An intelligent navigation beacon and AIS data cooperative navigation safety improvement method

Abstract

The application belongs to the technical field of navigation safety analysis, and specifically discloses an intelligent navigation mark and AIS data cooperative navigation safety improvement method, which comprises the following steps: collecting channel images through a navigation mark visual sensor and preprocessing, and generating an initial detection frame through multi-scale detection; dynamically configuring an NMS threshold value according to the physical size of a ship, and outputting a final detection frame with a physical size label; extracting a ship visual feature vector, assigning a feature update strategy to multiple ships according to the physical size, predicting the position through a historical trajectory when the ship moves out of the field of view, and triggering re-identification; and finally fusing the visual tracking result and real-time AIS data to generate a channel risk early warning and a navigation strategy. The application solves the problems of missing of a front-end perception layer, missing of an NMS mechanism leading to missed detection, design of a shielded target re-identification and blank of a re-identification triggering logic in the prior art, and comprehensively improves the accuracy and reliability of navigation safety guarantee.

Description

Technical Field

[0001] This invention belongs to the field of navigation safety analysis technology, and more specifically, relates to a method for improving navigation safety through the collaboration of intelligent navigation marks and AIS data. Background Technology

[0002] With the growth of global shipping volume and port expansion leading to increased waterway density and frequent vessel convergence, coupled with the higher demands placed on navigation safety by special waterways such as bridges and construction zones, traditional navigation marks and AIS operations alone are insufficient to cope with dynamically changing risks. Therefore, coordinated operation is necessary to achieve accurate early warning.

[0003] Prior art 1, such as the virtual electronic waterway design method and system based on AIS technology disclosed in Chinese invention patent application No. 202510341710.9, corrects the ship's position through multi-source data fusion, generates an initial waterway based on Bézier curves, and achieves dynamic adjustment through environmental correction terms and Kalman filtering, finally transmitting the data via VHF+5G dual-channel transmission. This ensures safe navigation even without physical navigational aids and is suitable for ship navigation and safety assurance in extreme environments such as freezing temperatures and typhoons.

[0004] Prior art 2, such as the ship route planning method, apparatus, electronic device and storage medium disclosed in Chinese invention patent application No. 202111282692.X, involves rasterizing the navigation chart and marking obstacle grids based on wind speed and wave height thresholds. For non-ocean-going scenarios, the shortest path is generated using the grid A* algorithm. For ocean-going scenarios, obstacles on the great circle route are detected in segments and locally replanned. Dynamic updates are triggered by the ship's position to avoid weather threats in real time and improve the safety and efficiency of ship navigation.

[0005] Regarding the existing technical solutions mentioned above, it is clear that the current focus on backend route planning still has the following problems: 1. Existing patents have serious deficiencies in the front-end perception layer, especially lacking accuracy in obstacle recognition and robustness in tracking, such as the lack of depth design for re-identifying occluded targets.

[0006] 2. None of the existing patents mention the NMS mechanism. They use a fixed threshold by default. When ships are densely packed, the detection boxes are excessively merged, which suppresses distant targets of small boats and easily leads to missed detections.

[0007] 3. Patent 1 does not involve target tracking in its route planning, and Patent 2 only predicts the AIS ship position through Kalman filtering. Neither of them involves re-identification triggering logic and cannot handle the situation when the target is occluded. Summary of the Invention

[0008] In view of this, in order to solve the above problems, a method for improving navigation safety by coordinating intelligent navigation marks and AIS data is proposed.

[0009] The objective of this invention can be achieved through the following technical solution: This invention provides a method for improving navigation safety through the collaboration of intelligent navigation marks and AIS data. The method includes: S1, acquiring a continuous image sequence of the waterway through a visual sensor mounted on the navigation mark and performing preprocessing.

[0010] S2. Identify ship targets in the image using a multi-scale target detection model and generate an initial set of detection boxes.

[0011] S3. Perform target determination operation on the initial detection box set: dynamically configure the NMS threshold according to the physical size of the ship target, wherein the physical size consists of physical length and width.

[0012] The final detection box with physical size label is output based on the configured NMS threshold.

[0013] S4. Extract the visual feature vectors of the ship targets within the final detection box, and assign feature update strategies to multiple ship targets that appear simultaneously based on the physical size labels.

[0014] S5. When the ship target moves out of the field of view, predict the current location area based on its historical movement trajectory, trigger ship target re-identification, and then return to step S3.

[0015] S6. Integrate visual tracking results of ship targets with real-time AIS data to generate waterway risk warnings and navigation strategies.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention can capture continuous image sequences and preprocess them by means of the visual sensor mounted on the navigation mark, and combine them with a multi-scale target detection model to more accurately capture ship targets in the waterway, solve the problem of insufficient accuracy of the front-end perception layer in identifying obstacles, and significantly improve the perception capability of various types of ships.

[0017] (2) By dynamically configuring the NMS threshold according to the physical size of the ship target, the present invention outputs the final detection box with the physical size label, which can avoid excessive merging of detection boxes when ships are dense, prevent small boats and distant targets from being suppressed, solve the problem of missed detection caused by fixed threshold, and improve the integrity of target detection.

[0018] (3) This invention extracts the visual feature vector of the ship target and assigns feature update strategies to multiple ships that appear at the same time based on physical size labels. At the same time, it combines the mechanism of predicting the location area based on the historical trajectory and triggering re-identification when the target moves out of the field of view. This can effectively handle the situation where the target is occluded, make up for the lack of target tracking and re-identification triggering logic in the prior art, and enhance the robustness of target tracking.

[0019] (4) By integrating the visual tracking results of ship targets with real-time AIS data, this invention effectively combines front-end precise perception with back-end data, making the generated waterway risk warning and navigation strategy more comprehensive and reliable, thereby improving the ability to cope with dynamic changes in risks and ensuring the safety and efficiency of ship navigation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall implementation process of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the main process for determining the physical dimensions of this invention.

[0023] Figure 3 This is a schematic diagram of the physical length determination process of the present invention.

[0024] Figure 4 This is a schematic diagram of the physical width determination process of the present invention.

[0025] Figure 5 This is a schematic diagram of the NMS threshold dynamic configuration process of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1 As shown, the present invention provides a method for improving navigation safety through the collaboration of intelligent navigation marks and AIS data. The method includes: S1, acquiring a continuous image sequence of the waterway through a visual sensor mounted on the navigation mark and performing preprocessing.

[0028] Understandably, the visual sensors on smart navigation aids, such as high-definition cameras, continuously capture images of the waterway at a set frame rate, for example, 30 frames per second, thereby obtaining a continuous sequence of images. These images cover information such as ships in the waterway, the surrounding environment, and potential obstacles.

