Navigation safety improvement method based on cooperation of intelligent navigation mark and AIS data

Abstract

The invention belongs to the technical field of navigation safety analysis, and particularly discloses an intelligent navigation mark and AIS (Automatic Identification System) data collaborative navigation safety improvement method, which comprises the following steps: acquiring a navigation channel image through a navigation mark visual sensor, preprocessing the navigation channel image, and generating an initial detection frame through multi-scale detection; an NMS threshold value is dynamically configured according to the physical size of the ship, and a final detection frame with a physical size label is output; extracting ship visual feature vectors, distributing feature updating strategies for multiple ships according to physical sizes, predicting positions through historical tracks when the ships move out of a visual field, and triggering re-identification; and finally, a visual tracking result and real-time AIS data are fused, and a channel risk early warning and navigation strategy is generated. According to the method, the problems of missing of a front-end sensing layer, missing detection caused by the fact that an NMS mechanism is not involved, shielding of target re-identification design and re-identification trigger logic blank and the like in the prior art are solved in a targeted manner, and the accuracy and reliability of navigation safety guarantee are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of navigation safety analysis, and specifically relates to a navigation safety improvement method based on intelligent navigation mark and AIS data cooperation. BACKGROUND

[0002] With the increase of global shipping volume and the expansion of ports leading to the increase of channel density, ships frequently intersect, and bridges, construction areas and other special waters have higher requirements for navigation safety. The independent operation of traditional navigation marks and AIS cannot cope with the dynamic changes in risks, so precise early warning needs to be achieved through cooperation.

[0003] Prior art 1, such as the virtual electronic channel design method and system based on AIS technology disclosed in Chinese patent application No. 202510341710.9, corrects the ship position through multi-source data fusion, generates an initial channel based on a Bezier curve, and realizes dynamic adjustment through an environmental correction term and Kalman filtering, and finally transmits through VHF+5G dual channels. Further, it ensures that physical navigation marks can also be safely navigated, and is suitable for ship navigation and safety protection in extreme environments such as ice and typhoon.

[0004] Prior art 2, such as the ship route planning method and device, electronic equipment and storage medium disclosed in Chinese patent application No. 202111282692.X, rasterizes the navigation chart and identifies obstacle grids based on wind speed and wave height thresholds, generates the shortest path using the grid A* algorithm for non-ocean scenarios, and detects the great circle route obstacles in segments for ocean scenarios and locally re-plans, triggers dynamic updates through ship position to avoid weather threats in real time, and improves the safety and efficiency of ship navigation.

[0005] For the above-mentioned prior art solutions, it is obvious that the current focus is on route planning on the front end and the back end, and there are still the following problems: 1. The existing patent has a serious lack in the front-end perception layer, especially the lack of obstacle identification accuracy and tracking robustness, such as the lack of deep design for occluded target re-identification.

[0006] 2. None of the existing patents mentions the NMS mechanism, and the default is to use a fixed threshold. When ships are dense, the detection frame is excessively merged, causing small ships to be suppressed far away, which can easily lead to missed detection.

[0007] 3. Patent 1 does not involve target tracking, and patent 2 only predicts the AIS ship position through Kalman filtering. Both of them do not involve re-identification trigger logic and cannot handle the situation when the target is occluded. SUMMARY

[0008] In view of this, in order to solve the above problems, a navigation safety improvement method based on intelligent navigation mark and AIS data cooperation is proposed.

[0009] The object of the present application can be achieved by the following technical solutions: The present application provides a method for improving navigation safety by cooperating intelligent navigation marks with AIS data, which comprises the following steps: S1, collecting continuous image sequences of waterways by a vision sensor mounted on a navigation mark and performing preprocessing.

[0010] S2, identifying ship targets in the images by a multi-scale target detection model to generate an initial detection frame set.

[0011] S3, performing a target determination operation on the initial detection frame set: dynamically configuring an NMS threshold value according to the physical size of the ship target, which consists of physical length and width.

[0012] Outputting a final detection frame with a physical size label based on the configured NMS threshold value.

[0013] S4, extracting a visual feature vector of the ship target in the final detection frame, and assigning a feature update strategy to multiple ship targets appearing simultaneously based on the physical size label.

[0014] S5, when the ship target moves out of the field of view, predicting the current position area based on its historical motion trajectory, triggering re-identification of the ship target and returning to step S3.

[0015] S6, fusing the visual tracking result of the ship target with real-time AIS data to generate a waterway risk warning and a navigation strategy.

[0016] Compared with the prior art, the present application has the following advantages: (1) The present application can more accurately capture ship targets in waterways by collecting continuous image sequences by a vision sensor mounted on a navigation mark and preprocessing, and combining a multi-scale target detection model, which solves the problem of insufficient obstacle recognition accuracy of the front-end perception layer and significantly improves the perception ability of various ships.

[0017] (2) The present application dynamically configures an NMS threshold value according to the physical size of the ship target, outputs a final detection frame with a physical size label, which can avoid excessive merging of detection frames when ships are dense, prevent small ships 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) The present application extracts a visual feature vector of the ship target and assigns a feature update strategy to multiple ships appearing simultaneously based on the physical size label, and at the same time, cooperates with the mechanism of predicting the position area based on the historical trajectory when the target moves out of the field of view and triggering re-identification, which can effectively handle the situation of target occlusion, makes up for the lack of target tracking and re-identification triggering logic in the prior art, and enhances the robustness of target tracking.

