Acquisition and construction method of remote ship identification high-quality data set

Abstract

The invention relates to an acquisition and construction method for a remote ship identification high-quality data set. The method comprises the following steps: acquiring position data and course data of a ship in the acquisition data set; acquiring sea chart data of the current leg based on the ship position data; delimiting an ROI (Region of Interest) region related to ship navigation in the obtained nautical chart data; judging whether the delimited ROI appears in a shooting field of view of the ship camera or not, if so, performing remote image capture, otherwise, giving up capture, and continuing to judge; different types of ROI images are selected from captured far-end images for stratified sampling, and an initial data set is formed; performing quality evaluation on the initial data set image, and removing the image with low quality; annotating the rejected image in combination with electronic chart data, and marking all visible ships; and performing data processing on the annotated image to increase the diversity and quantity of a data set. According to the method, a high-quality data set can be provided, so that a more efficient and accurate remote ship identification model can be trained.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, and specifically relates to a method for collecting and constructing a high-quality dataset for identifying distant ships. Background Technology

[0002] With the rapid development of the global shipping industry, the density and complexity of maritime traffic are constantly increasing, making ship navigation safety a more prominent issue. Especially when the identification of distant vessels is not timely, it can lead to serious collisions, threatening the safety of ships, personnel, and waterway facilities. Therefore, how to achieve long-distance, high-precision vessel identification and real-time monitoring of waterway safety through technological means has become an important research direction in the field of maritime safety. However, the lack of existing high-quality datasets limits the performance of existing models. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for collecting and constructing a high-quality dataset for distant vessel identification. By combining the Regions of Interest (ROIs) in electronic nautical chart data, it determines whether the ROIs are within the field of view of the ship's camera. If so, it performs remote image capture. Then, it performs stratified sampling and image annotation of different types of ROIs in the captured remote images, and uses electronic nautical chart data to verify the accuracy of the annotations. This ensures that the dataset is both accurate and diverse, thereby enabling the training of a more efficient and accurate distant vessel identification model.

[0004] To achieve the above objectives, this invention discloses a method for collecting and constructing a high-quality dataset for identifying distant vessels, comprising the following steps:

[0005] S1. Obtain the ship's position and heading data from the dataset;

[0006] S2. Obtain nautical chart data for the current voyage based on vessel position data;

[0007] S3. Delineate the ROI areas related to ship navigation in the acquired nautical chart data;

[0008] S4. Determine whether the defined ROI area is within the field of view of the ship's camera. If yes, perform remote image capture; otherwise, abandon capture and continue to determine.

[0009] S5. Select different types of ROI region images from the captured remote images and perform stratified sampling to form the initial dataset;

[0010] S6. Evaluate the quality of the images in the initial dataset and remove low-quality images;

[0011] S7. Annotate the removed images in conjunction with electronic nautical chart data, and mark all visible vessels;

[0012] S8. Perform data processing on the annotated images to increase the diversity and quantity of the target dataset;

[0013] S9. Balance the image data of ships and non-ships in the target dataset.

[0014] Furthermore, in S3, the ROI area related to ship navigation refers to the area in the nautical chart data that is delineated based on the characteristics of routes, channels, and marine landmarks in the nautical chart data, according to the principle that the higher the density of features related to ship navigation, the greater the probability of the number and types of ships appearing.

[0015] Furthermore, the delineation of the ROI area includes two aspects: on the one hand, combining factors such as shipping routes, channel density, and marine landmark characteristics, the ROI area is delineated by identifying channels, ports, and areas commonly used by various types of vessels; on the other hand, the ROI area is determined by combining historical vessel traffic data, obtaining historical vessel location data, and correlating the probability of vessel appearance with nautical chart data.

[0016] Further, step S4, determining whether the defined ROI region appears within the field of view of the ship's camera, if so, performs remote image capture; otherwise, abandons capture and continues the determination, including:

[0017] S41. Determine the field of view information of the ship's camera;

[0018] S42. Convert the camera's field of view to the nautical chart coordinate system;

[0019] S43. Compare the ROI area marked on the nautical chart with the camera's field of view;

[0020] S44. Based on the comparison results, determine whether the defined ROI area appears within the field of view of the ship's camera. If so, perform remote image capture; otherwise, abandon capture and continue the judgment.