[0029] It should be added that the image preprocessing involves multiple steps, including grayscale conversion, filtering and denoising, and image enhancement. Filtering and denoising can employ methods such as Gaussian filtering and median filtering to remove noise interference from the image. Image enhancement can utilize techniques such as histogram equalization and contrast-limited adaptive histogram equalization.

[0030] S2. Identify ship targets in the image using a multi-scale target detection model and generate an initial set of detection boxes.

[0031] In one specific embodiment, the multi-scale target detection model can be a YOLO series model or a Faster R-CNN model. The model is trained on a waterway image dataset that includes ship categories, visual physical dimensions, and location annotations. Moreover, multi-scale target detection models such as Faster R-CNN and YOLO series can perform feature extraction and target recognition on images at different scales, making it easy to adapt to the different sizes and proportions of ships that may appear in the image.

[0032] It should be added that the model training can be achieved by collecting a large dataset of images of waterways containing ship targets, and labeling the ships in the images. The labeling includes the ship's category, such as cargo ship, passenger ship, fishing boat, etc., as well as its position and actual size information in the image. Using this labeled data, a selected target detection model is trained, and by continuously adjusting the model's parameters, it can accurately identify ship targets in the images.

[0033] Furthermore, the specific generation process of generating the initial detection box set is as follows: A1. The size and pixel standardization of the preprocessed image sequence are performed and input into the trained multi-scale target detection model.

[0034] A2. Extract multi-scale feature maps through convolutional neural networks and perform feature fusion. Based on the fused features, identify the visual length and visual width of ship targets.

[0035] A3. Based on the comparison results between the aspect ratio and the preset threshold, select a rectangular or square anchor frame group to generate a preset anchor frame.

[0036] A4. Based on the current wave height level, dynamically configure the crossover ratio threshold. For each identified ship target, calculate its crossover ratio with all preset anchor frames, and mark the anchor frames with crossover ratios greater than the first preset threshold as positive sample anchor frames.

[0037] A5. Import the current surge level and dynamically configure the crossover ratio threshold. For each identified ship target, calculate its crossover ratio with all preset anchor frames, and mark the anchor frames with crossover ratios greater than the corresponding configured crossover ratio threshold as positive sample anchor frames.

[0038] A6. Output the class probability of the ship target in each positive sample anchor frame through the convolutional layer, and record the positive sample anchor frames to which the ship target with the class probability exceeds the standard as candidate frames.

[0039] A7. Calculate the product of the class probability of the ship target in each candidate box and the corresponding intersection-union ratio as the confidence level of the corresponding candidate box.

[0040] A8. Filter out candidate boxes with confidence scores below the threshold, and summarize the remaining candidate boxes according to the scale of their feature maps to form an initial set of detection boxes.

[0041] It's worth noting that the size normalization in step A1 primarily involves scaling or padding the image based on the model design. For example, YOLOv5 defaults to 640×640, while Faster R-CNN commonly uses 800×1333, ensuring a fixed input size. For instance, a 1920×1080 waterway image can be scaled proportionally to 640×360, then padded with black pixels on both sides to reach 640×640 to prevent distortion of ship targets. Pixel normalization can be achieved by converting pixel values ​​from the 0-255 range to 0-1 or -1-1. For example, the YOLO series uses division by 255, while Faster R-CNN uses mean subtraction, eliminating the impact of lighting intensity differences on the model and accelerating parameter convergence.

[0042] For example, using YOLOv5, the input image is downsampled five times through the CSPDarknet53 backbone network. Each downsampling reduces the size by half, generating feature maps at five scales, such as 640×640, 320×320, 160×160, 80×80, 40×40, and 20×20. Smaller feature map scales result in a larger receptive field, corresponding to distant areas and suitable for detecting large ships. Larger scales result in a smaller receptive field, corresponding to near details and suitable for detecting small ships.

[0043] Understandably, the A2 step can be implemented by fusing multi-scale feature maps using Neck structures such as PANet. For example, combining 20×20 high semantic features with 80×80 high resolution features can preserve the detailed features of the ship's outline and texture, while also incorporating the contextual information of the waterway background, such as the water surface and shoreline, thereby improving the detection rate of small targets, such as fishing boats.

[0044] Furthermore, the process of identifying visual length and visual width in step A2 is as follows: Assuming that the focal length of the navigation beacon camera is 8mm, the installation height is 10 meters, and the downward angle is 30°, the visual length and visual width of the ship target can be obtained by calculating the actual distance corresponding to 1 pixel based on the pixel width and pixel length using the geometric projection formula.

[0045] It should be added that, in this scenario, the geometric projection formula for calculating the actual distance using pixel size can be based on the pinhole imaging principle and trigonometric function relationships, and the focal length must first be converted to a unit consistent with the installation height. The physical size of the pixel is determined by the camera sensor parameters.

[0046] The specific formula for the geometric projection is as follows: ,in, Indicates the actual distance. Indicates pixel size, Indicates the installation height. Indicates the angle of depression. Indicates focal length. This represents the physical size of a pixel. In this formula, The vertical projection distance from the camera to the target's location corresponds to the effective component of the object distance on the horizontal plane. Corresponding to the imaging scale parameters on the image plane, This ratio constitutes the conversion coefficient between pixel size and actual distance, which is ultimately obtained by multiplying the pixel size (i.e., the number of pixels occupied by the ship target in the image) by the conversion coefficient of actual distance.

[0047] Understandably, regarding step A3, if the aspect ratio exceeds the preset threshold, the target is determined to be closer to a rectangle, and a rectangular anchor frame group is selected to generate a preset anchor frame. If it does not exceed the threshold, the target is determined to be closer to a square, and a square anchor frame group is selected to generate a preset anchor frame.

[0048] The determination of the preset threshold corresponding to the aspect ratio can be based on the statistical distribution of the actual aspect ratio of the ship samples and combined with the iterative optimization of the model detection effect to distinguish between rectangular and square ship shapes. For example, if statistics show that 85% of cargo ships in the waterway have an aspect ratio of 5:8 and 15% of near-shore ships have an aspect ratio of 2:3, the initial threshold can be set to 3 and adjusted to 2.8 after verification. That is, rectangular anchor frame groups such as [100,20] and [150,30] are selected when the aspect ratio is greater than 2.8, and square anchor frame groups such as [50,40] and [80,60] are selected when the aspect ratio is less than or equal to 2.8.

