A ship collision avoidance method and system based on multi-source perception fusion

Abstract

The application provides a ship collision avoidance method and system based on multi-source perception fusion, and relates to the technical field of intelligent shipping and ship autonomous navigation. The collision avoidance method specifically comprises the following steps: collecting video images of ships and navigation marks in a channel by using a drone, calculating a background water flow velocity in the channel, and reconstructing a static background water flow field covering the channel; based on the video images, water dynamic force features around the ship are extracted; the extracted water dynamic force features are physically corrected by using the static background water flow field to obtain an intrinsic water dynamic force parameter set; based on the intrinsic water dynamic force parameter set and a real-time background flow field, an actual dynamic water dynamic force interference field of a target ship is constructed and predicted, a navigation decision including an expected trajectory and a predictive attitude adjustment instruction is generated; a residual error between an actual motion state of the ship and the navigation decision is monitored; when the residual error satisfies a preset triggering condition, the static background water flow field is updated and the intrinsic water dynamic force parameter set is online corrected.

Description

Technical Field

[0001] This invention belongs to the field of intelligent shipping and autonomous ship navigation technology, specifically relating to a method and system for environmental field reconstruction, multi-target situation assessment and autonomous collision avoidance decision-making using multi-source heterogeneous data in complex water environments such as narrow waterways in ports. Background Technology

[0002] With the continued growth of global trade, the traffic density in ports and busy nearshore waters is increasing daily, posing a serious challenge to ship navigation safety. Especially in narrow waterways such as port access channels, the navigation environment is extremely complex. Constrained by the geographical limitations of these waterways, variable microclimates, and high-density ship traffic, collisions, groundings, and reef strikes occur frequently. To improve navigation safety and efficiency, autonomous surface vessel technology has become a research hotspot in the maritime industry. Among these technologies, high-precision environmental perception and intelligent collision avoidance decision-making are the core technologies for achieving autonomous ship navigation.

[0003] Existing ship perception technologies primarily rely on sensors such as Automatic Identification Systems (AIS), marine radar, and visible / infrared cameras to acquire information on the position and motion of surrounding obstacles. However, these single or simple combined perception methods have significant limitations in complex port scenarios. AIS depends on the target ship's active broadcasting, which carries the risk of data delay, packet loss, and even human intervention or deception. Marine radar is easily interfered with by shorelines, waves, and rain / snow clutter in narrow waterways and has blind spots. While visual sensors can provide rich texture information, they are greatly affected by lighting and fog, and it is difficult to directly obtain precise distance and speed information of targets. More importantly, most existing multi-sensor fusion algorithms are limited to simple superposition of data or feature layers, lacking in-depth mining and cross-validation of the physical and logical relationships between different modalities, resulting in insufficient robustness in extreme environments.

[0004] Furthermore, when ships navigate in ports or narrow waterways, their motion depends not only on their own maneuvering commands but also on severe environmental loads such as wind, currents, and waves. Existing environmental sensing methods typically rely on shipborne anemometers, logs, or data reported from shore-based weather stations. However, shipborne sensors can only acquire localized, single-point data for the ship itself and cannot perceive the flow field distribution in the channel ahead or the surrounding waters; shore-based data suffers from low spatial resolution and delayed updates.

[0005] In collision avoidance decision-making, most existing automatic collision avoidance algorithms simplify ships as moving point masses on a two-dimensional plane, focusing primarily on geometric avoidance. This simplified model ignores the complex hydrodynamic effects generated when ships move as large entities in water. For example, large ships navigating in narrow waterways generate propeller wakes, as well as bank effects and inter-ship suction and discharge forces caused by the displacement of water. These hydrodynamic disturbances not only affect the ship's maneuverability but also pose serious safety threats to nearby smaller vessels. Current collision avoidance systems lack the ability to predict such hidden disturbance fields and cannot, like experienced human navigators, anticipate the disturbance effects of other vessels' wakes or wave-making on the ship's future trajectory.