[0021] Furthermore, the field-of-view information of the ship's camera in S41 includes the camera's field-of-view angle, as well as the camera's installation position and orientation.

[0022] Furthermore, in S5, the different types of ROI regions are calculated based on the probability of ship appearance from historical data, dividing the ocean area on the nautical chart into three levels: high probability ROI region, medium probability ROI region, and low probability ROI region.

[0023] Furthermore, the low-quality images in S6 include blurry images, exposure problems, distortion, occlusion, images that are too small, duplicate content, irrelevant backgrounds, compression artifacts, and damaged images.

[0024] Further, in step S7, the removed images are annotated using electronic nautical chart data, marking all visible vessels, including:

[0025] The removed images are annotated to mark all visible ships. The annotated images include the ship locations, ship bounding boxes, and types.

[0026] Verify the accuracy of the annotations by combining them with electronic nautical chart data to ensure that the annotated vessel positions are consistent with the nautical chart data.

[0027] Furthermore, the method also includes S10, evaluating the quality of the improved target dataset using cross-validation.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. By combining electronic nautical chart data and historical ship position data, this invention can more accurately determine the areas where ships may appear, thereby improving the representativeness of the dataset;

[0030] 2. This invention employs a stratified sampling method to ensure that the number of images selected from different probability regions is reasonable and reflects the actual distribution of ships;

[0031] 3. This invention ensures the accuracy and consistency of ship information annotation in images by annotating nautical chart data. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a method for collecting and constructing a high-quality dataset for identifying distant ships, as described in this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, further explanation is provided below in conjunction with the accompanying drawings and embodiments.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0035] As attached Figure 1 As shown, this invention provides a method for collecting and constructing a high-quality dataset for identifying distant ships, which includes the following steps:

[0036] S1. Obtain the ship's position and heading data from the dataset;

[0037] The vessel responsible for collecting the dataset is equipped with cameras or other image capture devices to acquire remote image data. Real-time positioning data of the vessel is obtained, including its position, heading, and other information.

[0038] Ship's position: The ship's current real-time latitude and longitude position is obtained through the ship's navigation system (such as GPS).

[0039] Ship's heading: Determine the ship's heading (the angle with due north), that is, the direction the bow is facing, which is very important for determining the field of view of the cameras on the ship.

[0040] S2. Obtain nautical chart data for the current voyage based on vessel position data;

[0041] Nautical chart data is a vital information resource for marine navigation and navigation safety, primarily used to describe the geographical, physical, and environmental characteristics of marine waters. Nautical chart data includes a wide range of information, such as water depth data, channel markers, lighthouses, buoys and navigation marks, currents and tides, latitude and longitude grids (used to denote geographical locations), collision avoidance markers, marine meteorological information, boundaries, and national borders.

[0042] As an officially released standardized electronic nautical chart, ENC provides authoritative, comprehensive, and up-to-date data support. SENC, as an internally optimized data format for specific ship navigation systems, achieves more efficient display and operation functions through real-time processing and dynamic updates of ENC data. In a ship's SENC navigation system, acquiring chart data for the current segment of the voyage based on the ship's position and heading facilitates the processing of limited but practical chart data, improves the timely analysis and acquisition of images of distant ships, enhances recognition efficiency and response speed, and ensures the safety and efficiency of subsequent ship navigation.

[0043] S3. Delineate the ROI areas related to ship navigation in the acquired nautical chart data;

[0044] ROI (Region of Interest) refers to a region in nautical chart data that is delineated based on features such as shipping routes, channels, and marine landmarks. The higher the density of features related to ship navigation, the more likely there will be a greater number and variety of ships.

