based on open and closed source information fusion according to claim 4, characterized in that, The reliability value obtained from this fusion assessment was generated through the following steps: Using the target location in the closed-source observation data as the true location reference, calculate the absolute error between the fused target location and the true location; A confidence interval with a pre-set confidence level is constructed based on the uncertainty measure of state estimation, and it is determined whether the true location falls within the confidence interval; Based on the statistical percentage of the absolute error and the statistical percentage of the true location falling within the confidence interval, the evaluation level of this fusion is determined by referring to the preset evaluation level classification standard. The reliability value is obtained based on the preset mapping relationship between the evaluation level and the reliability value.

6. The target tracking and movement prediction method based on open and closed source information fusion according to claim 1, characterized in that, The step of calculating the posterior probability of the target going to each destination based on the updated track weights to obtain the motion prediction result further includes: Define the prior distributions of the destination set and arrival time; The likelihood probability under a given destination is calculated using numerical integration. By combining the likelihood probability and the prior probability of the destination, the posterior probability of each destination is calculated according to Bayes' theorem, and the optimal destination and its corresponding arrival time are selected.

7. The target tracking and movement prediction method based on open and closed source information fusion according to claim 1, characterized in that, The step of using the uncertainty metric output based on state estimation to characterize the confidence index of prediction reliability further includes: The uncertainty measure of the state estimate is decomposed into eigenvalues, and a confidence ellipse is constructed based on the obtained eigenvalues. The statistical proportion of the target's true location falling into the confidence ellipse is used as the confidence index.

8. A target tracking and movement prediction system based on open and closed source information fusion, characterized in that, include: An open-source information processing module is used to acquire and process open-source information, and generate and filter fuzzy observation values. An open-source information reliability quantification module is used to calculate the initial weights of each fuzzy observation based on multiple benchmarks, and to fuse the initial weights using a conflict confidence allocation rule to obtain the final weights. The open-source and closed-source information fusion module is used to combine closed-source observations and filtered fuzzy observations in chronological order to generate possible tracks, and to update the weights of the possible tracks based on the dynamic model containing the destination state of the target and the final weights. The motion prediction and confidence assessment module is used to calculate the posterior probability of the target going to each destination based on the updated track weights to obtain the motion prediction result, and output a confidence index to characterize the reliability of the prediction based on the uncertainty measure of the state prediction.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor is used to execute the computer program stored in the memory to implement the target tracking and movement prediction method based on open and closed source information fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the target tracking and movement prediction method based on open and closed source information fusion as described in any one of claims 1 to 7.

Images