[0006] In summary, overcoming the shortcomings of existing technologies in deep fusion of multi-source heterogeneous data, accurate reconstruction of microscopic environmental flow fields, prediction of hydrodynamic interactions, and closed-loop correction of perception-decision, and developing an autonomous collision avoidance method and system for ships that can adapt to the complex environments of ports and narrow waterways, is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0007] This invention discloses a ship collision avoidance method based on multi-source perception fusion, which includes the following steps: By using drones to collect video images of ships and navigation marks in the waterway, the background water flow velocity in the waterway is calculated, and the static background water flow field covering the waterway is reconstructed. Based on video images, hydrodynamic features around the ship are extracted; the extracted hydrodynamic features are physically corrected using the static background water flow field to obtain the intrinsic hydrodynamic parameter set. Based on the intrinsic hydrodynamic parameter set and the real-time background flow field, the actual dynamic hydrodynamic disturbance field of the target ship is constructed and predicted, and a navigation decision containing the desired trajectory and predictive attitude adjustment commands is generated. Monitor the residual between the ship's actual motion state and the navigation decision; when the residual meets the preset trigger conditions, update the static background water flow field and make online corrections to the intrinsic hydrodynamic parameter set.

[0008] The steps for using drones to collect video images of ships and navigation marks within a waterway include: using ship detection boxes extracted based on a feature pyramid network, mapping the endpoints of the detection boxes from the image coordinate system to the world geographic coordinate system to obtain the ship... First geographical coordinates With tail geographic coordinates ; Calculate ships length And the Froude number, which characterizes a ship's wave-making ability. : in, To track the ground velocity vector obtained by the displacement of the ship's center of mass within a continuous time window, This is the acceleration due to gravity.

[0009] In the step of extracting hydrodynamic features around a ship based on video images, in order to eliminate navigation marks that are interfered with by the ship, the ship length is used. and Froude's number Determine the vessel Dynamic influence domain Its longitudinal length threshold With horizontal width threshold The calculation formula is: in, The preset longitudinal hydrodynamic attenuation constant, It is the transverse diffusion constant; For any navigation mark Calculate the position vector from the ship to the navigation mark. ,in The geographical coordinates of the navigation beacon. The coordinates of the ship's center of mass; when the position vector Falling into the dynamic influence domain If the time limit is reached, the navigation mark is determined to be a disturbed navigation mark; otherwise, it is determined to be an isolated navigation mark.

[0010] The steps for extracting hydrodynamic features around a ship based on video images also include: calculating the relationship between the UAV and the ship. The observation azimuth angle is used to extract visual hydrodynamic features in the head-on or tail-on observation interval; When the ship is in the bow observation range, extract the half-diffusion angle of the wave crest lines on both sides of the bow relative to the ship's longitudinal axis. And calculate the observed wave intensity : in, The set of detected peak line pixels, The number of pixels. The image grayscale gradient vector; When in the tail-chasing observation range, the wake region is segmented using color space transformation, and the observed wake length in the opposite direction of the ship's course is extracted. .

[0011] The steps for physically correcting the extracted visual hydrodynamic features to obtain the intrinsic hydrodynamic parameter set include: Obtaining ships from static background water flow field Background velocity vector at the centroid And calculate the relative velocity vector of the ship. ; Using relative velocity vectors to observe the half-diffusion angle After performing flow field correction, the intrinsic half-diffusion angle, which does not change with the flow field, is obtained. The calculation formula is as follows: The step of physically correcting the extracted visual hydrodynamic features to obtain the intrinsic hydrodynamic parameter set also includes: Based on the convection effect correction model, the length of the observed wake is determined using the relative velocity vector. By performing transport corrections to eliminate visual errors of downstream stretching or upstream compression, the intrinsic wake impact length is obtained. : in, The turbulence intensity attenuation index, The flow coupling coefficient is the coefficient of mass. This is a unit vector in the direction of the stern. Intrinsic half-diffusion angle and intrinsic wake impact length Together they constitute the intrinsic hydrodynamic parameter set.