[0045] Specifically: shipping routes are the paths taken by ships and typically have high ship density, especially near busy shipping lanes and ports. Channels are passageways for ships, usually consisting of narrow bodies of water; their width and depth are crucial for ship passage. The denser the channels, the greater the number and type of ships. Marine markers, such as lighthouses, navigational aids, and buoys, are usually located in areas requiring special attention from ships, such as channel bends, danger zones, and shallow waters. These areas are often where ships congregate because they rely on these markers for navigation. Areas with dense markers are also likely to be areas with dense ship activity, especially near ports, channel junctions, or special sea areas (such as coastal channels). Areas with dense features generally indicate a higher probability of ship presence because shipping routes, channels, and markers often indicate key shipping areas. Frequent ship activity in these areas suggests they are potential areas of high ship activity. At the intersection of channels and shipping routes, there is usually a greater variety of ship types, as these areas are likely to be frequented by large cargo ships, tankers, container ships, and other types of vessels.

[0046] For the delineation of ROI areas, one approach is to combine characteristics such as shipping routes, channel density, and marine landmarks to identify commonly used areas of channels, ports, or various types of vessels. Another approach is to combine historical vessel traffic data, obtain historical vessel location data, and correlate the probability of vessel appearance with nautical chart data to determine ROI areas.

[0047] Nautical chart data with ROI areas includes features such as shipping routes, channels, and marine landmarks, along with their corresponding geographic (latitude and longitude) coordinates. Important landmarks on nautical charts, such as shipping routes and channels, typically include geographic (latitude and longitude) coordinate information. Shipping routes are usually marked on nautical charts as a series of coordinate points, which help ships determine the optimal navigation route. The boundaries, width, and centerline of channels are clearly marked with coordinates, which is crucial for navigation and collision avoidance. The locations of marine landmarks (such as lighthouses and buoys) are usually also provided with accurate latitude and longitude coordinates, serving as reference points for ship navigation.

[0048] The boundaries of a Region of Interest (ROI) are defined by a series of latitude and longitude coordinates, which can be extracted from electronic nautical charts. An ROI can be described as a polygon, rectangle, or a circular area enclosed by multiple latitude and longitude points, or it can be a specific target.

[0049] S4. Determine whether the defined ROI area is within the field of view of the ship's camera. If yes, perform remote image capture; otherwise, abandon capture and continue to determine.

[0050] Determining whether a Region of Interest (ROI) in the acquired nautical chart data appears within the camera's field of view involves converting the camera's field of view to the nautical chart coordinate system and then comparing the ROI with the camera's field of view. If there is spatial overlap, the ROI is considered to be within the camera's field of view, and image capture is performed to generate collectable data. Otherwise, if the ROI is not within the camera's field of view, capture is abandoned. Specifically, this includes:

[0051] S41. Determine the camera's field of view information.

[0052] Camera field of view: This determines the camera's field of view, which is typically the horizontal and vertical angles. The field of view determines the size of the area that the camera can cover.

[0053] Camera installation location: Obtain the camera's installation location and orientation. This typically includes the camera's bearing relative to the ship (ship's heading information), as well as the camera's elevation or depression angle.

[0054] S42. Convert the camera's field of view to the nautical chart coordinate system;

[0055] Field of view calculation: Based on the camera's installation location and field of view angle, calculate the area visible to the camera. This can be achieved by using geometric projection to convert the field of view from the camera's coordinate system to a geographic coordinate system (latitude and longitude) suitable for nautical charts. Assuming the camera is facing a certain direction and the field of view angle is a certain value, the geometric boundary points of the field of view can be calculated using a mathematical model.

[0056] Field of view boundary: Based on the ship's current position and heading, combined with the field of view angle, the area that the camera can cover is calculated. The boundary of the field of view will then be a geometric shape composed of latitude and longitude coordinates. It is usually a slanted rectangle rather than a square. The slant of the field of view boundary is mainly due to the scale differences between latitude and longitude and the influence of the Earth's three-dimensional shape. The Earth is not a perfect sphere, but a slightly flattened ellipsoid, which is one of the fundamental reasons for this effect. The boundary of the field of view will vary in each heading direction.

[0057] Assuming the ship's current position: latitude 30°N, longitude 45°E, the camera's field of view is 30°, and the camera's installation height is 10 meters, with a heading of 0° (due north), the area covered by the camera can be calculated using the following steps:

[0058] 1. Calculate the camera's line of sight.