[0019] (4) The application fuses the visual tracking results of the ship target and the real-time AIS data, effectively combines the front-end accurate perception and the back-end data, makes the generated channel risk early warning and navigation strategy more comprehensive and reliable, and further improves the ability to cope with dynamic change risks, and guarantees the safety and efficiency of ship navigation. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 The whole implementation process of the present application is shown in the schematic diagram.

[0022] Figure 2 The main process of determining the physical size of the present application is shown in the schematic diagram.

[0023] Figure 3 The process of determining the physical length of the present application is shown in the schematic diagram.

[0024] Figure 4 The process of determining the physical width of the present application is shown in the schematic diagram.

[0025] Figure 5 The process of dynamically configuring the NMS threshold of the present application is shown in the schematic diagram. DETAILED DESCRIPTION

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

[0027] Please refer to Figure 1 As shown in the figure, the present application provides a navigation safety improvement method of intelligent navigation mark and AIS data cooperation, which comprises the following steps: S1, acquiring a continuous image sequence of the channel water area by the vision sensor carried on the navigation mark and preprocessing.

[0028] Understandably, the vision sensor carried on the intelligent navigation mark, such as high-definition camera, will continuously shoot the channel water area according to the set frame rate, for example, 30 frames per second, so as to obtain continuous image sequence. These images cover the information of ships, surrounding environment and possible obstacles on the channel.

[0029] It should be noted that the image preprocessing involves multiple operation steps such as grayscale processing, filter denoising and image enhancement. The filter denoising can remove noise interference in the image by using Gaussian filtering, median filtering and other methods. The image enhancement can use histogram equalization, contrast-limited adaptive histogram equalization and other techniques.

[0030] S2, identifying a ship target in the image by a multi-scale target detection model to generate an initial detection box set.

[0031] In one specific embodiment, the multi-scale target detection model can be a YOLO series model or a FasterR-CNN model, which is trained by a channel image data set containing ship categories, visual physical dimensions and position labels, and the multi-scale target detection model such as FasterR-CNN and YOLO series can extract features and identify targets in images at different scales, which is convenient for adapting to different sizes and proportions of ships that may appear in images.

[0032] It should be noted that the model training can be performed by collecting a large number of channel water image data sets containing ship targets, and labeling the ships in the images. The labeling content includes the categories of the ships such as cargo ships, passenger ships and fishing ships, as well as the position and actual size information of the ships in the images. The selected target detection model is trained using these labeled data, and the parameters of the model are continuously adjusted to accurately identify the ship targets in the images.

[0033] Further, the specific generation process of generating the initial detection box set is as follows: A1, performing size and pixel standardization on the preprocessed image sequence and inputting it into the trained multi-scale target detection model.

[0034] A2, extracting multi-scale feature maps by a convolutional neural network and performing feature fusion to identify the visual length and visual width of the ship target based on the fused features.

[0035] A3, according to the comparison result of the aspect ratio and the preset threshold, selecting a rectangular or square anchor box group to generate a preset anchor box.

[0036] A4, based on the current wave level, dynamically configuring the intersection-over-union threshold, calculating the intersection-over-union of each identified ship target with all preset anchor boxes, and marking the anchor box with an intersection-over-union greater than a first preset threshold as a positive sample anchor box.

[0037] A5, importing the current wave level and dynamically configuring the intersection-over-union threshold, calculating the intersection-over-union of each identified 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 as a positive sample anchor box.

[0038] A6, output the class probability of the ship target in each positive sample anchor frame through the convolutional layer, and mark the positive sample anchor frame corresponding to the ship target with a class probability exceeding the standard as a candidate frame.

[0039] A7, calculate the product of the class probability of the ship target in each candidate frame and the corresponding intersection-over-union as the confidence of the corresponding candidate frame.

[0040] A8, filter out candidate frames with a confidence lower than a threshold, and aggregate the remaining candidate frames after filtering according to the scale of their feature maps to form an initial detection frame set.

[0041] It should be noted that the size normalization in step A1 is mainly based on model design to scale or pad the image, such as YOLOv5 default 640x640, FasterR-CNN commonly used 800x1333, to ensure fixed input size. For example, after scaling the channel image of 1920x1080 to 640x360 at the same proportion, black pixels are filled on both sides to 640x640 to avoid deformation of the ship target. Pixel normalization can be done by converting pixel values from 0-255 range to 0-1 or -1-1, such as YOLO series using division by 255, FasterR-CNN using mean subtraction, to eliminate the influence of light intensity difference on the model and speed up parameter convergence.

[0042] Illustratively, taking YOLOv5 as an example, through the CSPDarknet53 backbone network, the input image is down-sampled 5 times, each time the size is reduced by 1 / 2, generating 5 scale feature maps, such as 640x640, 320x320, 160x160, 80x80, 40x40 and 20x20. The smaller the scale of the feature map, the larger the receptive field, which can correspond to a large field of view in the distance and is suitable for detecting large ships. The larger the scale, the smaller the receptive field, which is suitable for detecting small ships.