[0012] The steps to calculate the background water velocity within the channel include: Extract arbitrary navigation marks from isolated navigation marks. average displacement vector Establish a drift-velocity mapping function and calculate the background velocity vector at that location. : in, The equivalent horizontal stiffness coefficient of the navigation mark. For the density of water, The drag coefficient of the water section of the navigation mark. To effectively immerse the projected area, This refers to the wind-induced drift component calculated using wind field data.

[0013] The steps for generating a navigation decision that includes the desired trajectory and predictive attitude adjustment commands include: calculating the lateral wave impact moment that the ship will experience when traversing the actual dynamic hydrodynamic disturbance field of the target vessel at the meeting point. ; Calculate the required predictive anti-disturbance pressure rudder angle As the attitude adjustment command for navigation decision-making, its calculation formula is: in, This is the rudder effectiveness coefficient of the ship. This is the ship's speed relative to the water.

[0014] This invention also discloses a ship collision avoidance system based on multi-source perception fusion, the system comprising: Background water field reconstruction module: Uses drones to collect video images of ships and navigation marks in the waterway, calculates the background water flow velocity in the waterway, and reconstructs the static background water flow field covering the waterway. Parameter set reconstruction module: Based on video images, extract hydrodynamic features around the ship; use static background water flow field to physically correct the extracted hydrodynamic features to obtain the intrinsic hydrodynamic parameter set; The measurement generation module: Based on the intrinsic hydrodynamic parameter set and the real-time background flow field, it constructs and predicts the actual dynamic hydrodynamic disturbance field of the target ship, and generates a navigation decision containing the desired trajectory and predictive attitude adjustment commands. Online correction module: monitors the residual between the ship's actual motion state and the navigation decision; when the residual meets the preset trigger conditions, it updates the static background water flow field and corrects the intrinsic hydrodynamic parameter set online.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned ship collision avoidance method based on multi-source perception fusion.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned ship collision avoidance method based on multi-source perception fusion.

[0017] The ship collision avoidance method and system based on multi-source perception fusion provided by this invention has achieved significant technological progress and beneficial effects in complex navigation environments with high density and strong interference, such as ports and narrow waterways.

[0018] This invention significantly improves the accuracy and completeness of perception of dynamic hydrological environments in ports. It reveals the true hydrodynamic disturbance capabilities of navigating vessels to the surrounding waters. In port navigation, the danger posed by a vessel lies not only in its physical location but also in the hydrodynamic disturbances it generates, such as wave-making and wakes. Existing technologies often rely solely on the vessel's ground speed or raw visual characteristics for judgment, leading to serious misjudgments in the presence of background currents. This invention introduces a reconstructed background flow field to physically correct the hydrodynamic features such as ripples and wakes extracted by UAV vision, calculating an intrinsic set of hydrodynamic parameters that remains unchanged with the environmental flow field. This enables the system to accurately quantify the true disturbance potential of any target vessel, regardless of its complex navigation conditions, such as downstream, upstream, or crosscurrent, thus providing an objective and stable physical basis for subsequent risk assessment.

[0019] Based on the aforementioned precise perception and profound understanding, this invention generates a forward-looking intelligent navigation decision that balances geometric safety and hydrodynamic stability. Traditional collision avoidance methods primarily focus on avoiding geometric collisions between ship hulls, neglecting the risks of hydrodynamic interactions during close encounters. This invention utilizes a decoupled set of intrinsic hydrodynamic parameters, combined with real-time background flow fields, to accurately predict the actual range and intensity of the disturbance field generated by the target vessel at the moment of future encounter. The dynamically constructed, asymmetric safety domain, along with the generated attitude adjustment commands including predictive anti-disturbance rudder angles, transforms collision avoidance decision-making from passive evasion to proactive control. This not only effectively prevents major safety hazards such as severe rolling due to wave-making or stall due to wake turbulence, but also makes the ship's navigation trajectory smoother and more stable, significantly improving navigation safety and efficiency in high-density traffic flows.