[0049] It should be added that the aspect ratio and anchor frame size are matched by calculating similarity and setting a matching similarity threshold. Anchor frame sizes that exceed the matching similarity threshold are used as matching anchor frame sizes. The similarity between the ship pixel size and each preset anchor frame size is usually calculated using the intersection-union ratio (IUU) method or the Euclidean distance method. The IUU method is obtained by treating the ship pixel size as a rectangle and calculating the IUU with the anchor frame size. Both the IUU method and the Euclidean distance method use existing similarity calculation methods. Their specific calculation formulas are existing empirical formulas and will not be shown.

[0050] For example, a matching similarity threshold is set, such as an intersection-union ratio (IU) greater than or equal to 0.6, or an Euclidean distance less than 50 pixels, and anchor frames that meet the threshold are selected. If there are multiple candidate anchor frames, the one with the highest similarity is selected as the matching result. For example, if the IU of a ship's pixel size (90,18) with the anchor frame (80,16) is 0.8, which is higher than the IU of the anchor frame (100,20) with 0.75, then the former is matched.

[0051] It is important to note that if partial occlusion of the vessel leads to inaccurate estimation of its actual size (e.g., only the bow is identified), the pixel size will be corrected by combining the average size of similar vessels in historical trajectories or size data broadcast by AIS before matching. If the pixel size of the same vessel differs at different distances, the matching process will automatically associate it with feature maps of the corresponding scale. For example, a small vessel in the distance will be matched with a small-scale anchor frame, corresponding to a high-resolution feature map.

[0052] It should also be noted that the preset anchor frames are designed based on the pixel dimensions of common ships in the waterway before model training. They need to cover ships of different sizes and shapes. For example, the anchor frames for small ships are typically [30,6] or [40,8], for medium-sized ships [80,16] or [100,20], and for large ships [200,40] or [300,60]. The aspect ratio is 0.2. For example, [30,6] refers to a width of 30 pixels and a length of 6 pixels.

[0053] In one specific embodiment, the specific execution rule for dynamically configuring the cross-union ratio (CUI) threshold in step A4 is as follows: Surge level data of the waterway is collected in real time by environmental sensors mounted on the navigation beacon and imported into the navigation system. Assuming the navigation system has a preset surge level threshold of 3, when a surge level greater than 3 is detected, it is determined to be a high-wave environment. At this time, the ship target in the image is prone to blurring or edge distortion due to wave movement. The CUI threshold is dynamically adjusted to 0.5 to increase the number of positive sample anchor frames and avoid missed detections. When the surge level is less than or equal to 3, the waterway environment is stable, and the ship target image is clear. The CUI threshold is maintained at 0.7 to strictly screen high-quality anchor frames, improve detection accuracy, and achieve adaptive optimization of anchor frame matching under different surge conditions. Furthermore, the specific value is based on the degree of influence of surges on ship image features and the optimal balance value obtained through extensive experimental verification.

[0054] It should also be noted that, generally, anchor boxes with an intersection-over-union (IoU) ratio greater than a threshold such as 0.5 are marked as positive samples and are responsible for predicting the target; IoU between 0.2 and 0.5 are ignored samples; and those less than 0.2 are negative samples and are responsible for predicting the background. A first preset threshold of 0.5 can be set as a specific reference value, and can be dynamically adjusted according to the actual scenario.

[0055] Understandably, in step A6, when outputting the class probability of the ship target through the convolutional layer, the classification accuracy is also optimized using the cross-entropy loss function. The cross-entropy loss function is an existing function and will not be described in detail here.

[0056] Understandably, the confidence threshold in step A7 needs to be set in conjunction with the characteristics of the ship detection scenario and the model performance. It should be determined by statistically analyzing the confidence distribution of candidate boxes and verifying it through experiments. For example, first, the confidence distribution of ship targets and background in the sample is statistically analyzed to determine the critical value for distinguishing real targets from false detections, such as 0.2. Then, the detection accuracy and recall rate under different thresholds are tested on the validation set. Finally, a threshold that balances the false negative rate and the false positive rate is selected, such as 0.25-0.3. In complex environments such as high waves, the threshold can be appropriately reduced, such as to 0.2, to reduce false negatives, while in stable scenarios such as clear skies, the threshold can be increased, such as to 0.35, to reduce false positives.

[0057] Understandably, the reference example summarized in step A8 is as follows: 30 candidate boxes for medium-sized ships are retained on the 80×80 feature map, and 15 candidate boxes for large ships are retained on the 20×20 feature map. The final set contains 45 initial detection boxes, each of which contains location coordinates, confidence score and predicted class information.

[0058] It should be noted that by adapting multi-scale features to different ship sizes, the anchor frame mechanism covers the diversity of target shapes, and finally, invalid candidate boxes are initially eliminated through confidence screening. This process provides basic data for subsequent NMS optimization, ensuring the recall rate and initial accuracy of ship target detection.

[0059] S3. Perform target determination operation on the initial detection box set: dynamically configure the NMS threshold according to the physical size of the ship target, wherein the physical size consists of physical length and width.

[0060] The final detection box with physical size label is output based on the configured NMS threshold.

[0061] Understandably, the label includes a physical length value, a physical width value, a size level identifier, and an assigned NMS threshold.

[0062] Specifically, please refer to Figures 2 to 4 As shown, the specific process for determining the physical dimensions is as follows: B1. Analyze the ship's AIS signal to obtain the AIS length and AIS width.

[0063] B2. Calculate the length difference between the visual length and the AIS length, and the width difference between the visual width and the AIS width.

[0064] B3. If the length difference is less than or equal to the preset first length difference threshold, then the physical length is the AIS length.

[0065] B4. If the length difference is greater than the first preset length difference threshold and less than or equal to the preset second length difference threshold, then detect the current visibility, fuse the AIS length and visual length according to the visibility matching length fusion rule, and output the fused length as the physical length, and use the fused length as the physical length of the ship target.

[0066] B5. If the length difference is greater than the preset second length difference threshold, collect the current wave height and match the preset wave height compensation model, and use the final length output by the model as the physical length.

[0067] B6. If the width difference is less than or equal to the preset first width difference threshold, then the physical width is the AIS width.

[0068] B7. If the width difference is greater than the first preset width difference threshold and less than or equal to the preset second width difference threshold, the AIS width and visual width are fused according to the preset width fusion rule, and the physical width is the fused width.

[0069] B8. If the width difference is greater than the width difference threshold, then the physical width is the visual width.

[0070] B9. Based on the physical length and physical width, output the final physical dimensions.