[0043] Understandably, the multi-scale feature maps can be fused through the Neck structure such as PANet in step A2, such as combining the high semantic feature of 20x20 with the high resolution feature of 80x80, which can not only retain the contour, texture and other detailed features of the ship, but also integrate the context information of the channel background such as water surface and shoreline, thereby improving the detection rate of small targets such as fishing boats.

[0044] Further, the identification process of visual length and visual width in step A2 is as follows: assuming that the known beacon camera focal length is 8mm, the installation height is 10m and the depression angle is 30°, then the visual length and visual width of the ship target can be obtained by calculating the actual distance corresponding to 1 pixel based on pixel width and pixel length through the geometric projection formula.

[0045] It is necessary to supplement that in this scenario, the geometric projection formula for calculating the actual distance through the pixel size can be based on the pinhole imaging principle and the trigonometric function relationship, and the focal length needs to be converted into a unit consistent with the installation height, and the pixel physical size is determined by the camera sensor parameters.

[0046] The specific formula of the geometric projection formula is as follows: , wherein, represents the actual distance, represents the pixel size, represents the installation height, represents the depression angle, represents the focal length, represents the pixel physical size. In the formula, represents the effective component of the vertical projection distance of the camera to the target ground on the horizontal plane, corresponds to the imaging scale parameter on the image plane, The ratio constitutes the conversion factor of the pixel size and the actual distance, and finally the pixel size, i.e. the number of pixels occupied by the ship target in the image, is multiplied by the conversion factor of the actual distance to obtain the result.

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

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

[0049] It is necessary to supplement that the matching of the aspect ratio and the anchor frame size is achieved by calculating the similarity and setting a matching similarity threshold, and the anchor frame size that exceeds the matching similarity threshold is taken as the matching anchor frame size, wherein the similarity between the ship pixel size and each preset anchor frame size is usually calculated by using the intersection over union method or the Euclidean distance method, the intersection over union method calculates the intersection over union by regarding the ship pixel size as a rectangle and calculating the intersection over union with the anchor frame size, and both the intersection over union method and the Euclidean distance method use existing similarity calculation methods, and the specific calculation formula is an existing empirical formula, which is not shown.

[0050] Exemplarily, a matching similarity threshold is set, such as an intersection-over-union greater than or equal to 0.6, and for example, a Euclidean distance less than 50 pixels, and an anchor box satisfying the threshold is selected. If there are multiple candidate anchor boxes, the one with the highest similarity is taken as the matching result. For example, the intersection-over-union of a certain ship pixel size (90, 18) and an anchor box (80, 16) is 0.8, which is higher than the intersection-over-union of 0.75 with an anchor box (100, 20), so the former is matched.

[0051] It should be noted that if the actual size estimation is inaccurate due to partial occlusion of the ship, such as only the bow is identified, the average size of the same type of ship in the historical trajectory or the size data broadcast by AIS is combined to correct the pixel size before matching. If the pixel size of the same ship is different at different distances, the corresponding scale feature map is automatically associated during matching, such as a small ship matching a small scale anchor box at a distance, corresponding to a high resolution feature map.

[0052] It should also be noted that the preset anchor box is designed based on the pixel size statistics of common ships in the channel before model training, and needs to cover ships of different scales and shapes, for example: small ship anchor box is usually [30, 6], [40, 8], medium ship anchor box is [80, 16], [100, 20], large ship anchor box: [200, 40], [300, 60], wherein the width-length ratio is 0.2. The said values, such as [30, 6], refer 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 intersection-over-union threshold in the A4 step is: real-time acquisition of channel surge level data by environmental sensors carried by navigation marks and import into the navigation system, assuming that the navigation system presets a surge level threshold of 3, when the detected surge level is greater than 3, it is determined to be a high wave environment, at this time the ship target in the image is easy to produce shape blur or edge distortion due to wave rocking, the intersection-over-union threshold is dynamically adjusted to 0.5 to increase the number of positive sample anchor boxes, avoiding missed detection. When the surge level is less than or equal to 3, the channel environment is stable, the ship target is imaged clearly, the intersection-over-union threshold remains 0.7 to strictly screen high-quality anchor boxes, improve detection accuracy, and realize adaptive optimization of anchor box matching under different surge conditions. And the specific value is the optimal balance value obtained based on the influence of surge on ship image features and a large number of experimental verification.

[0054] It should also be noted that, under normal circumstances, anchor boxes with an intersection-over-union greater than a threshold such as 0.5 are marked as positive samples responsible for predicting the target, those with an intersection-over-union between 0.2 and 0.5 are ignored samples, and those with an intersection-over-union less than 0.2 are negative samples responsible for predicting the background. That is, 0.5 can be set as the first preset threshold for specific reference value, and specific dynamic adjustment can also be made according to the actual scene.

[0055] Understandably, the category probability of the ship target output by the convolution layer in the A6 step also includes optimizing the classification accuracy by using a cross-entropy loss function. The cross-entropy loss function is a prior function, and thus will not be described here.

[0056] Understandably, the setting of the confidence threshold in the A7 step needs to be combined with the characteristics of the ship detection scene and the model performance, and is determined by statistical analysis of the confidence distribution of the candidate box and experimental verification. For example, the confidence distribution of the ship target and the background in the sample is first counted to distinguish the critical value of the real target and the false detection, which is the initial value such as 0.2. Then, the detection accuracy and recall rate under different threshold values are tested on the verification set. Finally, the threshold value that balances the missed detection rate and the false detection rate is selected, such as 0.25-0.3. In complex environments such as high waves, the threshold value can be appropriately reduced, such as 0.2, to reduce missed detection. In stable scenes such as clear weather, the threshold value can be increased, such as 0.35, to reduce false detection.