[0020] Port environments are complex and ever-changing, making it difficult for any fixed model to cover all operating conditions. This invention, by monitoring the execution residuals of navigation commands in real time, can keenly detect deviations between the model and reality. Utilizing the vessel itself as a mobile truth sensor, the system can not only calibrate and refine the background flow field map online, but also continuously optimize its understanding of hydrodynamic interference models from different types of ships based on actual responses during rendezvous. This closed-loop correction mechanism allows the overall performance of the system to continuously improve with operating time, constantly adapting to new environments and objectives, ultimately achieving a more robust and reliable autonomous navigation capability that surpasses human experience. Attached Figure Description

[0021] Figure 1 This is a flowchart of the ship collision avoidance based on multi-source perception fusion according to the present invention. Detailed Implementation

[0022] This embodiment is applied to port access channels with restricted waterway characteristics. These areas have complex hydrological environments, including significant periodic tidal currents and non-constant wind fields influenced by topography. Floating navigational aids are deployed on both sides of the channel to mark its boundaries. The system's physical architecture comprises two main parts: a shipborne main control subsystem and an airborne visual perception subsystem. These two subsystems communicate in real-time via a high-bandwidth, low-latency wireless data link.

[0023] The airborne visual perception subsystem mainly consists of one or more industrial-grade multi-rotor unmanned aerial vehicles (UAVs) accompanying the ship. These UAVs preferably employ a six-rotor flight platform with a wind resistance rating of level six or higher to ensure hovering and flight stability in the complex micro-meteorological environment of the port. Regarding the perception payload configuration, the UAV is equipped with a three-axis stabilization gimbal mounted on its underside. The gimbal integrates a high-resolution visible light optical imaging module, preferably a 4K resolution zoom camera with a frame rate of at least 60fps. Its field of view (FOV) is adjustable, used to acquire high-definition video streams within the waterway in real time. In addition, the UAV integrates a high-precision inertial measurement unit (IMU) to measure the relative airflow speed and direction at the UAV's altitude in real time during flight. Combined with the UAV's own attitude and ground speed vectors provided by the IMU, absolute wind field data is calculated. The UAV is also equipped with an RTK-GNSS positioning module to obtain its own absolute geographical location information, providing a benchmark for the georeferencing of visual data. The flight strategy for the unmanned aerial vehicle is set to fly alongside the ship or hover in a fixed position within an airspace 0.5 to 1 nautical mile in front of the bow of the ship. The flight altitude is preferably set between 50 and 150 meters above sea level to obtain the best overhead view covering the entire channel section.

[0024] The core of the shipborne main control subsystem is a high-performance edge computing processing terminal. This terminal preferably adopts a CPU and GPU heterogeneous computing architecture, capable of real-time decoding of multiple high-definition video streams, optical flow feature extraction, and hydrodynamic numerical calculation. The shipborne main control subsystem is physically connected to the ship's integrated navigation system via the shipborne local area network, reading the ship's motion status data and dynamic parameters in real time. The motion status data includes at least the heading value obtained via fiber optic or satellite compass, the speed over water obtained via a speedometer, and the heading at check (COG) obtained via dual-antenna GNSS. Furthermore, the shipborne main control subsystem is connected to an electronic chart display and information system, used to visually overlay and display the calculated flow field layers, navigation advice commands, and predicted collision avoidance paths on the electronic chart for use by the pilot or autopilot. Regarding communication links, the shipborne end is equipped with a directional high-gain microwave communication antenna or a 5G CPE terminal, establishing a point-to-point or cellular network-based data transmission channel with the airborne end, ensuring millisecond-level synchronous transmission of visual image data and control commands.

[0025] In this invention, floating navigation aids in the waterway environment, specifically referring to standard light buoys or cylindrical buoys anchored to a bottom boulder by anchor chains, are considered passive sensing objects. These buoys have known geometric dimensions, shape parameters, and design waterline height, and are treated as distributed surface probes during system operation.