[0059] Line-of-sight distance (LAS) refers to the maximum horizontal distance that a camera can see. The camera is 10 meters high, and the Earth's radius is approximately 6,371 kilometers. We use the following formula to calculate the camera's LAS:

[0060]

[0061] Where: R Earth =6,371,000 meters (Earth's radius), h=10 meters (camera height).

[0062] Substituting the values ​​into the calculation, the distance is approximately 11,292 meters. Therefore, the camera's line-of-sight distance d is approximately 11.3 kilometers.

[0063] 2. Calculate the camera's field of view.

[0064] The camera's field of view is 30°, so we can calculate the width of the field of view over the ground. Because the field of view is 30°, the camera's field of view forms a fan-shaped area.

[0065] The field of view w of a camera can be calculated using the following formula:

[0066]

[0067] Where: d = 11,292 meters (view distance), θ = 30° (field of view angle).

[0068] Substituting the values ​​into the calculation, the camera's field of view, w, is approximately 6.048 kilometers.

[0069] Considering a heading of 0° (due north), the latitude and longitude coordinates of the four boundary vertices of the ship's current field of view can be calculated using the following steps.

[0070] Calculation steps:

[0071] The upper and lower boundaries in the due north direction: the height of the field of view is moved north and south respectively (11.3 km), so the latitude of the upper and lower boundaries is calculated.

[0072] East-west boundary: The field of view is 6.048 kilometers wide. It is evenly distributed in the east and west directions, and the longitude of the east and west boundaries is calculated separately.

[0073] Calculation formula: Each degree of latitude is approximately 111 kilometers. The length of each degree of longitude varies with latitude. At 30° latitude, each degree of longitude is approximately 111 × cos(30°) ≈ 96.3.

[0074] Calculate the upper and lower boundaries (latitude):

[0075] Northward: Latitude increases by approximately 0.1018° (11.3 / 111 ≈ 0.1018°)

[0076] Southward: Latitude decreases by approximately 0.1018° (11.3 / 111 ≈ 0.1018°)

[0077] Calculate the east-west boundary (longitude):

[0078] Eastward: Longitude increases by 6.048 / 96.3 ≈ 0.0628°

[0079] Heading west: Longitude decreases by approximately 0.0628° (6.048 / 96.3).

[0080] Therefore, the latitude and longitude coordinates of the four corner vertices of the camera's field of view are as follows:

[0081] Northeast corner: Latitude 30.1018°N, Longitude 45.0628°E;

[0082] Northwest corner: Latitude 30.1018°N, Longitude 44.9372°E;

[0083] Southeast corner: Latitude 29.8982°N, Longitude 45.0628°E;

[0084] Southwest corner: Latitude 29.8982°N, Longitude 44.9372°E.

[0085] These vertices represent the four corners of the field of view, and the area enclosed by the lines connecting the four vertices is the camera's field of view at this time.

[0086] S43, Comparison of ROI region with camera field of view;

[0087] Overlap detection between field of view and ROI: Once the latitude and longitude coordinate boundaries of the camera's field of view are determined, they can be compared with the latitude and longitude coordinate boundaries of the ROI on the nautical chart. Specifically, it is necessary to determine whether the coordinates of the ROI are within the camera's field of view. Spatial overlap detection can be used: using geometric methods, it is determined whether the coordinates of the ROI fall within the camera's field of view boundary. This can be achieved by calculating whether the points in the ROI are within the geometric range of the field of view.

[0088] Furthermore, as the ship's position and orientation are constantly changing, the camera's field of view will also change accordingly. Therefore, it is necessary to acquire the ship's position and heading in real time and dynamically calculate the camera's field of view area in order to continuously compare it with the ROI area.

[0089] In addition, during data collection, it is necessary to ensure that the images captured of distant ships include different types of vessels, such as merchant ships, fishing boats, and warships, as well as different weather conditions, such as sunny days, rainy days, foggy days, and nighttime.