[0071] Understandably, the difference in step B2 is calculated as follows: the relative deviation between the AIS length and the visual length is calculated to obtain the length difference, and the relative deviation between the AIS width and the visual width is calculated to obtain the width difference.

[0072] Understandably, the visibility mentioned in step B4 can be obtained directly using meteorological sensors such as visibility meters installed near navigation marks or cameras to acquire real-time visibility data. If no dedicated sensor is provided, visibility conditions can also be indirectly inferred by analyzing image features captured by the camera, such as image contrast, edge sharpness, and the degree of blurring of distant targets, and by combining image sharpness evaluation algorithms, such as sharpness detection based on the Laplacian operator and atmospheric light estimation based on fog image models.

[0073] It is also understood that the specific rules of the preset length fusion rule in step B4 are as follows: 1) When the visibility is greater than 1000 meters, the first allocated length fusion weight is matched, and the weighted sum of the visual length, AIS length and fusion weight is used as the fusion length.

[0074] For example, in the first allocation length fusion weight, the visual length and AIS length are set to 0.7 and 0.3, respectively.

[0075] 2) When the visibility is less than or equal to 1000 meters, the second allocation length fusion weight is matched, and the weighted sum of the visual length, AIS length and fusion weight is used as the fusion length.

[0076] For example, in the second allocation length fusion weight, the visual length and AIS length are set to 0.4 and 0.6, respectively.

[0077] It is important to note that, based on the influence of visibility on the accuracy of visual length detection, extensive experimental verification has determined that 1000 meters is the critical value for visual imaging clarity. When the distance is greater than 1000 meters, the reliability of visual data is high, while when the distance is less than or equal to 1000 meters, the accuracy decreases due to interference such as fog. The fusion weight allocation of 0.7 and 0.3, and 0.4 and 0.6 is determined by comparing the error distribution of visual and AIS data under different visibility conditions. Through multiple fusion experiments, the optimal ratio that minimizes the fusion length error has been selected, which can achieve a reasonable adaptation of data reliability under different visibility conditions.

[0078] Understandably, the specific content of the wave height compensation model described in step B5 is as follows: The AIS length and visual length are respectively denoted as... and Calculate the compensation length , , This is the current wave height.

[0079] If the wave height is less than or equal to 4 meters, the compensation length is directly output as the final length.

[0080] If the wave height is greater than 4 meters, initiate a binocular vision remeasurement, which involves capturing a sequence of images of the ship's waterline. The remeasurement length is calculated based on a draft-pixel ratio model obtained by establishing a proportional relationship between the actual value of the ship's draft and the corresponding pixel length in the image. The actual value of the ship's draft is obtained through AIS data or ship files.

[0081] Calculate the final length , .

[0082] It should be noted that the specific values ​​mentioned above, such as 0.2, 3, 4 meters, 0.6, and 0.4, are all determined based on the scenario characteristics of the fusion of ship visual inspection and AIS data, combined with a large number of sample statistics and experimental verification. 0.2 is an empirical coefficient for balancing the weights of visual and AIS data. 3 meters and 4 meters are the critical values ​​for the impact of wave height on the accuracy of visual inspection. The deviation is controllable within 3 meters, and needs to be further verified above 4 meters. 0.6 and 0.4 are the fusion weights of the compensation length and the binocular remeasured length in high wave conditions. The whole system is optimized through multiple sets of comparative experiments to minimize the length detection error under different wave conditions.

[0083] It should be noted that wave height can be directly measured by using wave sensors deployed in the water, such as acoustic Doppler wave meters and pressure wave sensors. Wave height can also be retrieved by analyzing and inverting sea surface reflection signals using radar remote sensing technology. In scenarios with video monitoring, wave height can also be estimated by combining image processing algorithms to identify sea surface wave characteristics from sequential images.

[0084] Understandably, the preset width fusion rule in step B7 is as follows: if the detected ship speed is less than or equal to a preset ship speed threshold, the first allocated width fusion weight is matched; if the detected ship speed is greater than the preset ship speed threshold, the second allocated width fusion weight is matched. The weighted sum of the visual width, AIS width, and fusion weight is used as the fusion width degree. For example, the visual width and AIS width in the first and second allocated width fusion weights are 0.6 and 0.4, and 0.3 and 0.7, respectively.

[0085] It should be noted that experiments have verified that 5 knots is the critical value for the stability of visual imaging of ship width. That is, as a preset ship speed threshold, when it is less than or equal to 5 knots, the width imaging is clear and the error is small. When it is greater than 5 knots, the visual detection deviation is easily increased due to motion trailing.

[0086] Understandably, the fusion weight allocation of 0.6 and 0.4, and 0.3 and 0.7 is the optimal ratio that minimizes the fusion width error by analyzing the error distribution of visual width and AIS width at different flight speeds and screening through multiple fusion experiments.

[0087] It should be noted that the setting of various preset thresholds in the process of determining physical dimensions is based on the statistical distribution of differences between ship visual and AIS size data, the degree of influence of the environment on data reliability, and experimental verification. Specifically, this can be achieved by analyzing the difference patterns between visual and AIS length or width in a large number of samples.

[0088] For example, a first difference threshold, such as 5% for length and 8% for width, can be set as the critical value for reliable AIS data; that is, differences within this range are directly represented by AIS values. A second difference threshold, such as 15% for length and 20% for width, can be set as the critical value for fusion or compensation. Differences exceeding this range indicate significant data conflicts.

[0089] It should be noted that the above physical dimension setting process fully combines the stable data of AIS signals with the real-time nature of visual recognition. Through differential threshold grading processing, AIS data, fused data combined with visibility, model output length and AIS data combined with wave height compensation, fused data with preset rules, and visual width are used at different differential levels to achieve accurate determination of the ship's physical dimensions, taking into account both data reliability and environmental adaptability.

[0090] It should also be noted that the NMS thresholds corresponding to the physical length and physical width can be obtained by referring to the preset NMS threshold allocation table based on the physical length and physical width.

[0091] Understandably, the preset NMS threshold allocation table is divided into multiple ranges based on the numerical range of the ship's physical length and physical width. For example, the physical length can be divided into multiple ranges such as 0-30m, 30-60m, 60-100m, and over 100m, and the physical width can be divided into multiple ranges such as 0-10m, 10-20m, 20-50m, and over 50m. The horizontal direction represents the width range, and the vertical direction represents the length range. The NMS threshold is set for the cross cells according to the principle that the larger the size, the larger the threshold. Furthermore, the range division and threshold value are determined in conjunction with the actual ship size distribution and the characteristics of the detection frame.

