Vessel positioning method based on multi-node collaboration

Abstract

Disclosed in the present invention is a vessel positioning method based on multi-node collaboration. The method comprises: on the basis of a denoised multi-source data set, using a convolutional neural network to extract features of different dimensions, such as distance-velocity information and shape-thermal features, resulting from sensor heterogeneity, so as to obtain a preliminary feature vector group; for the preliminary feature vector group, if the feature dimensions are inconsistent, unifying temporal and spatial scales by means of an alignment transformation, so as to obtain an aligned feature vector group; using a weighted fusion mechanism to integrate dispersed information from the aligned feature vector group, so as to obtain a comprehensive fused feature representation; on the basis of the comprehensive fused feature representation, using a clustering method in a multi-target dense scenario to separate distinctive patterns of vessels, so as to obtain a separated target feature subset; and for the separated target feature subset, if there is a high risk of identity misjudgment, performing matching by means of comparison with a preset identity template, so as to determine identity labels of the vessels.

Description

A ship positioning method based on multi-node cooperation TECHNICAL FIELD

[0001] The application belongs to the technical field of ship positioning, and particularly relates to a ship positioning method based on multi-node cooperation. BACKGROUND

[0002] In modern marine environments, the accuracy of ship positioning technology is crucial for maritime safety, shipping management, and national defense strategy. With the increasing frequency of global marine activities, whether it is commercial shipping, fishing operations, or military operations, efficient and accurate determination of ship position and identification of target identity are required. However, the marine environment is complex and variable, with sea clutter, weather interference, and multi-target dense scenarios posing extremely high requirements for positioning and identification technology. Traditional positioning methods can handle single data sources, but in complex scenarios, relying on a single sensor or simple data processing methods has become difficult to meet the needs. Therefore, developing technology that can integrate multiple sensor data and achieve intelligent cooperative positioning has become a key to improving marine target monitoring capabilities.

[0003] Existing ship positioning technology mainly relies on a single type of sensor, such as radar or optical equipment, to detect and estimate the position of the target through direct signal processing. However, these methods often struggle when faced with the integration of multi-source heterogeneous data. For example, radar signals are easily disturbed by strong sea waves or bad weather, resulting in target signal ambiguity; optical images are difficult to capture clear features at night or in foggy conditions; infrared data is sensitive to heat sources, but is easily disturbed in complex backgrounds. These limitations make existing methods lack effective means in multi-sensor data cooperative processing, unable to fully utilize the complementary advantages of multiple data, and difficult to achieve stable and accurate positioning in dynamic marine environments.

[0004] The deeper challenge lies in how to extract and fuse key features from multi-source heterogeneous data. Different sensors produce data with completely different physical characteristics and forms, such as radar signals that emphasize target distance and speed information in time series; optical images that are suitable for describing target shape and texture in spatial pixel information; and infrared data that focuses on thermal characteristics. The differences in format, dimension, and time scale of these data make direct integration extremely difficult. Simple superposition or manually designed fusion rules often fail to capture the deep associations between data, resulting in insufficient feature extraction and affecting the accuracy of target identification.

[0005] The above problems not only affect real-time monitoring, but also can cause serious consequences in critical tasks, such as positioning deviation in maritime rescue or military operations, which can lead to misallocation of resources or inaccurate strategic judgments. Therefore, how to integrate multi-source sensor data, mine deep correlation features, and realize dynamic modeling of ship trajectories on this basis in a variable and high-noise marine environment has become a technical problem to be solved. SUMMARY

[0006] The present application provides a ship positioning method based on multi-node cooperation to solve the above problems existing in the prior art.

[0007] To achieve the above purpose, the present application provides a ship positioning method based on multi-node cooperation, comprising the following steps:

[0008] Obtain the multi-node original signal sequence and image data, remove the noise caused by strong signal interference by using a preprocessing method, and obtain a denoised multi-source data set;

[0009] Extract different dimension features of the denoised multi-source data set by using a convolutional neural network, and obtain a preliminary feature vector group;

[0010] When the feature dimensions of the preliminary feature vector group are inconsistent, unify the time and space scales by alignment transformation, and obtain an aligned feature vector group;

[0011] Integrate the dispersed information in the aligned feature vector group by using a weighted fusion mechanism, and obtain a comprehensive fusion feature representation;

[0012] According to the comprehensive fusion feature representation, use a clustering method to separate the unique patterns of each ship, and obtain a separated target feature subset;

[0013] According to the separated target feature subset, determine whether there is a high risk of identity misjudgment, and if there is, match by comparing a preset identity template to determine the identity label of each ship;

[0014] According to the identity label of each ship and the comprehensive fusion feature representation, use Kalman filtering to correct the trajectory prediction difference, obtain a predicted trajectory sequence, and determine the real-time position of the ship according to the predicted trajectory sequence.

[0015] Optionally, the denoised multi-source data set comprises:

[0016] Collect the original signal sequence and image data by using radar, optical and infrared sensors, remove noise interference by using a filtering algorithm, and obtain a first denoised data set;

[0017] If the intensity of the signal sequence in the first denoised data set is lower than a preset threshold, process the signal sequence by using a mean filtering algorithm to obtain a second denoised signal sequence;

[0018] If the definition of the image data in the first denoising data set is lower than the preset threshold, the image data is processed by using a Gaussian filtering algorithm to obtain second denoising image data;

[0019] The second denoising signal sequence and the second denoising image data are fused by using a principal component analysis algorithm to obtain a first fusion data set;

[0020] According to the first fusion data set, a weighted average method is used to integrate the features of the radar signal, the optical image and the infrared data to obtain a second fusion data set;

[0021] If the dimension of the feature vector in the second fusion data set is higher than the preset threshold, a dimension reduction algorithm is used for processing to obtain a third fusion data set;

[0022] The third fusion data set is classified by using a clustering algorithm to obtain a multi-source data set.

[0023] Optionally, the obtaining of the preliminary feature vector group comprises:

[0024] An initial feature vector group is extracted from the multi-source data set by using a convolutional neural network, the initial feature vector group contains distance information features, speed information features and shape heat features, and a preliminary feature set is obtained;

[0025] If the dimension of the feature vector in the preliminary feature set exceeds the preset threshold, a principal component analysis algorithm is used for dimension reduction processing on the preliminary feature set to obtain a dimension-reduced feature set;

[0026] According to the dimension-reduced feature set, a support vector machine algorithm is used to classify the feature vectors to determine the feature categories, and a classified feature set is obtained;

[0027] For the classified feature set, a weighted average method is used to integrate the distance information features, the speed information features and the shape heat features to obtain a final feature vector group.