[0090] S5. Select different types of ROI region images from the captured remote images and perform stratified sampling to form the initial dataset;

[0091] Different types of ROIs are defined based on the probability of ship appearance calculated from historical data, dividing ocean areas in nautical charts into three levels: high-probability ROIs, medium-probability ROIs, and low-probability ROIs. According to the ROI classification, stratified sampling is performed from distant images captured in that sea area to ensure a reasonable number of images are selected from high-probability, medium-probability, and low-probability ROIs. Selecting images from different types of regions allows for the acquisition of diverse datasets.

[0092] S6. Evaluate the quality of the images in the initial dataset and remove low-quality images;

[0093] Low-quality images include blurry images, exposure problems, distortion, occlusion, images that are too small, duplicate content, irrelevant backgrounds, compression artifacts, and corrupted images. Specifically:

[0094] Blurry images: Due to factors such as inaccurate focus, motion blur, or lens blur, the details in the image are unclear and difficult to identify.

[0095] Exposure issues: Images are overexposed or underexposed, resulting in information loss or difficulty in identification.

[0096] Distortion: Due to lens or shooting angle issues, images exhibit geometric distortion, such as fisheye effect and barrel distortion.

[0097] Occlusion: Key information in an image is obscured, preventing the target from being fully displayed.

[0098] Too small a size: The image resolution is low and cannot provide enough information for training.

[0099] Duplicate content: Images that are highly similar to other images increase the redundancy of the dataset.

[0100] Irrelevant background: An image background that is too complex or contains irrelevant information may interfere with the training of the model.

[0101] Compression artifacts: Due to excessive compression, images may appear blocky or distorted.

[0102] Damaged image: The image file is corrupted, causing it to fail to display properly or contain obvious errors.

[0103] The low-quality images mentioned above are usually caused by shooting conditions, equipment problems, or improper post-processing. By removing low-quality images, the overall quality of the dataset can be improved, thereby improving the model training effect.

[0104] S7. Annotate the selected images in conjunction with electronic chart data, and mark all visible ships;

[0105] Annotate the selected images, marking all visible vessels. Annotated images include marking vessel locations, vessel bounding boxes, and types. Verify the accuracy of the annotations using electronic chart data, ensuring that the marked vessel locations are consistent with the chart data.

[0106] S8. Perform data processing on the annotated images to increase the diversity and quantity of the target dataset;

[0107] Data processing is performed on the annotated images, including geometric transformations, color space transformations, and noise addition. Specifically:

[0108] Geometric transformations: These are geometric transformations performed on images, such as rotation, scaling, and translation, to increase the diversity of the dataset.

[0109] Color space transformation: This involves transforming the color space of an image, such as converting from RGB (red, green, blue) to HSV (hue, saturation, value), enhancing the model's feature separation capabilities. The RGB color space mixes hue, saturation, and value, making some color features difficult to analyze independently. For example, in image segmentation or object detection tasks, changes in lighting can interfere with the model's color perception. HSV separates hue (H), saturation (S), and value (V), allowing the model to more accurately identify color features without being affected by changes in brightness. This is especially important when processing images under uneven or low-light conditions. In complex environments (such as foggy, rainy, or nighttime), RGB images may lose features due to lighting and noise interference. By converting to a color space like HSV, the model can more robustly handle changes in brightness and color distortion, improving the model's task specificity. The HSV color space is more suitable for color segmentation tasks because its hue (H) channel is sensitive to color changes, enabling more precise segmentation of regions with specific colors. In object detection scenarios, HSV can more accurately identify targets, thereby improving detection accuracy.

[0110] Adding noise involves adding different types of noise to an image, such as Gaussian noise and salt-and-pepper noise. Gaussian noise is a type of statistical noise whose probability density function follows a Gaussian distribution (normal distribution). The steps to generate Gaussian noise are as follows: determine the mean (usually 0) and standard deviation (which controls the intensity of the noise), use a random number generator to generate a noise matrix of the same size as the image according to the Gaussian distribution, and add the generated noise matrix to the original image.

[0111] Salt and pepper noise refers to randomly appearing black and white pixels in an image. The steps to generate salt and pepper noise are as follows: determine the proportion of noise (i.e., the proportion of noise pixels in the image), randomly select pixels in the image, and set the values ​​of these pixels to the maximum value (white) or the minimum value (black).