[0057] Understandably, the reference example summarized in the A8 step is as follows: 30 medium-sized ship candidate boxes are retained on the 80x80 feature map, 15 large-sized ship candidate boxes are retained on the 20x20 feature map, and the final set contains 45 initial detection boxes, each of which contains position coordinates, confidence, and predicted category information.

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

[0059] 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 being composed of a physical length and a width.

[0060] Outputting a final detection box with a physical size label based on the configured NMS threshold value.

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

[0062] Specifically, referring to Figures 2 to 4 As shown in the figure, the specific determination process of the physical size is as follows: B1, analyzing the ship AIS signal to obtain an AIS length and an AIS width.

[0063] B2, calculating the length difference degree of the visual length and the AIS length and the width difference degree of the visual width and the AIS width.

[0064] B3, if the length difference degree is less than or equal to a preset first length difference threshold value, the physical length is the AIS length.

[0065] B4, if the length difference degree is greater than the first preset length difference threshold and less than or equal to the preset second length difference threshold, detecting the current visibility, fusing the AIS length and the visual length according to the length fusion rule matched with the visibility, and outputting the fused length as the physical length.

[0066] B5, if the length difference degree is greater than the preset second length difference threshold, collecting the current wave height and matching the preset wave height compensation model, and outputting the final length of the model as the physical length.

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

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

[0069] B8, if the width difference degree is greater than the second width difference threshold, the physical width is the visual width.

[0070] B9, based on the physical length and the physical width, outputting the final physical size.

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

[0072] Understandably, the visibility in B4 step can directly obtain real-time visibility data by using a meteorological sensor such as a visibility meter installed near a navigation mark or a camera. If a dedicated sensor is not equipped, the visibility condition can also be indirectly inferred by analyzing the image features collected by the camera, such as image contrast, edge sharpness, and blur degree of distant targets, and combining image sharpness evaluation algorithms such as Laplacian-based sharpness detection and atmospheric light estimation based on foggy image model.

[0073] Understandably, the specific rules of the preset length fusion rule in B4 step are as follows: 1) when the visibility is greater than 1000 meters, the first assigned length fusion weight is matched, and the weighted sum of the visual length and the AIS length with the fusion weight is taken as the fused length.

[0074] Exemplarily, the visual length and the AIS length in the first assigned length fusion weight are respectively taken as 0.7 and 0.3.

[0075] 2) match the second distribution length fusion weight when the visibility is less than or equal to 1000 meters, and take the weighted sum of the visual length and the AIS length and the fusion weight as the fusion length.

[0076] Exemplarily, the visual length and the AIS length in the second distribution length fusion weight are respectively 0.4 and 0.6.

[0077] It should be noted that, based on the influence law of the visibility on the visual length detection accuracy, it is determined through a large number of experimental verifications that 1000 meters is the critical value of the visual imaging definition, the visual data reliability is high when greater than 1000 meters, and the accuracy is reduced when less than or equal to 1000 meters due to interference such as fog, and the fusion weight distribution of 0.7 and 0.3, 0.4 and 0.6 is the optimal proportion that minimizes the fusion length error through comparison of the error distribution of the visual and AIS data under different visibilities, and is screened out through multiple fusion experiments, which can realize the reasonable adaptation of the data reliability under different visibilities.

[0078] Understandably, the specific content of the wave height compensation model in the B5 step is as follows: let the AIS length and the visual length be respectively and , calculate the compensation length , , as the current wave height.

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

[0080] If the wave height is greater than 4 meters, start binocular visual re-measurement, that is, capture the ship waterline image sequence, calculate the re-measurement length based on the draft-pixel ratio model of the draft depth actual value and the corresponding pixel length of the draft depth in the image.

[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 determined based on the scene characteristics of ship visual detection and AIS data fusion, 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 of the influence of wave height on visual detection accuracy, the deviation is controllable within 3 meters, and verification needs to be strengthened above 4 meters, 0.6 and 0.4 are the fusion weights of the compensation length and the binocular re-measurement length in high wave environment, and the whole is optimized through multiple comparison experiments to minimize the length detection error under different wave conditions.

[0083] It should be noted that the wave height can be directly measured by using wave sensors such as acoustic Doppler wave meters, pressure wave sensors, etc. deployed in the water area, and can also be analyzed and inverted by radar remote sensing technology on the sea surface reflection signal, and in the scene with video monitoring, the sea surface fluctuation characteristics can also be identified from the sequence images by combining image processing algorithm to estimate the wave height.

[0084] Understandably, the preset width fusion rule in B7 is that if the detected ship speed is less than or equal to the preset ship speed threshold, the first distribution width fusion weight is matched, if the detected ship speed is greater than the preset ship speed threshold, the second distribution width fusion weight is matched, the weighted sum of the visual width and the AIS width and the fusion weight is taken as the fusion width, and the first and second distribution width fusion weights are taken as 0.6 and 0.4 and 0.3 and 0.7, respectively.