[0092] Further, please refer to Figure 5 As shown, the specific configuration process of the dynamically configured NMS threshold is as follows: E1. Determine the length classification standard and width classification standard according to the ship's AIS type. The levels include extra-large, large, medium and small.

[0093] Understandably, the AIS type classification standard is set based on the differences in different ship operating environments, size characteristics, and collision risks, and is used to adapt to the different requirements of ship target detection for threshold strictness. Taking the physical length classification standard as an example, the classification example is as follows: For ocean-going cargo ships, the length is greater than the length threshold of ultra-large ships and is classified as ultra-large; for large ships, the minimum length ≤ length < the length threshold of ultra-large ships and is classified as large. For inland waterway ships, the length ≥ the length threshold of large inland waterway ships and is classified as large; for medium-sized ships, the minimum length ≤ length < the length threshold of large inland waterway ships and is classified as medium. For fishing boats, the length < the length threshold of small ships and is classified as small.

[0094] E2. The physical length and physical width are classified separately. If there is an extra-large level in the length level or the width level, the initial allocation threshold is the first preset NMS threshold.

[0095] It should be added that the specific value of the first preset NMS threshold needs to be set in combination with the size characteristics, density and detection accuracy requirements of ultra-large ships in the ship target detection scenario. It is usually set to a low value, such as 0.3-0.5. Its setting mainly considers the large-area overlap characteristics that ultra-large ships are prone to generate in the image, the error tolerance requirement to avoid missed detection, and the differentiated adaptation with the detection threshold of other ship levels, so as to accurately screen candidate boxes and reduce false deletions.

[0096] In ship target detection, the NMS threshold applicable to the current ship is dynamically determined using the aforementioned rules to optimize the candidate box filtering effect. For example, for very large ships, a stricter NMS threshold, such as 0.25, is used to reduce redundant boxes. When the length classification is more stringent, the threshold is appropriately lowered to enhance filtering accuracy. That is, 0.25 can be used as a reference value for the first preset NMS threshold.

[0097] E3. Otherwise, determine the initial allocation threshold according to the following rules: E31. If the length and width levels are the same, take the average of the corresponding NMS thresholds as the initial allocation threshold.

[0098] The method of selecting the average of the two values ​​as the final threshold takes into account the characteristics of both length and width dimensions, making the final threshold more in line with the overall inspection requirements of ships of this class.

[0099] E32. If the length level is higher than the width level, when the length-corresponding NMS threshold is greater than the width-corresponding NMS threshold, the length-corresponding NMS threshold is taken as the initial NMS threshold; otherwise, the difference between the length-corresponding NMS threshold and the preset compensation interference effect value is taken as the initial NMS threshold.

[0100] Understandably, when the length level is higher than the width level, the features of ship targets in the length dimension are more significant and have a greater impact on detection accuracy. Therefore, the length-corresponding NMS threshold is used as the benchmark. If the length-corresponding NMS threshold is not greater than the width-corresponding NMS threshold, it indicates that there may be interference factors in the width dimension, such as ship roll or shooting angle deviation. By subtracting the preset compensation interference impact value to adjust the threshold, the impact of interference on target selection can be reduced, ensuring that more realistic candidate boxes are retained.

[0101] It should be added that the preset compensation interference impact value is usually set based on the degree of width interference commonly found in ship inspection scenarios, such as the impact of different wave heights and changes in ship attitude on width recognition. It is generally an empirical value such as 0.05-0.2. The specific value needs to be calibrated in combination with the deviation pattern of length and width thresholds in historical inspection data.

[0102] E33. If the length level is lower than the width level, take the NMS threshold corresponding to the width as the initial NMS threshold.

[0103] The width level is higher, and its threshold is more lenient, which meets the requirements of higher-level detection tolerance and does not require additional adjustment.

[0104] E34. Determine whether the visibility exceeds the preset threshold. If it does, the final NMS threshold is the initial assigned threshold.

[0105] Understandably, the aforementioned visibility preset threshold is typically set to 1000 meters.

[0106] E35. Otherwise, calculate the compensation value according to the ratio of visibility to the preset threshold, and use the sum of the initial allocated threshold and the compensation value as the final NMS threshold.

[0107] Understandably, the compensation value is obtained by setting a maximum permissible compensation value and multiplying the ratio of visibility to its preset threshold by the maximum permissible compensation value. Furthermore, the setting of the maximum permissible compensation value needs to consider the effective adjustment range of the NMS threshold in the ship inspection scenario, the impact of extreme visibility values ​​on detection accuracy, and the balance between compensation effect and false positive or false negative rates in historical data. It is typically set to 0.05-0.1, and the specific value needs to be calibrated through multiple experiments.

[0108] Furthermore, configuring the NMS threshold also includes the following steps: establishing a sliding window with a preset frame length, and statistically analyzing the maximum, minimum, and average number of ship targets within the window.

[0109] If the ratio of the difference between the maximum and minimum values ​​to the average number of target ships exceeds a preset threshold, then the preset safety NMS threshold will be used as the final allocation NMS threshold.

[0110] Understandably, the preset frame length needs to be set in conjunction with the motion characteristics of the ship target in the video sequence, such as sailing speed and turning frequency, and also needs to be considered in terms of the dynamic changes in the detection scene and computational resource limitations, typically ranging from 5 to 30 frames. For example, for high-speed ships or dynamic and complex scenes, a shorter frame length, such as 5 to 10 frames, can be set to quickly capture fluctuations in the number of targets, while a longer frame length, such as 15 to 30 frames, can be set to smooth out instantaneous noise and balance real-time performance with statistical stability.

[0111] It is also understandable that the proportion exceeding the preset threshold is usually set to 30%, which corresponds to the critical value of the fluctuation of the number of targets in consecutive frames, and is used to determine whether the detection is interfered with.

[0112] It should be noted that the preset security NMS threshold can usually be set to 0.5, and the security threshold of 0.5 is an empirical value. It is neither too strict to detect missed detections nor too lenient to detect false positives. In abnormal scenarios, it prioritizes the detection of at least the target to avoid the system making decision-making errors due to misjudgment.

[0113] S4. Extract the visual feature vectors of the ship targets within the final detection box, and assign feature update strategies to multiple ship targets that appear simultaneously based on the physical size labels.

[0114] Specifically, the visual feature vector is generated through a dual-channel convolutional network, and the specific generation steps are as follows: the first channel extracts the geometric topological features of the hull outline through the directional gradient histogram operator.

[0115] The second channel extracts texture-invariant features of the ship's waterline using a local binary mode operator.

[0116] The features extracted from the dual channels are fused to generate rotation- and scale-invariant feature vectors, which are then used as the visual feature vectors.