[0112] By adding this noise to the training data, the model can learn to identify and extract features in noisy environments, improving its robustness, generalization ability, and adaptability under various noise conditions, thereby achieving better performance in practical applications.

[0113] Furthermore, generative adversarial networks (GANs) can be used to generate synthetic images from annotated images to augment the dataset. These images are created based on annotated images and nautical chart data. Examples include:

[0114] GAN Model Design: A generator network is designed to generate synthetic images from random noise and annotation information; a discriminator network is designed to distinguish between real and generated images. Annotation information (such as ship position and type) is encoded into a format usable by the generator, such as one-hot encoding or embedding vectors. Nautical chart data (including water depth, navigation marks, channel boundaries, etc.) is overlaid with the annotated image to form a composite image containing geographic information. The preprocessed annotated image, nautical chart data, and random noise are used as input to the generator. The generator and discriminator are trained alternately; the generator attempts to generate realistic images, while the discriminator attempts to distinguish between real and generated images. An appropriate loss function, such as binary cross-entropy loss, is used to optimize the model. Hyperparameters such as learning rate and batch size are adjusted based on performance during training. Using the trained generator, synthetic images are generated based on the new annotation information and nautical chart data. Post-processing is applied to the generated images, such as sharpening and color correction, to improve image quality. The generated synthetic images are integrated into the original dataset for model training. By following the steps above, GANs can be effectively used to generate synthetic images from annotated images and nautical chart data, thereby enhancing the dataset and improving the performance of ship behavior analysis and prediction models.

[0115] Furthermore, the enhanced data is cleaned to remove images of poor quality or with poor composite effects.

[0116] S9. Balance the image data of ships and non-ships in the target dataset;

[0117] Ensure the dataset contains a reasonable ratio of images of ships to non-ships, ideally close to 1:1. This is because if the dataset only contains images of ships, the model's training and performance will be negatively impacted as follows:

[0118] Model bias: The model only learns the features of ships and cannot distinguish between ships and non-ships. This will cause the model to fail to classify correctly when faced with a test set containing non-ship images, because the model has not learned the features of non-ships.

[0119] Overfitting: Because the model only encounters images of ships, it may overfit the features of those images, resulting in poor performance on new, unseen images of ships. The model's generalization ability will be very limited.

[0120] Performance cannot be evaluated: Without non-ship images, it is impossible to accurately evaluate the model's performance in real-world scenarios, as real-world data typically includes both ships and non-ships.

[0121] Practical limitations: In real-world applications, the model needs to be able to distinguish between ships and non-ships. If the model fails to learn the characteristics of non-ships, it will not be able to work reliably in real-world environments.

[0122] To build an effective model for recognizing distant ships, it is necessary to ensure that the dataset includes images of both ships and non-ships, so that the model has higher generalization ability and robustness in ship and non-ship classification tasks.

[0123] S10. Use cross-validation to evaluate the quality of the improved target dataset.

[0124] It also includes using cross-validation to evaluate the quality of the constructed target dataset and taking corresponding measures to improve the quality of the dataset.

[0125] Furthermore, it also includes storing the processed data in a structured database and indexing the data, including location information, time information, weather conditions, sea conditions, etc., to optimize dataset performance and facilitate subsequent model use.

[0126] As another aspect of the invention, furthermore, fine-grained annotations can be performed on the ships in the constructed target dataset images, including information such as ship type, size, heading, and speed. A ship trajectory dataset containing temporal information is constructed for distant ship behavior analysis and prediction. This includes: calculating the ship's speed and heading based on position and timestamp; calculating the ship's acceleration based on speed and timestamp; identifying and extracting waypoints such as start point, end point, and turning point; calculating the ship's distance between different waypoints; and calculating statistical features such as average speed, average acceleration, maximum speed, and maximum acceleration within a certain time window. The entire ship trajectory is segmented according to time or waypoints, with each segment representing an independent trajectory fragment. For each trajectory segment, the time-series data is organized into a sequence, such as: (Position 1, Speed ​​1, Heading 1, Time 1), (Position 2, Speed ​​2, Heading 2, Time 2), ..., (Position n, Speed ​​n, Heading n, Time n). Based on the ship's behavior patterns, such as sailing, berthing, anchoring, docking, and departing, a behavior label and a predicted future position or speed label are generated for each trajectory segment. This constructs a ship trajectory dataset containing time-series information for the analysis and prediction of distant ship behavior. This dataset can provide important data support for fields such as maritime safety, traffic management, and shipping optimization.