[0085] It should be noted that the experiment verifies that 5 knots is the critical value of the stability of the visual imaging of the width of the ship, that is, as a preset ship speed threshold, when less than or equal to 5 knots, the width imaging is clear and the error is small, and when greater than 5 knots, the visual detection deviation increases due to motion smearing.

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

[0087] It should be noted that the setting of each preset threshold in the physical size determination process is based on the difference distribution statistics of the visual and AIS size data of the ship, the influence degree of the environment on the data reliability and the experimental verification, and the difference law of the visual and AIS length or width in a large number of samples can be analyzed.

[0088] Exemplarily, the first difference threshold such as 5% of the length and 8% of the width can be set as the critical value of the AIS data reliability, that is, the difference is within this range, and the AIS value is directly used. The second difference threshold, such as 15% of the length and 20% of the width, is set as the critical value that needs to be fused or compensated. The difference beyond this range indicates that the data conflict is significant.

[0089] It should be noted that the above setting process of the physical size fully combines the stable data of the AIS signal and the real-time of the visual recognition, and through the difference threshold grading processing, the AIS data, the fusion data combined with the visibility, the model output length combined with the AIS data, the preset rule fusion data, and the visual width are used respectively under different difference degrees, so as to realize the accurate determination of the physical size of the ship, and the data reliability and environmental adaptability are considered.

[0090] It should be further explained that the preset NMS threshold allocation table can be consulted based on the physical length and the physical width to obtain the NMS threshold corresponding to the physical length and the physical width.

[0091] Understandably, the preset NMS threshold allocation table is divided into multiple interval ranges according to the numerical ranges of the physical length and the physical width of the ship. For example, the physical length can be divided into 0-30m, 30-60m, 60-100m, 100m and above, and the physical width can be divided into 0-10m, 10-20m, 20-50m, 50m and above. The width interval is transverse, the length interval is longitudinal, the NMS threshold is set in the cross cell according to the principle that the larger the size, the larger the threshold, and the interval division is combined with the actual ship size distribution and the detection box characteristics to determine the threshold value.

[0092] Further, referring to Figure 5 As shown in the figure, the specific configuration process of the dynamic configuration NMS threshold is: E1, determining the length grading standard and the width grading standard according to the ship AIS type, the level includes super large, large, medium and small.

[0093] Understandably, the AIS type grading standard is set according to the differences of different ship operation environment, size characteristics and collision risk, which is used to adapt to the different needs of ship target detection for threshold strictness. For example, the length grading standard is as follows: for ocean freighter, the length is greater than the super large ship length threshold, which is super large, the minimum length of large ship is ≤ length < super large ship length threshold, which is large, for inland river ship, length ≥ large inland river ship length threshold, which is large, medium ship minimum length ≤ length < large inland river ship length threshold, which is medium, for fishing boat, length < small ship length threshold, which is small.

[0094] E2, grading the physical length and the physical width respectively, if there is a super 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 the super large ship in the ship target detection scene. Usually, the value is low, such as 0.3-0.5. The setting mainly considers the large area overlapping characteristics of super large ship in image, the fault tolerance requirement of avoiding missing detection and the differentiated adaptation of detection threshold with other level ships, in order to accurately select the candidate box and reduce the false deletion.

[0096] In the ship target detection, the NMS threshold suitable for the current ship is dynamically determined by the above rules to optimize the screening effect of the candidate box in the target detection, for example, for super large ships, a stricter NMS threshold, such as 0.25, is used to reduce redundant boxes. When the length classification is more strict, the threshold is appropriately reduced to enhance the filtering accuracy, etc. That is, the 0.25 can be used as a reference value of the first preset NMS threshold.

[0097] E3, otherwise, the initial allocation threshold is determined according to the following rules: E31, if the length and width levels are consistent, the average of the NMS thresholds corresponding to the two is taken as the initial allocation threshold.

[0098] Among them, the average of the two is taken as the final threshold. This processing takes into account the characteristics of the length and width dimensions, making the final threshold more suitable for the overall detection needs of the ship of this level.

[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 allocation NMS threshold, otherwise, the difference between the length corresponding NMS threshold and the preset compensation interference value is taken as the initial allocation NMS threshold.

[0100] Understandably, when the length level is higher than the width level, the characteristics of the ship target in the length dimension are more prominent, and the influence on the detection accuracy is greater, so the length corresponding NMS threshold is taken as the reference. If the length corresponding NMS threshold is not greater than the width corresponding NMS threshold, it means that there may be interference factors in the width dimension, such as ship roll, shooting angle deviation, etc. By subtracting the preset compensation interference value, the threshold can be adjusted to reduce the influence of interference on target screening and ensure that the candidate box that is more consistent with the actual situation is retained.

[0101] It should be noted that the preset compensation interference value is usually set according to the common width interference degree in the ship detection scene, such as the influence amplitude of different wave heights and ship attitude changes on width recognition, which is generally an empirical value such as 0.05-0.2. The specific value needs to be calibrated in combination with the deviation law of the length and width thresholds in the historical detection data.

[0102] E33, if the length level is lower than the width level, the width corresponding NMS threshold is taken as the initial allocation NMS threshold.

[0103] The width level is higher, and its threshold itself is more relaxed, which meets the demand of high level for detection fault tolerance, and does not need to be adjusted additionally.

[0104] E34, it is judged whether the visibility exceeds the preset threshold, if yes, the final NMS threshold is the initial allocation threshold.