[0028] Optionally, the obtaining of the aligned feature vector group comprises:

[0029] The time scale and the space scale of each feature vector in the preliminary feature vector group are obtained, wherein the time scale represents the distribution range or the sampling frequency of the feature vector in the time dimension, and the space scale represents the distribution range or the resolution of the feature vector in the space dimension;

[0030] The time scale difference and the space scale difference between the feature vectors are calculated by using a statistical analysis method, the statistical analysis method includes but is not limited to analysis of variance and correlation analysis or standard deviation calculation, to quantify the degree of scale difference;

[0031] If there is a time scale or a spatial scale inconsistency, the eigenvectors are transformed by a linear interpolation method to obtain an aligned eigenvector group.

[0032] Optionally, the obtaining of the comprehensive fusion feature representation comprises:

[0033] From the aligned eigenvector group, the attribute distribution of each eigenvector is obtained, the mean and variance are calculated by a statistical analysis method, and a quantitative description of the feature distribution is obtained.

[0034] According to the quantitative description of the feature distribution, a preset weighting mechanism is used to allocate weights, and a weighted average method is used to integrate the dispersed information to obtain the preliminary fusion eigenvector.

[0035] If the fusion accuracy of the preliminary fusion eigenvector is lower than a preset threshold, the weight allocation is adjusted by a linear regression method, and the weighted average is recalculated to obtain the optimized fusion eigenvector.

[0036] From the optimized fusion eigenvector, a service correlation feature is extracted, the features are grouped by a clustering method, and the comprehensive fusion feature representation is obtained.

[0037] Optionally, the obtaining of the separated target feature subset comprises:

[0038] The comprehensive fusion features are grouped by a density clustering method to obtain an initial clustering result.

[0039] According to the initial clustering result, a kernel density estimation method is used to smooth the feature distribution in each cluster to obtain a smooth clustering feature group.

[0040] If the inter-cluster distance of the smooth clustering feature group is lower than a preset threshold, the inter-cluster boundary is adjusted by an Euclidean distance calculation method to determine an optimized clustering feature group.

[0041] According to the optimized clustering feature group, a weighted average method is used to integrate the feature information within the cluster to obtain an integrated feature subset.

[0042] If the feature consistency of the integrated feature subset is lower than a preset threshold, the feature subset is completed by a linear interpolation method to obtain a completed feature subset.

[0043] According to the completed feature subset, a principal component analysis method is used to reduce the dimension of the feature subset to obtain a reduced dimension feature subset.

[0044] According to the reduced dimension feature subset, a cosine similarity calculation method is used to evaluate the uniqueness of each ship feature subset to determine the final target feature subset.

[0045] Optionally, the determining of the identity label of each ship comprises:

[0046] obtaining an initial template set by acquiring preset identity templates from the template database through the target feature subset;

[0047] if the feature consistency of the initial template set and the target feature subset is lower than a preset threshold, adjusting the template set by a cosine similarity calculation method to obtain an optimized template set;

[0048] calculating the matching degree of the target feature subset and each template by a feature comparison method according to the optimized template set to obtain a matching degree score;

[0049] if the matching degree score is higher than a preset threshold, assigning a corresponding identity label to the target feature subset by a label assignment method to obtain a preliminary identity label;

[0050] calculating an identity misjudgment risk by a risk assessment method according to the preliminary identity label to obtain a misjudgment risk value;

[0051] if the misjudgment risk value is higher than a preset threshold, performing a neighbor comparison between the target feature subset and samples in the template database by a k-nearest neighbor algorithm to obtain a corrected identity label;

[0052] updating the ship identity by a label assignment method according to the corrected identity label to determine a final ship identity label.

[0053] Optionally, the determining the real-time position of the ship according to the predicted trajectory sequence comprises:

[0054] obtaining a target motion state from the ship identity label and the comprehensive fusion feature, initializing a state estimation vector by a Kalman filtering method to obtain an initial state estimation;

[0055] extracting real-time motion information from sensor observation data according to the initial state estimation, calculating a target predicted position by a Kalman filtering prediction step to obtain a predicted trajectory point;

[0056] if the deviation of the predicted trajectory point and the sensor observation data exceeds a preset threshold, adjusting the state estimation vector by a Kalman filtering update step to obtain a corrected state estimation;

[0057] optimizing the predicted trajectory point by a trajectory smoothing processing method according to the corrected state estimation to obtain a smoothed trajectory sequence;

[0058] extracting continuous motion features from the smoothed trajectory sequence, adjusting the trajectory point by a dynamic correction process to obtain an optimized trajectory sequence;

[0059] if the matching degree of the optimized trajectory sequence and the sensor observation data is lower than a preset threshold, performing a secondary correction on the trajectory sequence by a least square method to obtain a final trajectory sequence;

[0060] According to the final trajectory sequence, the positioning information of the ship is updated to determine the real-time position of the ship.

[0061] Compared with the prior art, the present application has the following advantages and technical effects:

[0062] The application discloses a ship positioning method based on multi-node cooperation, solves the problems of multi-source sensor data heterogeneity, strong signal interference, scattered features and large positioning deviation in business scenarios. The application removes the noise of the original signal sequence and image data through preprocessing, extracts distance speed and shape heat features by using a convolutional neural network, forms a preliminary feature vector group, and solves the problem of inconsistent feature dimensions by aligning and transforming the space-time scale. Based on a weighted fusion mechanism, scattered information is integrated to generate a comprehensive fusion feature representation, then a clustering method is used to separate the unique patterns of each ship, a preset identity template is matched to reduce the risk of identity misjudgment, and finally a Kalman filter is used to correct the trajectory prediction error to obtain a high-precision predicted trajectory sequence. The application significantly improves the accuracy of ship identity recognition and the accuracy of trajectory prediction in a multi-target dense scene, reduces the positioning deviation, and provides an efficient solution for target tracking in a complex marine environment. BRIEF DESCRIPTION OF DRAWINGS

[0063] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and their description together with the drawings serve to explain the application. In the drawings:

[0064] FIG. 1 is a method flowchart of an embodiment of the application. DETAILED DESCRIPTION

[0065] It should be noted that the embodiments and features in the application can be combined with each other without conflict. The application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0066] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0067] Embodiment one

[0068] As shown in FIG. 1, in this embodiment, a ship positioning method based on multi-node cooperation is provided, including the following steps:

[0069] The multi-node original signal sequence and image data are obtained, and a preprocessing method is used to remove the noise caused by strong signal interference to obtain a denoised multi-source data set;

[0070] extracting different dimension features of the denoised multi-source data set through a convolutional neural network to obtain a preliminary feature vector group;

[0071] When the feature dimensions of the preliminary feature vector group are inconsistent, aligning transformation is used to unify the time and space scales to obtain an aligned feature vector group;

[0072] The dispersed information in the aligned feature vector group is integrated through a weighted fusion mechanism to obtain a comprehensive fusion feature representation;

[0073] According to the comprehensive fusion feature representation, a clustering method is used to separate the unique patterns of each ship to obtain a separated target feature subset;

[0074] According to the separated target feature subset, it is determined whether there is a high risk of identity misjudgment, and when there is, the preset identity template is matched to determine the identity label of each ship;

[0075] According to the identity label of each ship and the comprehensive fusion feature representation, Kalman filtering is used to correct the trajectory prediction difference to obtain a predicted trajectory sequence, and the real-time position of the ship is determined according to the predicted trajectory sequence.

[0076] Specifically, the following steps are included:

[0077] In step S101, the original signal sequence and image data collected by the radar, optical and infrared sensors are preprocessed to remove noise caused by signal interference, and a denoised multi-source data set is obtained.

[0078] Specifically, the original signal sequence and image data collected by the radar, optical and infrared sensors are preprocessed to remove noise interference, and a first denoised data set is obtained. If the intensity of the signal sequence in the first denoised data set is lower than the preset threshold, the mean filtering algorithm is used to process the signal sequence to obtain a second denoised signal sequence. If the clarity of the image data in the first denoised data set is lower than the preset threshold, the Gaussian filtering algorithm is used to process the image data to obtain a second denoised image data. The second denoised signal sequence and the second denoised image data are fused by the principal component analysis algorithm to obtain a first fusion data set. According to the first fusion data set, the features of the radar signal, optical image and infrared data are integrated by the weighted average method to obtain a second fusion data set. If the dimension of the feature vector in the second fusion data set is higher than the preset threshold, the dimension reduction algorithm is used for processing to obtain a third fusion data set. The third fusion data set is classified by the clustering algorithm to obtain a multi-source data set.

[0079] Specifically, in the multi-source sensor data processing scene, the original signal and image data collected by the radar, optical and infrared sensors are often used for target detection and recognition.

[0080] For example, radar sensors capture the distance and velocity information of a target, optical sensors provide high-resolution images, and infrared sensors reflect the thermal characteristics of the target. These data are susceptible to noise interference, such as multipath effects in radar signals or changes in illumination in optical images.

[0081] Specifically, filtering algorithms are used to remove noise to generate a first denoised data set. Mean filtering can be applied to radar signal sequences by smoothing noise by averaging adjacent data.

[0082] For example, an outlier of 100.5 meters in a radar range sequence can be adjusted to a more consistent 100 meters using mean filtering, generating a second denoised signal sequence. Gaussian filtering is suitable for processing optical image data. Assuming the image sharpness is below a preset threshold of 0.8 (a sharpness index based on edge detection), Gaussian filtering smooths pixel values ​​through weighted averaging, preserving edge information and generating a second denoised image data. This processing effectively improves the accuracy of subsequent feature extraction.

[0083] In this embodiment, principal component analysis (PCA) is used to fuse the second denoised signal sequence and image data. Assuming that radar signals provide range and velocity features, optical images extract edge features, and infrared data provide temperature features, PCA maps these high-dimensional features to a low-dimensional space through linear transformation, generating a first fused data set.

[0084] For example, the fused data may retain 90% of the variance, generating a comprehensive feature vector that includes target distance, contour, and temperature. This step reduces data redundancy and improves computational efficiency.

[0085] For example, a weighted average method integrates features from the first fused dataset. Assuming the radar signal has high reliability, its weight is set to 0.5, while the optical and infrared signals each have a weight of 0.25, generating the second fused dataset. This weighting method highlights key features and balances the contributions of multiple data sources. If the feature vector dimension exceeds a preset threshold of 10, a dimensionality reduction algorithm, such as linear discriminant analysis, is used to further compress it to 8 dimensions, generating the third fused dataset, ensuring that computational complexity remains controllable.

[0086] Understandably, clustering algorithms such as K-means are used to classify a third fused dataset. Suppose the system needs to distinguish between a vehicle, a pedestrian, or an obstacle ahead; the clustering algorithm divides the data into three categories based on the distance, contour, and temperature features of the feature vectors.

[0087] The above processing steps significantly improve data quality and target recognition accuracy.

[0088] Step S102: Based on the denoised multi-source data set, a convolutional neural network is used to extract different dimensional features brought about by sensor heterogeneity, such as distance and speed information and shape and thermal features, to obtain a preliminary feature vector set.

[0089] Specifically, a convolutional neural network is used to extract an initial feature vector set from a multi-source dataset. This initial feature vector set includes distance information features, velocity information features, and shape and thermal features, resulting in a preliminary feature set. If the dimension of the feature vectors in the preliminary feature set exceeds a preset threshold, principal component analysis (PCA) is used to reduce the dimensionality of the preliminary feature set, resulting in a dimensionality-reduced feature set. Based on the dimensionality-reduced feature set, a support vector machine (SVM) algorithm is used to classify the feature vectors, determining the feature categories and obtaining a categorized feature set. For the categorized feature set, a weighted average method is used to integrate the distance information features, velocity information features, and shape and thermal features, resulting in the final feature vector set.

[0090] For example, in the field of multi-source data processing, convolutional neural networks are used to extract initial feature vector sets from multi-source data sets acquired by radar, optical, and infrared sensors.