[0117] Understandably, in the scenario of the ship's waterline, texture invariance features are specifically manifested in the inherent stability of the texture pattern of the waterline region, such as the direction of lines, the alternation of light and dark, and the repetitive structure of local details. That is, no matter if the ship rotates or scales in the image, or the brightness changes due to changes in lighting, the core texture features of the waterline, such as the relative distribution of lines and the periodicity of texture, can remain consistent and thus be stably extracted by the algorithm, providing a reliable texture basis for tasks such as ship recognition and matching.

[0118] Understandably, during the fusion process, the geometric and topological features of the hull outline extracted from the first channel, such as the relative positions of vertices and the distribution of outline curvature, are first normalized using an affine transformation to eliminate the influence of rotation and scale changes on the outline structure. Simultaneously, the waterline texture-invariant features of the second channel, such as the local texture gradient direction and the frequency features of repeating patterns, are processed using multi-scale pyramid sampling and rotation-invariant filtering to ensure the consistency of texture information under different angles and scaling ratios. Subsequently, the two are combined through feature concatenation, and feature mapping and dimensionality compression are performed using convolutional layers to output the final feature vector. The multi-scale pyramid sampling and rotation-invariant filtering involved in the fusion process all employ existing techniques and will not be described in detail here.

[0119] It should be noted that the directional gradient histogram operator and the local binary mode operator are existing operators and will not be described in detail.

[0120] Furthermore, the specific allocation steps of the feature update strategy are as follows: calculate the update weight of each ship target based on its physical size.

[0121] The updated priority weight is obtained by matching a preset compensation coefficient with the size level identifier and then compensating for the updated weight based on the compensation coefficient.

[0122] The preset feature vector update frequency is matched based on the updated priority weight corresponding to each ship target, and the matched preset feature vector update frequency is assigned to the corresponding ship target.

[0123] Understandably, the specific calculation process of the updated weight is as follows: the current ship speed is collected, and the speed, physical length and physical width are normalized respectively. The normalization results are then weighted and summed based on preset weights to output the final updated weight. The normalization can be performed using the minimum-maximum normalization method. Narrow-body ships are given priority due to their high flexibility and the difficulty in predicting potential risks. The final normalized physical width is obtained by subtracting the physical width value after minimum-maximum normalization from 1.

[0124] It should be noted that in scenarios such as ship supervision and obstacle avoidance, the physical length of a ship is usually more directly related to navigation safety and regulatory priority. Large ships have greater inertia and are more difficult to maneuver, resulting in a much higher risk of collision or regulatory complexity than small ships. The width of a ship is a static attribute related to its adaptability to the water environment, and it has more basic reference value for the stable determination of priority than the dynamic speed. Therefore, the influence weight of physical length is set as follows: > influence weight of physical width > influence weight of speed.

[0125] Understandably, the preset compensation coefficient is set based on the correlation between the ship size class identifier, such as large, medium, and small, and the priority deviation in actual scenarios. For different size classes, the deviation between the actual priority and the basic weight of the ship in terms of collision risk and regulatory requirements can be analyzed in historical data. For example, large ships need a higher priority compensation coefficient of 1.2 due to their greater inertia, while small ships are easily overlooked and the compensation coefficient is set to 0.8. Through experimental verification, the priority weight coefficient value that can reduce the gap between theoretical weight and actual needs is selected.

[0126] In one specific embodiment, the preset feature vector update frequency can be determined based on the priority weight of the updated ship target, combined with the timeliness requirements of the feature vector and the cost of computing resources. High-priority ships need more timely feature updates to accurately track dynamic changes, such as setting it to 50ms per update. Medium-priority ships balance real-time performance and resource consumption, such as setting it to 100ms per update. Low-priority ships reduce the update frequency to save resources, such as setting it to 200ms per update. By dividing the frequency into multiple levels and matching the weight range, the timeliness of the features of high-priority targets can be guaranteed, while avoiding low-priority targets from occupying too many computing resources, thus achieving optimized resource allocation.

[0127] S5. When the ship target moves out of the field of view, predict the current location area based on its historical movement trajectory, trigger ship target re-identification, and then return to step S3.

[0128] Specifically, the prediction process for the current location area is as follows: extract the historical motion trajectory point set of the ship target and calculate the heading undulation.

[0129] If the heading fluctuation is less than the set angle threshold, the output straight-line extension prediction model is the final prediction model; if the heading fluctuation is greater than or equal to the set angle threshold, the output maneuvering and steering prediction model is the final prediction model.

[0130] Based on the predicted location output by the prediction model, a rectangular prediction region is generated with the predicted location as the center, and the rectangular prediction region is marked as the predicted current location region.

[0131] Understandably, heading erraticity can be calculated by first using the arctangent function to calculate the heading angle of adjacent trajectory points, i.e., the angle between the line connecting the two points and the reference direction, such as due north or horizontal to the right of the image. Then, the degree of erraticity of all adjacent heading angles is calculated. The standard deviation of the angle or the maximum angle difference is usually used as the heading erraticity index to quantify the stability of the ship's motion direction.

[0132] Understandably, the angle threshold is adjusted according to the type of vessel; for example, the threshold for large cargo ships that turn slowly can be set to a smaller value. The steering flexibility threshold for small speedboats can be set to a relatively large value, such as... .

[0133] Understandably, the side length of the region is directly proportional to the velocity, as shown in the formula: , The scaling factor is typically set between 0.5 and 2, and is adjusted according to the required prediction accuracy. For faster speeds... Take a larger value to expand the coverage area; when the speed is slow... To improve positioning accuracy, a smaller value is selected. For speed, Let be the side length of the region.

[0134] It should be noted that the straight-line extension prediction model is built upon the assumption of uniform linear motion of the ship: First, the position coordinates and timestamps of the most recent consecutive frames (e.g., frames 10-20) are extracted from the historical trajectory point set. The equation of motion for the straight line is fitted using the least squares method, and the average velocity vector, including the velocity components in the horizontal and vertical directions, is calculated. Then, starting from the current frame position, the predicted position for the next moment is directly calculated by combining the aforementioned velocity vector with the prediction time interval (e.g., one frame).

[0135] It should also be noted that the maneuvering prediction model is constructed in the following way: the steering angle acceleration of the ship target's historical trajectory is recorded, and when the ship target has a steering behavior before disappearing, an adaptive Kalman filter based on the steering angle acceleration is used to predict the trajectory.

[0136] Specifically, the triggering of ship target re-identification includes: T1, extracting the visual feature vector of the re-reappearing ship target in the prediction area, and calculating the cosine similarity between the vector and the historical visual feature vector.