[0105] Understandably, the above visibility preset threshold is usually 1000 meters.

[0106] E35、Otherwise, calculate a compensation value in proportion to the visibility and a preset threshold value, and take the sum of the initial allocation threshold value and the compensation value as the final NMS threshold value.

[0107] Understandably, the compensation value is obtained by setting the maximum allowable compensation value and multiplying the proportion of the visibility to the preset threshold value by the maximum allowable compensation value. The setting of the maximum allowable compensation value needs to be combined with the effective adjustment range of the NMS threshold value in the ship detection scene, the influence degree of the extreme value of the visibility on the detection accuracy, and the balance relationship between the compensation effect and the false detection or missed detection rate in the historical data, and is usually set to 0.05-0.1. The specific value needs to be calibrated through multiple experiments.

[0108] Further, the configuration of the NMS threshold value further includes the following steps: establishing a sliding window according to a preset frame length, and counting the maximum value, the minimum value and the average value of the number of ship targets in the window.

[0109] If the proportion of the difference between the maximum value and the minimum value to the average number of ship targets exceeds a preset threshold value, a preset safe NMS threshold value is taken as the final allocated NMS threshold value.

[0110] Understandably, the setting of the preset frame length needs to be combined with the motion characteristics of the ship target in the video sequence, such as the sailing speed, the steering frequency, etc., and needs to be combined with the dynamic change degree of the detection scene and the limitation of the calculation resources, and is usually taken as 5-30 frames. For example, a high-speed sailing ship or a dynamic complex scene can be set to a shorter frame length such as 5-10 frames to quickly capture the target quantity fluctuation, and a low-speed or stable scene can be set to a longer frame length such as 15-30 frames to smooth the instantaneous noise and balance the real-time performance and the statistical stability.

[0111] Understandably, the preset threshold value in the proportion exceeding the preset threshold value is usually set to 30%, which corresponds to the critical value of the target number fluctuation of the continuous frames, and is used to determine whether the detection is disturbed.

[0112] It should be noted that the preset safe NMS threshold value can be usually set to 0.5, and the safe threshold value 0.5 is an empirical value, which is neither too strict in missed detection nor too loose in false detection, and in an abnormal scene, at least the target is detected in priority to avoid the decision-making error caused by the false judgment of the system.

[0113] S4, extract the visual feature vector of the ship target in the final detection frame, and assign a feature updating strategy to multiple ship targets appearing at the same time based on the physical size label.

[0114] Specifically, the visual feature vector is generated by a double-channel convolutional network, and the specific generation steps are as follows: the first channel extracts the hull contour geometric topology feature by a direction gradient histogram operator.

[0115] The second channel extracts texture invariant features of the ship waterline through a local binary pattern operator.

[0116] The fused features extracted by the two channels generate a rotation and scale invariant feature vector as the visual feature vector.

[0117] Understandably, in the scene of the ship waterline, the texture invariant features specifically represent the texture patterns of the waterline region, such as the inherent stability of the line direction, the alternating law of light and shade, the repeated structure of local details, that is, no matter the ship rotates, scales, or the brightness changes due to changes in illumination, the core texture features of the waterline, such as the relative distribution of lines and the periodicity of textures, remain consistent, so they can be stably extracted by the algorithm, providing reliable texture basis for ship recognition, matching, and other tasks.

[0118] Understandably, when the fusion processing is performed, the geometric topological features of the ship hull contour extracted by the first channel, such as the relative positions of the vertices and the curvature distribution of the contour, are subjected to affine transformation normalization to eliminate the effects of rotation and scale changes on the contour structure. Meanwhile, the texture invariant features of the waterline extracted by the second channel, such as the local texture gradient direction and the frequency features of the repeated patterns, are subjected to multi-scale pyramid sampling and rotation invariance filtering to ensure the consistency of the texture information under different angles and scaling ratios. Subsequently, the two are combined through feature splicing, and a convolution layer is used for feature mapping and dimension compression to output the final feature vector. The multi-scale pyramid sampling and rotation invariance filtering involved in the fusion processing are all existing technologies, which will not be described in detail here.

[0119] It should be noted that the histogram of oriented gradients operator and the local binary pattern operator are both existing operators, which will not be described in detail.

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

[0121] Based on the size level identification, a preset compensation coefficient is matched, and the update weight is compensated based on the compensation coefficient to obtain an updated priority weight.

[0122] Based on the corresponding updated priority weight of each ship target, a preset feature vector update frequency is matched, and the matched preset feature vector update frequency is allocated to the corresponding ship target.

[0123] Understandably, the specific calculation process of the updated weight is as follows: the current ship motion speed is collected, and the motion speed, physical length and physical width are normalized respectively, and the normalized results are weighted and summed based on the preset weight to output the final updated weight. The normalization can be performed by the minimum-maximum normalization method. Because narrow-body ships are highly flexible and it is difficult to predict potential risks, they need to be given priority. The physical width value obtained by the minimum-maximum normalization is subtracted by 1 as the final normalized physical width.