[0091] Specifically, convolutional neural networks (CNNs) extract features from radar signals' range information, optical images' shape characteristics, and infrared data's thermal features through multiple layers of convolution and pooling operations. Assume the range information extracted from radar data includes the distance vector between the target and the sensor, such as numerical values ​​like 10 meters or 15 meters; the shape features of the optical image capture the target's contour information, such as edge curvature; and the thermal features of the infrared data reflect the target's temperature distribution, such as a heat intensity value of 300K. The CNN scans the raw data using convolutional kernels to generate an initial feature vector set containing these features. This method efficiently extracts multidimensional features, providing a rich data foundation for subsequent processing.

[0092] In this embodiment, if the dimension of the initial feature vector group exceeds a preset threshold, for example, if the dimension exceeds 100, then principal component analysis (PCA) is used for dimensionality reduction. PCA projects high-dimensional features into a low-dimensional space through linear transformation, preserving the main information.

[0093] For example, the initial feature vector set contains 120 dimensions of distance, velocity, and thermal features. After principal component analysis, the dimensionality is reduced to 50 dimensions, retaining more than 90% of the data variance. This dimensionality reduction operation reduces computational complexity while maintaining the effectiveness of the features, providing a more concise data representation for subsequent classification tasks.

[0094] For example, for a dimensionality-reduced feature set, the Support Vector Machine (SVM) algorithm is used for feature classification to determine the feature category. SVM classifies features into different categories by finding the optimal hyperplane, such as classifying targets into moving targets and stationary targets.

[0095] Specifically, assuming the dimensionality reduction feature set contains the target's speed features, such as speed values ​​of 5 m / s and 10 m / s, combined with distance and thermal features, the support vector machine can map the features to a high-dimensional space through a kernel function to find the classification boundary.

[0096] For example, targets with a speed greater than 8 meters per second are classified as moving targets. This classification method can clearly distinguish the characteristics of different types of targets, providing a clear categorical basis for subsequent feature integration.

[0097] In this embodiment, a weighted average method is used to integrate distance information features, velocity information features, and shape thermal features for the classification feature set to generate the final feature vector set.

[0098] For example, by setting the weight of distance feature to 0.4, velocity feature to 0.3, and thermal feature to 0.3, a weighted average is calculated to generate a comprehensive feature vector. Assuming a target has a distance feature of 12 meters, a velocity feature of 6 meters per second, and a thermal feature of 310K, a weighted average yields a comprehensive feature vector representing the target's overall characteristics. This integration method balances the importance of multiple feature sources, generating a unified feature representation and providing a more comprehensive perspective for subsequent data analysis.

[0099] Understandably, the above methods closely address the core needs of multi-source data processing. Through convolutional neural networks for feature extraction, principal component analysis for dimensionality reduction, support vector machine classification, and weighted average integration, a complete processing chain is formed from raw data to the final feature vector set. Each step is designed specifically for the characteristics of multi-source data, ensuring comprehensive feature extraction and accurate classification.

[0100] For example, in real-world scenarios, when processing radar, optical, and infrared data, the above methods can effectively integrate multimodal information to generate high-quality feature vector sets, providing reliable support for target recognition and tracking tasks.

[0101] Step S103: For the initial feature vector group, if the feature dimensions are inconsistent, the time and space scales are unified through alignment transformation to obtain the aligned feature vector group.

[0102] Specifically, for the collected preliminary feature vector set, the temporal and spatial scales of each feature vector are obtained. The temporal scale represents the distribution range or sampling frequency of the feature vector in the time dimension, and the spatial scale represents the distribution range or resolution of the feature vector in the spatial dimension. Statistical analysis methods are used to calculate the temporal and spatial scale differences between each feature vector. The statistical analysis methods include, but are not limited to, analysis of variance, correlation analysis, or standard deviation calculation, to quantify the degree of scale difference. If there are inconsistencies in the temporal or spatial scales, the feature vectors are aligned using a linear interpolation method to obtain an aligned feature vector set.

[0103] For example, in the field of sensor data processing, obtaining the temporal and spatial scales is a crucial step for the initial feature vector set collected. The temporal scale reflects the distribution characteristics of the feature vectors over time, such as the frequency or time span of data acquisition by the sensor. The spatial scale describes the distribution of the feature vectors in space, such as the physical coverage area or resolution of the sensor. Taking a radar sensor as an example, assuming its range feature vector set comes from pulse signals at different time points, the temporal scale can be determined by analyzing the sampling frequency; for example, 100 samplings per second results in a temporal scale of 0.01 seconds. The spatial scale can be determined by the radar's resolution; for example, a sampling point every 0.5 meters results in a spatial scale of 0.5 meters. This analysis lays the foundation for subsequent feature alignment.

[0104] In this embodiment, statistical analysis methods are used to quantify the temporal and spatial scale differences between feature vectors.

[0105] For example, for feature vector sets collected by multiple sensors, the degree of dispersion in time scale can be calculated using analysis of variance (ANOVA). Assuming sensor A has a sampling frequency of 100Hz and sensor B has a sampling frequency of 50Hz, ANOVA can reveal a time scale difference of 0.01 seconds versus 0.02 seconds. Similarly, calculating the standard deviation of spatial scale can quantify the difference in resolution between different sensors. For instance, if sensor A has a spatial resolution of 0.5 meters and sensor B has a spatial resolution of 1 meter, the standard deviation reflects the degree of inconsistency in spatial scale. These statistical results provide a basis for subsequent alignment transformations.

[0106] Specifically, if the time scale or spatial scale is inconsistent, a linear interpolation method is used for alignment transformation.

[0107] For example, for feature vector sets with inconsistent time scales, assuming sensor A's sampling points are 0, 0.01, and 0.02 seconds, and sensor B's are 0, 0.02, and 0.04 seconds, a new feature value for sensor B can be generated at 0.01 seconds through linear interpolation, calculated as a time-weighted average based on adjacent points. This method ensures alignment in the time dimension, generating a feature vector set with a uniform time scale. Similarly, for the spatial scale, assuming sensor A has a distance resolution of 0.5 meters and sensor B has a distance resolution of 1 meter, feature values ​​with a resolution of 0.5 meters can be generated through interpolation, ensuring consistency in the spatial scale. The aligned feature vector set is consistent in both time and space, facilitating subsequent classification or fusion processing.

[0108] In this embodiment, the implementation of linear interpolation needs to take into account the physical meaning of the sensor data.