[0127] This invention acquires nautical chart data with Regions of Interest (ROIs) and optimizes image data acquisition by utilizing this data. Because nautical chart data contains a wealth of information on marine landmarks, routes, and channels, ROIs can be quickly delineated from a distance based on the features within the chart data. Image data can then be acquired and subsequently extracted in high, medium, and low ROI areas, significantly improving the probability, classification, and clarity of captured ship images. This provides a high-quality dataset for training distant ship image recognition models and also enhances the accuracy of later application tasks such as ship monitoring, collision avoidance decision-making, and traffic flow prediction, ensuring a safe navigation environment.

[0128] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0129] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for collecting and constructing a high-quality dataset for identifying distant ships, characterized in that, It includes the following steps: S1. Obtain the ship's position and heading data from the dataset; S2. Obtain nautical chart data for the current voyage based on vessel position data; S3. Delineate the ROI areas related to ship navigation in the acquired nautical chart data; S4. Determine whether the defined ROI area is within the field of view of the ship's camera. If yes, perform remote image capture; otherwise, abandon capture and continue to determine. S5. Select different types of ROI region images from the captured remote images and perform stratified sampling to form the initial dataset; S6. Evaluate the quality of the images in the initial dataset and remove low-quality images; S7. Annotate the removed images in conjunction with electronic nautical chart data, and mark all visible vessels; S8. Perform data processing on the annotated images to increase the diversity and quantity of the target dataset.

2. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 1, characterized in that, In S3, the ROI area related to ship navigation refers to the area in the nautical chart data that is delineated based on the characteristics of the route, channel, and marine landmarks in the nautical chart data, according to the principle that the higher the density of features related to ship navigation, the greater the probability of the number and types of ships appearing.

3. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 2, characterized in that, The delineation of the ROI area includes two aspects: one is to combine factors such as shipping routes, channel density, and marine landmark characteristics to delineate the ROI area by identifying channels, ports, and areas commonly used by various types of ships; the other is to combine historical ship traffic data, obtain historical ship location data, and correlate the probability of ship appearance with nautical chart data to determine the ROI area.

4. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 3, characterized in that, S4, determining whether the defined ROI region appears within the field of view of the ship's camera, if yes, performs remote image capture; otherwise, abandons capture and continues the determination, including: S41. Determine the field of view information of the ship's camera; S42. Convert the camera's field of view to the nautical chart coordinate system; S43. Compare the ROI area marked on the nautical chart with the camera's field of view; S44. Based on the comparison results, determine whether the defined ROI area appears within the field of view of the ship's camera. If so, perform remote image capture; otherwise, abandon capture and continue the judgment.

5. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 4, characterized in that, The field-of-view information of the ship's camera in S41 includes the camera's field-of-view angle, installation position, and orientation.

6. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 1, characterized in that, The different types of ROI regions in S5 are based on the probability of ship appearance calculated from historical data, dividing the ocean areas on the nautical chart into three levels: high probability ROI regions, medium probability ROI regions, and low probability ROI regions.

7. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 1, characterized in that, Low-quality images in S6 include blurry images, exposure problems, distortion, occlusion, images that are too small, duplicate content, irrelevant backgrounds, compression artifacts, and damaged images.

8. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 1, characterized in that, In step S7, the removed images are annotated using electronic nautical chart data, marking all visible vessels, including: The removed images are annotated to mark all visible ships. The annotated images include the ship locations, ship bounding boxes, and types. Verify the accuracy of the annotations by combining them with electronic nautical chart data to ensure that the annotated vessel positions are consistent with the nautical chart data.

9. The method for collecting and constructing a high-quality dataset for distant vessel identification according to claim 1, characterized in that, The method also includes S10, using cross-validation to evaluate the quality of the improved target dataset.

Images