[0124] It should be noted that in the scenes of ship supervision, obstacle avoidance, etc., the physical length of the ship is usually more directly related to the association of navigation safety and supervision priority. Large ships have large inertia and high maneuvering difficulty, and the collision risk or supervision complexity is much higher than that of small ships. The width of the ship is a static attribute related to the adaptability to the water environment. Compared with the dynamic motion speed, the stable determination of the priority has more basic reference value. Therefore, the influence weight of the physical length is set to be greater than that of the physical width, and the influence weight of the motion speed.

[0125] Understandably, the setting of the preset compensation coefficient is determined according to the association rule of the priority deviation of the ship size level identifier, such as large, medium and small levels, in the actual scene. For different size levels, the deviation between the actual priority and the basic weight of the ships of this level in the historical data in terms of collision risk, supervision demand, etc. can be analyzed. For example, large ships need a higher priority compensation coefficient of 1.2 due to their large inertia, and small ships need a compensation coefficient of 0.8 because they are easily ignored. Through experimental verification, the coefficient value that can narrow the gap between the theoretical weight and the actual demand is selected as the priority weight.

[0126] In one specific embodiment, the setting of the preset feature vector update frequency can be determined according to the priority weight of the ship target after updating, combined with the time effectiveness demand of the feature vector and the calculation resource cost. High-priority ships need more timely feature updates to accurately track dynamic changes, such as setting 50ms each time. Medium-priority ships balance real-time performance and resource consumption, such as setting 100ms each time. Low-priority ships reduce the update frequency to save resources, such as setting 200ms each time. By dividing multiple frequency levels and matching the weight interval, the feature time effectiveness of high-priority targets can be guaranteed, and excessive calculation resources of low-priority targets can be avoided, realizing the optimized allocation of resources.

[0127] S5, when the ship target moves out of the field of view, the current position area is predicted based on the historical motion trajectory, and the ship target re-identification is triggered to return to step S3.

[0128] Specifically, the prediction process of the current position area is as follows: the historical motion trajectory point set of the ship target is extracted, and the heading fluctuation degree is calculated.

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

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

[0131] Understandably, the heading fluctuation can be calculated by first calculating the heading angles of adjacent track points through the inverse tangent function, that is, the angle between the line connecting the two points and the reference direction such as due north or the horizontal right of the image, and then calculating the fluctuation degree of all adjacent heading angles. The angle standard deviation or maximum angle difference is usually used as the heading fluctuation index to quantify the stability of the ship's motion direction.

[0132] It is understandable that the angle threshold is set according to the type of ship. For example, the slow turning threshold for a large cargo ship can be set smaller, such as , the steering flexibility threshold of small speedboats can be set larger, such as .

[0133] It is understandable that the length of the area is proportional to the speed, and the formula is: , The proportional coefficient is usually 0.5-2, which is adjusted according to the prediction accuracy requirements. Take a larger value to expand the coverage, when the speed is slow Take a small value to improve positioning accuracy. For speed, is the side length of the region.

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

[0135] It should also be noted that the maneuvering steering prediction model is constructed in the following way: the steering angular acceleration of the historical trajectory of the ship target is recorded, and when there is a steering behavior before the ship target disappears, the trajectory is predicted using an adaptive Kalman filter based on the steering angular acceleration.

[0136] In another specific embodiment, the triggering of ship target re-identification includes: T1, for a ship target that reappears in the prediction area, extracting its visual feature vector when it reappears, 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 a preset first threshold value, it is determined that 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 a preset second threshold value and less than a preset first threshold value, the displacement deviation and the heading deviation of the current ship target are collected, and if both are less than a preset threshold value, it is determined that the same target, otherwise it is marked as a to-be-confirmed target.

[0139] T3, if the similarity is less than a preset second similarity threshold value: T4, 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-identification target and reported, otherwise it is marked as a re-identification target.

[0140] T5, when the ship target is a small size and the visibility is greater than a preset threshold value, and the wave height and the historical number of occlusions are both less than the corresponding preset values, it is marked as a re-identification target.

[0141] T6, when the condition is not established or the ship target is of other levels, submit the manual review queue, and if it is not responded within a timeout, it is automatically marked as a re-identification target.

[0142] T7, for the re-identification target, the ship body contour geometric topology features and the texture invariant features of the ship waterline are extracted to regenerate the visual feature vector.

[0143] It should be noted that the cosine similarity calculation can calculate the cosine similarity of the contour and the texture respectively, and the final cosine similarity is calculated by weighted summation, and the weight setting needs to be combined with the stability and discrimination of the ship features. The contour feature has high discrimination in ship type and size and is less affected by the environment, and can be given a higher weight such as 0.6-0.8. The texture feature is easily affected by light and occlusion, but can reflect the difference in details, and the weight can be set to 0.2-0.4. In actual application, the weight can be dynamically adjusted through sample training to adapt to the reliability of the features in different scenes.

[0144] It should be noted that the preset first and second similarity threshold values are respectively used as high and medium threshold values of the cosine similarity. The high threshold value is used to directly determine the same target, and the setting needs to be based on the stability of the visual feature vector, such as the consistency of the ship appearance and texture features, and is usually set to a higher value such as 0.8-0.9 to ensure the accuracy of the matching result under high similarity. The medium threshold value is set to filter targets that need to be further verified, and is set to be lower than the first threshold value such as 0.6-0.7 to balance the missed detection and misjudgment, so as to avoid directly excluding medium similarity targets and reserve space for subsequent displacement and heading deviation verification.