[0026] Next, this embodiment details the specific implementation process of the static background field construction step in the ship collision avoidance method based on multi-source sensing fusion. The core logic of this step is as follows: using physical navigation marks distributed within the port channel as a distributed sensor array, by establishing a spatial topological mapping relationship between the navigation marks and navigating ships, the navigation marks are divided into disturbed sets and isolated sets. A background current field reflecting natural hydrological characteristics is constructed based on the inversion data of the isolated sets, and the impact information of ships on the water field is extracted based on the residual analysis of the disturbed sets. S1.1 Navigation Marks-Ships Spatial Topology Mapping and State Clustering In the high-density navigation environment of port entry and exit channels, it is necessary to construct a dynamic correlation between navigation marks and ships based on video streams collected by unmanned aerial vehicles (UAVs). The system operates in real-time... Identify all sets of navigation marks within the field of view And all ships at sea ,in, For the i-th navigation mark, N is the total number of navigation marks. For the j-th ship, M represents the total number of ships.

[0027] At any moment Identify all sets of navigation marks within the field of view And all ships at sea During the process, considering that the visual environment in port and narrow waterway scenarios has the characteristics of uncertain sea-sky boundary, significant dynamic changes in illumination, and large differences in target scale (i.e., extremely small distant navigation marks and large nearby ships), this embodiment adopts a saliency multi-scale detection method based on sea-sky boundary constraints.

[0028] S1.1.1 Sealine segmentation and water area mask construction based on texture gradient energy To eliminate interference from shoreline buildings and cloud cover on water surface target detection and reduce computational redundancy, effective water areas are first extracted based on scene texture features. The port water surface has unique wave texture features, which are represented by high-frequency components in the frequency domain, while the sky region is represented by low-frequency flat areas.

[0029] Input time Original image frame Define the first in the image Column, No. The pixel grayscale value of the row Calculate its longitudinal gradient magnitude Constructing row-level texture energy functions : in, Image width, This is the mean of the gradient for that row. Due to the presence of water surface wave texture, The energy exhibits high-energy oscillations within the water region, while approaching zero in the sky region. The system determines the horizontal coordinates of the sea-line by searching for energy abrupt change points. : based on Generate a binarized water area of ​​interest mask Only retain The pixel region is used as the search space for subsequent detection, effectively filtering out false detection interference from non-water surface targets such as shore cranes.

[0030] S1.1.2 A set of navigational aids based on color-spatial saliency extract Port navigation marks (collection) Following the standards of the International Association of Lighthouse Authorities (IALA), to achieve a specific red (port) or green (starboard) color and a small image size, the image within the masked area needs to be converted to the Lab color space. Channel (red-green component) enhances the chromaticity characteristics of the beacon. Define pixels. Navigational beacon color saliency map for: in, For pixels of Channel value, and The water surface background of the current frame Channel mean and variance.

[0031] Secondly, in order to distinguish between floating debris and real navigational aids, for candidate connected domains Calculate candidate connected components Its neighboring ring zone brightness contrast : in This represents the average brightness within the area. Only when... And color salience At that time, the area was determined to be a candidate target for navigational aids.

[0032] Finally, morphological closing operations are performed on all candidate targets to extract their centroid pixel coordinates. This constitutes a set of navigation beacon pixels.

[0033] S1.1.3 Ship Set Based on Feature Pyramid Network (FPN) extract For ship targets (set) To address the problem of targets ranging from tugboats tens of meters to container ships hundreds of meters in length, and exhibiting significant differences in appearance, this invention employs a deep convolutional neural network based on Feature Pyramid Network (FPN) for multi-scale target detection.

[0034] The deep convolutional neural network architecture uses ResNet-101 as the backbone network to extract basic feature maps, and constructs a feature pyramid with four scales through top-down paths and lateral connections. , respectively corresponding to the original image resolution For any ship candidate box in the image. The network outputs its class probability. And the positional regression offset.