[0137] T2. If the similarity is greater than or equal to the preset first threshold, it is determined to be the same target, and the physical size label of the original target is inherited.

[0138] T3. If the similarity is greater than or equal to the preset second threshold and less than the preset first threshold, collect the heading deviation of the displacement deviation of the current ship target. If both are lower than the preset threshold, they are determined to be the same target; otherwise, they are marked as targets to be confirmed.

[0139] T3. If the similarity is less than the preset second similarity threshold:

[0140] T4. When the target vessel is large, the preset multimodal verification mechanism is invoked for multimodal verification. If the verification passes, it is marked as a suspected re-identified target and reported; otherwise, it is marked as a re-identified target.

[0141] T5. When the ship target is small and the visibility is greater than the preset threshold, and the wave height and the number of historical obstructions are both less than the corresponding preset values, it is marked as a re-identified target.

[0142] T6. If the conditions are not met or the ship target is of another level, submit to the manual review queue. If no response is received within the time limit, it will be automatically marked as a re-identified target.

[0143] T7. For the target identification, extract its hull outline geometric topological features and the texture-invariant features of the waterline to regenerate the visual feature vector.

[0144] It should be noted that the cosine similarity calculation can separately calculate the cosine similarity of the contour and the texture, and then obtain the final cosine similarity through weighted summation. Furthermore, the weight settings need to consider the stability and discriminative power of the ship features. Contour features have high discriminative power in terms of ship type and scale and are less affected by environmental interference, so they can be assigned a higher weight, such as 0.6-0.8. Texture features are easily affected by lighting and occlusion, but can reflect subtle differences, so their weight can be set to 0.2-0.4. In practical applications, the weights can be dynamically adjusted through sample training to adapt to the reliability of features in different scenarios.

[0145] It should be noted that the first and second similarity thresholds are preset as the high threshold and medium threshold for cosine similarity, respectively. The high threshold is used to directly identify the same target. The setting should be based on the stability of visual feature vectors, such as the consistency of features like ship appearance and texture. It is usually set to a high value, such as 0.8-0.9, to ensure the accuracy of matching results under high similarity. The medium threshold can be set to filter targets that need further verification. The setting should be lower than the first threshold, such as 0.6-0.7, to balance missed detections and false judgments. This avoids directly excluding targets with medium similarity and also leaves room for subsequent displacement and heading deviation verification.

[0146] Understandably, the preset threshold for displacement deviation is used to determine the acceptable range of deviation between the ship's reproduced position and the predicted position. The setting needs to be combined with the ship's sailing speed, the accuracy of the prediction model, and the image resolution, such as 5-20 pixels. For fast-moving ships, the threshold can be appropriately relaxed, while for low-speed ships, it should be tightened to ensure that the consistency of displacement helps to verify the identity of the target.

[0147] A preset threshold for heading deviation is used to determine the acceptable range of deviation between the ship's heading and its historical heading during a recurrence. A reference ship's steering characteristics are set, such as large ships turning slowly while small ships are more maneuverable; typically, this threshold is [value missing]. - .

[0148] The preset wave height value is used to limit the wave height conditions for re-identification of small vessels. The setting should take into account the impact of waves on the visual characteristics of small vessels, such as 0.5-1.5 meters, to avoid direct judgment when excessive wave height causes drastic changes in vessel attitude and feature distortion. The preset value for the number of historical occlusions is used to limit the occlusion conditions for re-identification of small vessels. The setting should take into account the impact of occlusion on feature continuity, such as 2-5 times, to avoid direct judgment when frequent occlusion causes historical features to become invalid.

[0149] The threshold for manual review timeout is used to set the maximum waiting time for manual review. The setting should be combined with the actual business response efficiency, such as 5-20 minutes, to balance the accuracy of manual review and the timeliness of system processing, and to ensure that the system will automatically mark the review if it is not reviewed in time, so as to avoid the process from stalling.

[0150] S6. Integrate visual tracking results of ship targets with real-time AIS data to generate waterway risk warnings and navigation strategies.

[0151] In one specific embodiment, the specific process of generating waterway risk warning and navigation strategy is as follows: Based on the real-time capture of the ship's position, heading, speed and other dynamic trajectories by visual sensors, the ship's position, heading, speed and other dynamic trajectories are matched and fused with data such as ship identity, planned route, and registered size provided by the AIS system in a spatiotemporal manner. For example, the deviation between the real-time position tracked by vision and the position reported by AIS is compared to identify abnormal trajectories.

[0152] If, after integration, it is found that a vessel deviates from the channel or there is a risk of collision, such as when two vessels cross course and the distance is less than the safety threshold or the vessel enters a restricted area, a graded risk warning will be generated. Combined with channel rules and vessel performance, specific navigation strategies will be output, such as suggesting that high-priority vessels slow down and give way, and that low-priority vessels adjust their course to safe waters, thereby achieving comprehensive monitoring and precise intervention of channel dynamics.

[0153] For example, a red alert indicates an imminent collision, while a yellow alert indicates a deviation from the flight path.

[0154] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for improving navigation safety by coordinating an intelligent navigation mark with AIS data, characterized in that, The method comprises: S1, collecting a continuous image sequence of a waterway by a visual sensor carried by a navigation mark and preprocessing the image sequence; S2, identifying a ship target in the image by a multi-scale target detection model to generate an initial detection box set; S3, performing a target determination operation on the initial detection box set: dynamically configuring an NMS threshold value according to the physical size of the ship target, the physical size consisting of a physical length and a width; outputting a final detection box with a physical size label based on the configured NMS threshold value; S4, extracting a visual feature vector of the ship target in the final detection box, and assigning a feature update strategy to multiple ship targets appearing simultaneously based on the physical size label; the visual feature vector is generated by a double-channel convolutional network, and the specific generation steps are as follows: a first channel extracts ship body contour geometric topology features by a direction gradient histogram operator; a second channel extracts texture invariant features of a ship waterline by a local binary pattern operator; and a feature vector that is invariant to rotation and scale is generated by fusing the features extracted by the two channels, as the visual feature vector; the specific assignment steps of the feature update strategy are as follows: calculating an update weight of each ship target based on the physical size; matching a preset compensation coefficient based on the size level identifier, and compensating the update weight based on the compensation coefficient to obtain an updated priority weight; matching a preset feature vector update frequency based on the updated priority weight of each ship target, and assigning the matched preset feature vector update frequency to the corresponding ship target; S5, when the ship target moves out of the field of view, predicting a current position area based on its historical motion trajectory, triggering re-identification of the ship target, and returning to step S3; S6, fusing the visual tracking result of the ship target and real-time AIS data to generate a waterway risk warning and a navigation strategy.