[0145] The preset threshold of displacement deviation is used to determine the acceptable range of deviation between the reappeared position of the ship and the predicted position. The setting needs to be combined with the ship's navigation speed, the accuracy of the prediction model, and the image resolution, such as 5-20 pixels. The speed of the ship can be appropriately relaxed, and the speed of the low-speed ship can be tightened to ensure that the target identity is verified by the displacement consistency.

[0146] The preset threshold of heading deviation is used to determine the acceptable range of deviation between the heading of the ship when it reappears and the historical heading. The setting is based on the turning characteristics of the ship, such as large ships turning slowly and small ships being flexible. Typically, it is - .

[0147] The preset value of wave height is used to limit the wave height condition for re-identification of small ships. The setting needs to consider the influence of waves on the visual features of small ships, such as 0.5-1.5 meters, to avoid directly determining when the ship's attitude changes dramatically and the features are distorted due to excessive wave height. The preset value of the number of historical occlusions is used to limit the occlusion condition for re-identification of small ships. The setting needs to consider the influence of occlusions on feature continuity, such as 2-5 times, to avoid directly determining when the historical features are invalid due to frequent occlusions.

[0148] The threshold of manual review timeout is used to set the maximum waiting time for manual review. The setting needs to 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, to ensure that the marking is automatically completed when the review is not timely, and to avoid process stagnation.

[0149] S6, fuse the visual tracking results of the ship target and the real-time AIS data to generate channel risk early warning and navigation strategy.

[0150] In one specific embodiment, the specific process of generating channel risk early warning and navigation strategy is as follows: based on the real-time capture of the position, heading, speed, and other dynamic trajectories of the ship by the visual sensor, the data is spatio-temporally matched and fused with the ship identity, planned route, registered size, and other data provided by the AIS system, such as comparing the deviation between the real-time position of visual tracking and the position reported by AIS to identify trajectory anomalies.

[0151] If the fusion finds that the ship deviates from the channel and there is a collision danger, such as the heading of two ships crossing and the distance being less than a safety threshold or entering a prohibited area, a graded risk early warning is generated, and specific navigation strategies are output in combination with channel rules and ship performance, such as suggesting that high-priority ships slow down to avoid and that low-priority ships adjust their heading to safe waters, thereby achieving comprehensive monitoring and precise intervention of the dynamic channel.

[0152] For example, red early warning indicates an imminent collision, and yellow early warning indicates deviation from the route.

[0153] The above merely illustrates and describes the concept of the present application, and those skilled in the art can make various modifications or supplements to the described specific embodiments or adopt similar ways to replace, as long as the modifications or supplements do not deviate from the concept of the present application or exceed the defined scope of the present application, and should belong to the protection scope of the present application.

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 being composed 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 updating strategy to multiple ship targets appearing simultaneously based on the physical size label; S5, when the ship target moves out of the field of view, predicting a current position area based on its historical motion trajectory, triggering ship target re-identification and returning to step S3; S6, fusing the visual tracking result of the ship target with 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 the 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 the visual width of the ship target based on the fused features; selecting a rectangular or square anchor box group to generate a preset anchor box according to the comparison result of the aspect ratio and a preset 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; importing the current surge level and dynamically configuring the 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 as a candidate box if the class probability exceeds the standard; 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: analyzing the ship AIS signal to obtain an AIS length and an AIS width; calculating the length difference degree of the visual length and the AIS length and the width difference degree of the visual width and the AIS width; if the length difference degree is less than or equal to a first preset length difference threshold value, the physical length is the AIS length; if the length difference degree is greater than the first preset length difference threshold value and less than or equal to a second preset length difference threshold value, detecting the current visibility, fusing the AIS length and the visual length according to the length fusion rule matched with the visibility, and outputting the fused length as the physical length; if the length difference degree is greater than the second preset length difference threshold value, collecting the current wave height and matching the wave height with a preset wave height compensation model, and outputting the final length of the model as the physical length. If the width difference degree is less than or equal to a preset first width difference degree 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 then the physical width is the 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 configuration of the NMS threshold is as follows: According to the type of the ship AIS, length classification standards and width classification standards are determined, 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 corresponding NMS thresholds of the two is taken as the initial allocation threshold; 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 allocation NMS threshold, otherwise, the difference between the length corresponding NMS threshold 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 width corresponding NMS threshold is taken as the initial allocation NMS threshold; If the visibility exceeds a preset threshold, 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: 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 proportion of the average number of ship targets 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 contains 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 visual feature vector is generated by a double-channel convolution network, and the specific generation steps are as follows: The first channel extracts the ship body contour geometric topology feature by a direction gradient histogram operator; The second channel extracts the texture invariant feature of the ship waterline by a local binary pattern operator; The features extracted by the double channels are fused to generate a rotation and scale invariant feature vector as the visual feature vector.

8. The method of claim 6, 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 specific allocation steps of the feature update strategy are as follows: Based on the physical size, the update weight of each ship target is calculated; Based on the size level identifier, a preset compensation coefficient is matched, and the updated priority weight is obtained by compensating the update weight based on the compensation coefficient; Based on the updated priority weight corresponding to each ship target, a preset feature vector update frequency is matched, and the matched preset feature vector update frequency is allocated to the corresponding ship target.

9. 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: The historical motion trajectory point set of the ship target is extracted and the 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.

10. 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