[0109] For example, interpolation of radar distance features needs to ensure that the interpolation points reflect the true physical distance, rather than simply filling in numerical values. Suppose sensor A measures a distance feature value of 10 at 0.5 meters, and sensor B measures feature values ​​of 8 and 12 at 0 meters and 1 meter, respectively. The interpolated feature value at 0.5 meters is approximately 10. This method preserves the physical properties of the data and improves the reliability of aligning feature vector groups. Furthermore, scale alignment can reduce the misclassification rate of subsequent classification algorithms and improve the accuracy of feature fusion.

[0110] Understandably, the advantage of the above method lies in solving the data inconsistency problem caused by the heterogeneity of multi-source sensors through scale analysis and alignment.

[0111] For example, the differences in temporal and spatial scales between the feature vector sets of radar and thermal imaging sensors can be unified into a consistent feature representation through the steps described above. This consistency provides high-quality input data for subsequent support vector machine classification or weighted average fusion, ensuring the coherence and accuracy of feature processing. The entire process, through scale analysis, statistical quantization, and alignment transformation, forms a rigorous logical chain, fully tapping the potential value of multi-source data.

[0112] Step S104: From the aligned feature vector group, a weighted fusion mechanism is used to integrate the scattered information to obtain a comprehensive fusion feature representation.

[0113] Specifically, from the aligned feature vector group, the attribute distribution of each feature vector is obtained, and the mean and variance are calculated using statistical analysis methods to obtain a quantitative description of the feature distribution. Based on the quantitative description of the feature distribution, weights are allocated using a preset weighting mechanism, and the dispersed information is integrated using a weighted average method to obtain the preliminary fused feature vector. If the fusion accuracy of the preliminary fused feature vector is lower than a preset threshold, the weight allocation is adjusted using a linear regression method, and the weighted average is recalculated to obtain the optimized fused feature vector. From the optimized fused feature vector, business-related features are extracted, and the features are grouped using a clustering method to obtain the comprehensive fused feature representation.

[0114] For example, when obtaining the attribute distribution of each feature vector in an aligned feature vector group, statistical analysis methods can be used to quantify the distribution characteristics of the features. Attribute distribution refers to the numerical distribution of feature vectors in a specific dimension, such as the frequency of change in the time dimension or the intensity range in the spatial dimension. One possible implementation is to choose the mean and variance as quantification indicators. The mean reflects the central tendency of the feature vectors, while the variance describes their dispersion.

[0115] For example, in a video analysis scenario, suppose the feature vector set is extracted from the motion trajectory features of objects in surveillance video. The time dimension represents the timestamp of each frame, and the spatial dimension represents the coordinates of the object on a two-dimensional plane. After statistical analysis, a certain feature vector has a time mean of 10 seconds and a variance of 2 seconds, a spatial mean of (50,50) pixels, and a variance of (5,5) pixels. This indicates the range of fluctuation in the motion trajectory in both time and space, which helps to determine the stability of the features in subsequent processing.

[0116] Specifically, when assigning weights using a preset weighting mechanism, weights can be designed based on the importance and reliability of the features.

[0117] For example, in the video analysis scenario described above, if some feature vectors come from high-resolution cameras, their data reliability is higher, and they can be assigned higher weights, such as 0.7; while the feature vectors from low-resolution cameras can be weighted at 0.3. The weighted average method integrates scattered information by summing the weighted sums of each feature vector, generating a preliminary fused feature vector.

[0118] For example, suppose there are three feature vectors with values ​​(10,20), (15,25), and (12,22) and weights of 0.4, 0.3, and 0.2, respectively. After calculating the weighted average, a preliminary fused feature vector (12.1, 22.1) is obtained. This method can effectively integrate multi-source information and improve the representativeness of the features.

[0119] In this embodiment, if the fusion accuracy of the initial feature vector fusion is lower than a preset threshold, for example, if the error between the fused features and the actual target exceeds 5%, the weights are adjusted using a linear regression method. Linear regression can analyze the correlation between each feature and the target based on historical data and reallocate the weights accordingly.

[0120] For example, in video analysis, if a feature vector is found to have a low correlation with the trajectory of a target object, its weight can be reduced from 0.4 to 0.2, and the weighted average can be recalculated to obtain an optimized fused feature vector (11.8, 21.8). This can improve the accuracy of the fused features.

[0121] For example, when extracting business-relevant features from an optimized fused feature vector, one can focus on attributes directly related to business objectives. In video analytics, business-relevant features might be the speed or direction of an object. Assuming the optimized fused feature vector represents the average displacement of an object over 10 seconds, the speed feature is extracted as 5 pixels per second by calculating the displacement difference. Then, clustering methods, such as K-means clustering, are used to group the features.

[0122] For example, speed features can be divided into two groups: "fast movement" and "slow movement." The fast movement group includes features with speeds greater than 4 pixels per second, while the slow movement group includes features with speeds less than or equal to 4 pixels per second. This results in a comprehensive and fused feature representation. This grouping method clearly distinguishes different motion patterns, facilitating subsequent business analysis, such as abnormal behavior detection.

[0123] It should be noted that the above method closely integrates temporal and spatial scale alignment to ensure that feature vector sets are analyzed at a uniform scale. Each step is implemented within the context of the video analysis scenario, avoiding interference from irrelevant business logic.

[0124] For example, both weight adjustment and clustering aim to improve the accuracy and business relevance of feature representations. This logically progressive processing approach, from attribute distribution to comprehensive feature representation integration, advances layer by layer, ensuring the efficiency and consistency of feature processing.

[0125] Step S105: Based on the comprehensive fusion feature representation, a clustering method is used to separate the unique patterns of each ship in a multi-target dense scene to obtain a subset of separated target features.

[0126] Specifically, the integrated features are grouped using density clustering to obtain initial clustering results. Based on the initial clustering results, the feature distribution within each cluster is smoothed using kernel density estimation to obtain smoothed cluster feature groups. If the inter-cluster distance of the smoothed cluster feature groups is lower than a preset threshold, the inter-cluster boundaries are adjusted using Euclidean distance calculation to determine optimized cluster feature groups. Based on the optimized cluster feature groups, the intra-cluster feature information is integrated using a weighted average method to obtain integrated feature subsets. If the feature consistency of the integrated feature subsets is lower than a preset threshold, the feature subsets are completed using linear interpolation to obtain completed feature subsets. Based on the completed feature subsets, the feature subsets are dimensionality-reduced using principal component analysis to obtain dimensionality-reduced feature subsets. Based on the dimensionality-reduced feature subsets, the uniqueness of each ship feature subset is evaluated using cosine similarity calculation to determine the final target feature subsets.