[0035] For detected ship targets Define a wake verification area behind its detection frame. Calculate the gray-level entropy within this region. : in grayscale The probability of its occurrence. The propeller churning of a moving ship creates a foam zone, resulting in a complex texture and significantly higher grayscale entropy in this area compared to other water surfaces. Only when the detection box confidence is high and At that time, the target was identified as a moving vessel, thus excluding moored vessels that do not cause flow field interference. Finally, the pixel coordinates of the center point of the vessel's bottom were extracted. .

[0036] S1.1.4 Target Geographic Relocation Based on Monocular Perspective Transformation In order to establish a spatial topological mapping, the aforementioned navigation marks must be... and ships The pixel coordinates are mapped to the world geographic coordinate system.

[0037] Based on the assumption that the port water surface is approximately a plane, a monocular inverse perspective projection model (IPM) is constructed using the real-time pose information of the UAV.

[0038] Let the position of the drone camera in the world coordinate system be... The position was obtained using RTK-GNSS, and the camera's rotation matrix is... The camera intrinsic parameter matrix is For any target point on the image plane Its corresponding world coordinate direction vector for: Since the target is located on the water surface, the constraints are satisfied. Calculate the position of the target in the world coordinate system : Wherein, scaling factor The scale uncertainty of monocular vision was eliminated by utilizing the altitude information of the drone.

[0039] This allows for the extraction of the beacon pixel set from the image. Convert to world coordinate set Set up ship pixels Convert to world coordinate set .

[0040] In order to determine any number navigation mark Whether or not it is subject to the vessels To address wave-making or wake interference, an anisotropic influence domain model based on the ship is established. Considering that the Kelvin wave system generated by a ship during navigation is mainly distributed on both sides of the stern, and the wake extends aft, Defined as having the ship's center of mass The region is a combination of a semi-ellipse and a fan-shaped area extending in the opposite direction from the bow, with the origin as the starting point. and hour, .

[0041] in, and navigation marks and ships Position coordinates in the world coordinate system; For ships The heading angle; The ship's velocity vector; This is the position vector pointing from the ship to the navigation mark. and These are the longitudinal length threshold and transverse width threshold of the influence domain, respectively, and the longitudinal length threshold and transverse width threshold are related to the ship. length and Froude's number related.

[0042] In identifying and eliminating hydrodynamic interference from navigating vessels on navigation aids, it is necessary to quantify the vessel's motion state and the associated hydrodynamic influence range. This is achieved by using the Feature Pyramid Network (FPN) extracted from S1.1.3 to generate vessel detection boxes. And the monocular inverse perspective transformation model (IPM) established in S1.1.4, for any ship identified within the field of view First, the front and rear endpoints of the detection bounding box in the image coordinate system are extracted, corresponding to the front pixels of the ship. With tail pixels Utilizing the real-time pose matrix of the UAV With intrinsic parameter matrix The above pixels are mapped to the world geographic coordinate system through inverse perspective transformation, with sea level as the reference point. Based on this, the geographical coordinates of the ship's bow were obtained. With tail geographic coordinates .

[0043] Based on this, the ship is calculated length : Meanwhile, in consecutive time windows The system tracks the geographic coordinate changes of the ship's center of gravity and calculates the ship's velocity vector relative to the ground. and heading angle Among them, heading angle Defined as the angle between the velocity vector and true north. Based on the obtained length and velocity magnitude, the Froude number of the ship is calculated. Used to characterize the dynamic intensity of waves made by ships: In the formula, Let be the acceleration due to gravity, and take . .

[0044] Define dynamic influence domain Its boundary is determined by the longitudinal length threshold. With horizontal width threshold Confirmed. These two thresholds have the following functional relationship with the ship's physical parameters: In the formula, The preset longitudinal hydrodynamic attenuation constant, in the application scenario of this invention, preferably takes a value within the range of [value range missing]. to , representing the main energy attenuation distance of the wake in the longitudinal direction; The transverse diffusion constant is preferably taken as follows: .