2. The method of claim 1, wherein the method comprises: The specific generation process of the initial detection box set is as follows: performing size and pixel standardization on the preprocessed image sequence, and inputting the image sequence into a trained multi-scale target detection model; extracting multi-scale feature maps by a convolutional neural network and performing feature fusion, identifying the visual length and visual width of the ship target based on the fused features; selecting a rectangular or square anchor box group according to the comparison result of the aspect ratio and a preset threshold value to generate a preset anchor box; based on the current wave level, dynamically configuring an intersection-over-union threshold value, for each identified ship target, calculating the intersection-over-union of the ship target with all preset anchor boxes, and marking the anchor box with an intersection-over-union greater than a first preset threshold value as a positive sample anchor box; based on the current wave level, dynamically configuring an intersection-over-union threshold value, for each identified ship target, calculating the intersection-over-union of the ship target with all preset anchor boxes, and marking the anchor box with an intersection-over-union greater than the corresponding configured intersection-over-union threshold value as a positive sample anchor box; outputting the class probability of the ship target in each positive sample anchor box by a convolutional layer, and marking the corresponding positive sample anchor box of the ship target with a class probability exceeding a threshold value as a candidate box; calculating the product of the class probability of the ship target in each candidate box and the corresponding intersection-over-union as the confidence of the corresponding candidate box; filtering out the candidate boxes with a confidence lower than a threshold value, and collecting the remaining candidate boxes after filtering according to the scale of their feature maps to form the initial detection box set.

3. The method of claim 2, wherein the method further comprises: The specific determination process of the physical size is as follows: AIS signals of the ship are analyzed to obtain an AIS length and an AIS width; a length difference degree of the visual length and the AIS length and a width difference degree of the visual width and the AIS width are calculated; if the length difference degree is less than or equal to a preset first length difference threshold, the physical length is the AIS length; if the length difference degree is greater than the first preset length difference threshold and less than or equal to a preset second length difference threshold, the current visibility is detected, the AIS length and the visual length are fused according to a length fusion rule matched with the visibility, and a fused length is output as the physical length; if the length difference degree is greater than the preset second length difference threshold, a current wave height is collected, a preset wave height compensation model is matched, and a final length output by the model is taken as the physical length; if the width difference degree is less than or equal to a preset first width difference threshold, the physical width is the AIS width; if the width difference degree is greater than the first preset width difference threshold and less than or equal to a preset second width difference threshold, the AIS width and the visual width are fused according to a preset width fusion rule, and the physical width is a fused width; if the width difference degree is greater than the second width difference threshold, the physical width is the visual width; based on the physical length and the physical width, a final physical size is output.

4. The method of claim 3, wherein the method further comprises: The specific configuration process of the dynamic NMS threshold is as follows: length classification standards and width classification standards are determined according to the AIS type of the ship, and the levels include super-large, large, medium and small; the physical length and the physical width are classified respectively, if there is a super-large level in the length level or the width level, the initial allocation threshold is a first preset NMS threshold; otherwise, the initial allocation threshold is determined according to the following rules: if the length level and the width level are consistent, the average of the NMS thresholds corresponding to the two is taken as the initial allocation threshold; if the length level is higher than the width level, when the NMS threshold corresponding to the length is greater than the NMS threshold corresponding to the width, the NMS threshold corresponding to the length is taken as the initial allocation NMS threshold, otherwise, the difference between the NMS threshold corresponding to the length and a preset compensation interference influence value is taken as the initial allocation NMS threshold; if the length level is lower than the width level, the NMS threshold corresponding to the width is taken as the initial allocation NMS threshold; it is judged whether the visibility exceeds a preset threshold, if yes, the final NMS threshold is the initial allocation threshold; otherwise, a compensation value is calculated according to the proportion of the visibility and the preset threshold, and the sum of the initial allocation threshold and the compensation value is taken as the final NMS threshold.

5. The method of claim 1, wherein the method further comprises: receiving AIS data from a plurality of AIS transmitters; and determining a position of the plurality of AIS transmitters based on the received AIS data. The configuration of the NMS threshold also includes the following steps: a sliding window is established according to a preset frame length, and the maximum value, the minimum value and the average value of the number of ship targets in the window are counted; if the difference between the maximum value and the minimum value and the average value exceed a preset threshold, a preset safe NMS threshold is taken as the final allocation NMS threshold.

6. The method of claim 1, wherein the method further comprises: The label includes a physical length value, a physical width value, a size level identifier and an allocated NMS threshold.

7. The method of claim 1, wherein the method further comprises: receiving AIS data from a plurality of AIS transmitters; and determining a position of the plurality of AIS transmitters based on the received AIS data. The prediction process of the current position area is as follows: a historical motion trajectory point set of the ship target is extracted, and a heading fluctuation degree is calculated; If the heading fluctuation is less than a set angle threshold, output the straight extension prediction model as the final prediction model, and if the heading fluctuation is greater than or equal to the set angle threshold, output the maneuvering turning prediction model as the final prediction model; Based on the prediction model, a predicted position is output, a rectangular prediction area is generated with the predicted position as the center, and the rectangular prediction area is marked as a predicted current position area.

8. The method of claim 7, wherein the method further comprises: The triggering of the ship target re-identification includes: For the ship target reappearing in the prediction area, a visual feature vector of the ship target when reappearing is extracted, and a cosine similarity between the vector and a historical visual feature vector is calculated; If the similarity is greater than or equal to a preset first threshold, it is determined that the same target is identified, and the physical size label of the original target is inherited; If the similarity is greater than or equal to a preset second threshold and less than the preset first threshold, a displacement deviation and a heading deviation of the current ship target are collected, and if both are less than a preset threshold, it is determined that the same target is identified, otherwise it is marked as a to-be-confirmed target; If the similarity is less than the preset second similarity threshold: When the ship target is a large size, a preset multi-modal verification mechanism is called to perform multi-modal verification, and if the verification is passed, it is marked as a suspected re-identified target and reported, otherwise it is marked as a re-identified target; When the ship target is a small size and the visibility is greater than a preset threshold, and the wave height and the historical number of occlusions are both less than corresponding preset values, it is marked as a re-identified target; When the conditions are not met or the ship target is of other sizes, a manual review queue is submitted, and if it is not responded within a timeout, it is automatically marked as a re-identified target; A visual feature vector is regenerated for the re-identified target.

Images