[0127] For example, in the business scenario of ship feature analysis, density clustering methods can be used to group integrated features. Density clustering identifies high-density regions in the feature space, grouping similar feature vectors into one class, making it suitable for handling irregularly shaped feature distributions. Suppose a ship dataset contains features such as position, speed, and heading; density clustering can group similar ship behavior patterns into the same cluster based on the Euclidean distance between features, such as identifying feature groups for ships berthing and those sailing. The initial clustering results may contain noisy points, such as abnormal navigation trajectories, requiring further optimization.

[0128] In this embodiment, the kernel density estimation method smooths the initial clustering results. Kernel density estimation smooths the feature distribution and reduces the influence of noise by introducing a Gaussian kernel function around each feature point.

[0129] For example, when processing the speed characteristics of ships within a cluster, the original data may show large speed fluctuations. Kernel density estimation can generate a smooth distribution curve, highlighting the main speed range, such as 10-15 knots, and weakening the influence of outliers.

[0130] Specifically, if the distance between clusters is below a preset threshold, such as 0.5 units, the cluster boundaries are adjusted using Euclidean distance calculation. Assuming two clusters represent near-shore and ocean-going navigation respectively, if their feature distributions overlap, the distance from each feature point to the cluster center can be recalculated, the cluster affiliation adjusted, and a clearer optimized cluster feature group generated. This process effectively distinguishes different navigation modes.

[0131] Optionally, a weighted average method can be used when integrating intra-cluster feature information.

[0132] For example, for the heading and speed characteristics of ships within a cluster, a weight of 0.6 for heading and 0.4 for speed are assigned, and a weighted average is calculated to generate an integrated feature subset. If the feature consistency is below a threshold, such as a correlation coefficient less than 0.8, missing feature values ​​are filled in using linear interpolation. For instance, if a ship's speed data is missing, it can be filled in by interpolating speed values ​​from consecutive time points, such as interpolating from 10 knots to 11 knots between 10 and 12 knots.

[0133] In this embodiment, principal component analysis is used for dimensionality reduction. For high-dimensional feature subsets, such as those containing multi-dimensional data like speed, heading, and position, principal component analysis is used to extract the principal components, retaining 90% of the variance, thus generating a low-dimensional feature subset and reducing the computational complexity of subsequent steps.

[0134] For example, after dimensionality reduction, only the two principal components of speed and heading are retained, which can still characterize ship behavior.

[0135] For example, cosine similarity calculation is used to evaluate the uniqueness of feature subsets. Assuming two sets of ship feature subsets represent cargo ships and fishing vessels respectively, the cosine similarity between feature vectors is calculated. A value of 0.9 indicates high similarity, requiring further adjustment of the feature extraction method; a value of 0.3 indicates high uniqueness, making it suitable as a target feature subset for subsequent ship classification tasks. This method ensures that feature subsets can effectively distinguish different ship types.

[0136] Step S106: For the separated target feature subset, if there is a high risk of identity misjudgment, the identity label of each ship is determined by comparing with the preset identity template.

[0137] Specifically, an initial template set is obtained by retrieving preset identity templates from a template database using a target feature subset. If the feature consistency between the initial template set and the target feature subset is lower than a preset threshold, the template set is adjusted using a cosine similarity calculation method to obtain an optimized template set. Based on the optimized template set, a feature comparison method is used to calculate the matching degree between the target feature subset and each template, resulting in a matching degree score. If the matching degree score is higher than a preset threshold, a corresponding identity label is assigned to the target feature subset using a label assignment method, resulting in preliminary identity labels. Based on the preliminary identity labels, a risk assessment method is used to calculate the identity misjudgment risk, resulting in a misjudgment risk value. If the misjudgment risk value is higher than a preset threshold, a k-nearest neighbor algorithm is used to perform a proximity comparison between the target feature subset and samples in the template database, resulting in corrected identity labels. Based on the corrected identity labels, a label assignment method is used to update the ship's identity, determining the final ship identity label.

[0138] Specifically, in multi-target, high-density scenarios, ship identification is a critical task, requiring accurate template matching and identity label allocation based on subsets of target features. The following analysis and examples, using specific implementations, focus on relevant technical topics to ensure the content closely aligns with the business domain of ship feature processing and identification.

[0139] For example, when retrieving a preset identity template from a template database, the database can be constructed based on the ship's radar signal characteristics, navigation patterns, etc. Assume the database contains 100 preset templates, each covering the ship's electromagnetic radiation characteristics, speed range, and heading pattern. The initial template set is selected from the database by feature vector comparison, choosing the 10 templates that are closest to the target feature subset.

[0140] It should be noted that the template database needs to be updated regularly to cover new ship features and ensure the coverage of the initial template set.

[0141] In this embodiment, if the feature consistency between the initial template set and the target feature subset is lower than a preset threshold (e.g., 0.8), the template set can be adjusted using the cosine similarity calculation method.

[0142] For example, a target feature subset might include a ship's radar cross-section of 200 square meters and a speed of 20 knots, while the average values ​​of the templates in the initial template set are 180 square meters and 18 knots, respectively. After calculating the cosine similarity, templates with similarity scores below 0.7 are removed, and more similar templates from the database are added to form an optimized template set. This method can improve the matching accuracy between templates and targets.

[0143] For example, a weighted feature matching method can be used for feature comparison. Assume the target feature subset contains three dimensions: radar signal strength, speed, and heading, with weights of 0.5, 0.3, and 0.2, respectively. Each template in the optimized template set is compared dimension-by-dimensionally with the target feature subset, and a matching score is calculated. If a template's score reaches 0.9, which is higher than the preset threshold of 0.85, a match is considered successful. This weighted approach highlights the contribution of key features to identity recognition.

[0144] In this embodiment, the label assignment method can be implemented through a rule engine. If the matching score is higher than a threshold, the system assigns the corresponding template identity label to the target feature subset, such as "cargo ship type A".

[0145] Specifically, if a subset of the target ship's features matches the "cargo ship type A" template, the system automatically assigns the tag and records the matching time and location for subsequent tracking.

[0146] It should be noted that tag assignment needs to be combined with the scene context, such as the marine environment, to avoid misassignment.

[0147] For example, risk assessment methods can calculate the risk of misjudgment through statistical analysis. Suppose a ship's initial identification label is "Cargo Ship Type A," but its speed is abnormally high, exceeding the template definition range. Based on historical data statistics, the system assesses the misjudgment risk of this label at 0.3, which is higher than the threshold of 0.2, triggering further verification. This method can effectively reduce the misjudgment rate.

[0148] In this embodiment, the k-nearest neighbor algorithm is used to correct identity labels.

[0149] For example, the target feature subset is compared with samples in the template database, and the five nearest neighbor samples are selected. If four of these samples point to "Patrol Ship Type B", the label is corrected to "Patrol Ship Type B".

[0150] Specifically, the k-nearest neighbor algorithm uses distance calculations in the feature space to select the sample closest to the target, ensuring the accuracy of label correction.

[0151] For example, the final ship identification label can be updated through multiple rounds of verification. The revised label "Patrol Ship Type B" is compared with the initial label; if the consistency is high, it is updated to the final label and recorded in the system log. This multi-round verification improves the robustness of identification and adapts to complex multi-target scenarios.

[0152] Step S107: Starting from the determined ship identification tags and integrated feature representations, Kalman filtering is used to correct the trajectory prediction error and obtain the predicted trajectory sequence, thereby reducing the problem of large positioning deviation.

[0153] Specifically, the target's motion state is obtained from the ship's identification tag and integrated features. A Kalman filter is used to initialize the state estimation vector, yielding an initial state estimate. Based on this initial state estimate, real-time motion information is extracted from sensor observation data. A Kalman filter prediction step is then used to calculate the target's predicted position, resulting in predicted trajectory points. If the deviation between the predicted trajectory points and the sensor observation data exceeds a preset threshold, a Kalman filter update step is used to adjust the state estimation vector, yielding a corrected state estimate. Based on the corrected state estimate, a trajectory smoothing method is used to optimize the predicted trajectory points, resulting in a smoothed trajectory sequence. Continuous motion features are extracted from the smoothed trajectory sequence, and the trajectory points are fine-tuned through a dynamic correction process, resulting in an optimized trajectory sequence. If the matching degree between the optimized trajectory sequence and the sensor observation data is lower than a preset threshold, the trajectory sequence is further corrected using the least squares method, yielding the final trajectory sequence. Based on the final trajectory sequence, the ship's positioning information is updated, determining the ship's real-time position.

[0154] For example, in real-time ship positioning scenarios, when acquiring the target's motion state based on ship identification tags and integrated features, information such as the ship's latitude, longitude, speed, and heading can be obtained through sensor data such as radar and AIS (Automatic Identification System). The Kalman filter method is used to initialize the state estimation vector, and its core lies in fusing multi-source data to construct the initial state.

[0155] For example, if a ship's initial position is 120.5° longitude, 30.2° latitude, with a speed of 15 knots and a heading of 45°, the predicted position at the next moment can be estimated based on historical motion data through the Kalman filter prediction step. For example, the predicted position after 1 second is 120.51° longitude and 30.21° latitude.

[0156] In this embodiment, data noise needs to be considered when extracting real-time motion information from sensor observation data.

[0157] For example, radar positioning may be affected by weather interference, resulting in a deviation of up to 50 meters. The Kalman filter prediction step combines historical trajectories with current observations to generate predicted trajectory points. If the deviation between the predicted trajectory points and the actual observations exceeds a preset threshold (e.g., 100 meters), the state estimation vector is adjusted through an update step.

[0158] For example, if the actual observed location is 120.52° longitude and 30.22° latitude, updating the state estimate through Kalman filtering can reduce the position deviation to within 20 meters, thereby improving the positioning accuracy.

[0159] Specifically, trajectory smoothing methods can employ sliding window techniques to optimize predicted trajectory points.

[0160] For example, the five most recent prediction points are taken and a smooth trajectory sequence is generated by weighted averaging to reduce the impact of abrupt changes. If the matching degree between the smooth trajectory sequence and the sensor data is lower than a threshold (e.g., matching degree is lower than 80%), a secondary correction is performed using the least squares method.

[0161] For example, given a trajectory sequence, sensor data indicates that the ship is sailing in a straight line, but the predicted trajectory deviates slightly. The least squares method can fit a trajectory that is more in line with the actual route, and the matching degree is improved to more than 90% after correction.

[0162] In this embodiment, when fine-tuning the trajectory points during the dynamic correction process, the ship's dynamic model can be incorporated.

[0163] For example, considering the ship's acceleration limitations, abnormal trajectory points can be adjusted to conform to actual navigation characteristics, avoiding the prediction of unreasonable sharp turns. After the final trajectory sequence is generated, the ship's real-time position can be updated.

[0164] For example, after correction, a ship's longitude was determined to be 120.53°, latitude 30.23°, speed 16 knots, and heading 46°, providing reliable data for subsequent navigation or monitoring.

[0165] For example, in the application of identity tags, by combining identity recognition scenarios in historical dialogues, the motion state can be ensured to be consistent with the identity tag by fusing features (such as ship size and radar reflection characteristics).

[0166] For example, a large cargo ship's tag requires a speed not exceeding 20 knots. If the predicted trajectory shows an abnormal speed, erroneous data can be dynamically corrected and eliminated to ensure the trajectory matches the ship's identity. This multi-faceted collaborative approach improves positioning accuracy and ensures consistency between the identity tag and the ship's movement status.

[0167] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A ship positioning method based on multi-node cooperation, characterized in that, The method comprises the following steps: Obtaining a multi-node original signal sequence and image data, removing noise caused by signal interference by using a preprocessing method to obtain a denoised multi-source data set; Extracting different dimension features of the denoised multi-source data set by a convolutional neural network to obtain a preliminary feature vector group; When the feature dimensions of the preliminary feature vector group are inconsistent, aligning and transforming the time and space scales to obtain an aligned feature vector group; Integrating the dispersed information in the aligned feature vector group by a weighted fusion mechanism to obtain a comprehensive fusion feature representation; Separating the unique modes of each ship by using a clustering method according to the comprehensive fusion feature representation to obtain a separated target feature subset; When there is a high risk of identity misjudgment, matching the preset identity template to determine the identity label of each ship; According to the identity label of each ship and the comprehensive fusion feature representation, the Kalman filter is used to correct the trajectory prediction difference to obtain a predicted trajectory sequence, and the real-time position of the ship is determined according to the predicted trajectory sequence.

2. The method of claim 1, wherein, The denoised multi-source data set comprises: Collecting original signal sequences and image data by radar, optical and infrared sensors, removing noise interference by using a filtering algorithm to obtain a first denoised data set; If the intensity of the signal sequence in the first denoised data set is lower than the preset threshold, the signal sequence is processed by using a mean filtering algorithm to obtain a second denoised signal sequence; If the definition of the image data in the first denoised data set is lower than the preset threshold, the image data is processed by using a Gaussian filtering algorithm to obtain second denoised image data; Fusing the second denoised signal sequence and the second denoised image data by a principal component analysis algorithm to obtain a first fusion data set; According to the first fusion data set, the features of the radar signal, optical image and infrared data are integrated by using a weighted average method to obtain a second fusion data set; If the dimension of the feature vector in the second fusion data set is higher than the preset threshold, the third fusion data set is obtained by processing by using a dimension reduction algorithm; Classifying the third fusion data set by a clustering algorithm to obtain a multi-source data set.

3. The method of claim 1, wherein, The preliminary feature vector group comprises: Extracting an initial feature vector group from the multi-source data set by using a convolutional neural network, the initial feature vector group containing distance information features, speed information features and shape heat features to obtain a preliminary feature set; If the dimension of the feature vector in the preliminary feature set exceeds the preset threshold, the preliminary feature set is processed by using a principal component analysis algorithm to obtain a dimension-reduced feature set; According to the dimension-reduced feature set, the feature vectors are classified by using a support vector machine algorithm to determine the feature categories to obtain a classified feature set; For the classified feature set, the distance information features, speed information features and shape heat features are integrated by using a weighted average method to obtain the final feature vector group.

4. The method of claim 1, wherein, The aligned feature vector group comprises: Obtaining the time scale and the space scale of each feature vector in the preliminary feature vector group, wherein the time scale represents the distribution range or sampling frequency of the feature vector in the time dimension, and the space scale represents the distribution range or resolution of the feature vector in the space dimension; The time scale difference and the space scale difference between each feature vector are calculated by statistical analysis methods, including but not limited to variance analysis and correlation analysis or standard deviation calculation, to quantify the degree of scale difference; If there is a time scale or space scale inconsistency, the feature vectors are aligned and transformed by a linear interpolation method to obtain an aligned feature vector group.

5. The method of claim 1, wherein, The comprehensive fusion feature representation includes: From the aligned feature vector group, the attribute distribution of each feature vector is obtained, the mean and variance are calculated by statistical analysis methods, and the quantitative description of the feature distribution is obtained; According to the quantitative description of the feature distribution, a preset weighting mechanism is used to assign weights, and a weighted average method is used to integrate the dispersed information to obtain the preliminary fusion feature vector; If the fusion accuracy of the preliminary fusion feature vector is lower than the preset threshold, the weight distribution is adjusted by a linear regression method, the weighted average is recalculated, and an optimized fusion feature vector is obtained; From the optimized fusion feature vector, the business relevance feature is extracted, the features are grouped by clustering method, and the comprehensive fusion feature representation is obtained.

6. The method of claim 1, wherein, The separated target feature subset includes: The comprehensive fusion features are grouped by a density clustering method to obtain an initial clustering result; According to the initial clustering result, the feature distribution in each cluster is smoothed by a kernel density estimation method to obtain a smoothed clustering feature group; If the inter-cluster distance of the smoothed clustering feature group is lower than the preset threshold, the inter-cluster boundary is adjusted by the Euclidean distance calculation method to determine the optimized clustering feature group; According to the optimized clustering feature group, the weighted average method is used to integrate the feature information within the cluster to obtain an integrated feature subset; If the feature consistency of the integrated feature subset is lower than the preset threshold, the feature subset is completed by a linear interpolation method to obtain a completed feature subset; According to the completed feature subset, the feature subset is reduced by a principal component analysis method to obtain a reduced feature subset; According to the reduced feature subset, the uniqueness of each ship feature subset is evaluated by a cosine similarity calculation method to determine the final target feature subset.

7. The method of claim 1, wherein, The identity label of each ship includes: From the target feature subset, a preset identity template is obtained from the template database to obtain an initial template set; If the feature consistency of the initial template set and the target feature subset is lower than the preset threshold, the template set is adjusted by a cosine similarity calculation method to obtain an optimized template set; According to the optimized template set, the feature comparison method is used to calculate the matching degree of the target feature subset and each template to obtain a matching score; If the matching score is higher than the preset threshold, the target feature subset is assigned a corresponding identity label by a label assignment method to obtain a preliminary identity label; According to the preliminary identity label, the risk assessment method is used to calculate the identity misjudgment risk to obtain a misjudgment risk value; If the misjudgment risk value is higher than the preset threshold, the k-nearest neighbor algorithm is used to compare the target feature subset with the samples in the template database to obtain a corrected identity label; According to the corrected identity label, the label assignment method is used to update the ship identity to determine the final ship identity label.

8. The method of claim 1, wherein, The determining the real-time position of the ship according to the predicted trajectory sequence comprises: obtaining a target motion state from a ship identity tag and a comprehensive fusion feature, initializing a state estimation vector by using a Kalman filtering method, and obtaining an initial state estimation; extracting real-time motion information from sensor observation data according to the initial state estimation, calculating a target predicted position by a Kalman filtering prediction step, and obtaining a predicted trajectory point; if a deviation between the predicted trajectory point and the sensor observation data exceeds a preset threshold, adjusting the state estimation vector by a Kalman filtering update step, and obtaining a corrected state estimation; optimizing the predicted trajectory point by using a trajectory smoothing processing method according to the corrected state estimation, and obtaining a smoothed trajectory sequence; extracting continuous motion features from the smoothed trajectory sequence, adjusting the trajectory point by a dynamic correction process, and obtaining an optimized trajectory sequence; if a matching degree of the optimized trajectory sequence and the sensor observation data is lower than a preset threshold, performing secondary correction on the trajectory sequence by using a least square method, and obtaining a final trajectory sequence; updating positioning information of the ship according to the final trajectory sequence, and determining the real-time position of the ship.

Images