[0045] For any navigation mark within the waterway Let its geographical coordinates be... Ships The coordinates of the centroid are The position vector from the ship to the navigation mark Defined as: System determines navigation mark Did it fall into the ship? Influence domain That is, to determine the vector Does it satisfy the constraint conditions of S1.1.4?

[0046] It should be noted that the ship's actual velocity relative to water cannot be directly obtained during initial system operation or before a background flow field is established. Therefore, in this step, the ship's velocity vector relative to ground, calculated using UAV tracking, is used. Approximate substitution for Froude number of water velocity and the area of ​​influence The initial estimate. To compensate for the estimation error caused by the lack of flow velocity, the longitudinal attenuation constant in this embodiment... and transverse diffusion constant Sufficient safety redundancy is preset to ensure maximum coverage of potentially disturbed areas and prevent navigation marks affected by waves from being misjudged as isolated navigation marks.

[0047] Based on the above determination, the system will set up the navigation beacon collection. Divided into two mutually exclusive subsets: the set of disturbed beacons The movement of navigation marks within this set is superimposed with the wave-making effect of ships. Isolated navigation mark set. The movement of the beacons within this group is driven only by the natural background water flow.

[0048] S1.2 Static background flow field inversion based on isolated navigation mark set The background current field is an objective manifestation of port tides and runoff under topographic constraints. This invention selects only isolated navigation mark sets. The data in the image is processed to obtain the background water flow velocity vector at that location. .

[0049] S1.2.1: Region of Interest (ROI) Delineation for Isolated Navigation Marks against Each isolated beacon The system in the current image frame In the middle, with the pixel centroid of the navigation mark as the center, a region of size is defined. The rectangular region of interest of pixels is denoted as . The purpose of setting the ROI is to limit the analysis scope to the immediate neighborhood of the navigation mark, where the current conditions best represent the background current at the location of the navigation mark.

[0050] exist To accurately track water surface movement driven by water flow, stable and identifiable visual feature points need to be extracted. Considering that the texture of the port water surface under different lighting and weather conditions is mainly manifested as local variations in brightness and gradient formed by tiny waves and ripples, this embodiment uses the Shi-Tomasi corner detection algorithm to extract feature points. For ROI For any pixel within the range, its Harris matrix Defined as: in, It is a neighborhood window centered on that pixel. and These are images The gradient in the horizontal and vertical directions. The Shi-Tomasi corner criterion is a matrix. smaller eigenvalues It must be greater than the preset threshold. : In each Extract a set of feature points from the water surface. ,in It is the first Each feature point in Time frame image The pixel coordinates in the image.

[0051] S2.1.2: Feature Point Tracking and Pixel Displacement Field Generation Obtain Feature point set at time Then, the system needs to process subsequent image frames. These points are tracked to calculate their pixel displacement. However, considering that fast water flow can cause large displacements of feature points between consecutive frames, the standard Lucas-Kanade optical flow method may fail.

[0052] Therefore, the present invention first constructs an image. and The Gaussian pyramid is used to calculate the initial optical flow starting from the low-resolution top-level image. The calculation result is then used as the initial value and passed down layer by layer to higher-resolution image layers for optimization, thus expanding the range of trackable motion. At each level of the pyramid, assuming constant brightness, that is: By expanding the Taylor series and neglecting higher-order terms, we can obtain the optical flow constraint equation: To solve underdetermined equations, the KLT algorithm uses a small neighborhood window. Optical flow of all pixels They are the same. The displacement vector is solved by minimizing the weighted squared error. : After multi-level pyramid calculations and optimizations, the final ROI is... Each initial feature point within Calculate its in New position of the moment This results in a set of pixel displacement vectors. ,in .

[0053] S2.1.3: Conversion of pixel displacement to world coordinate system velocity vector Received pixel displacement vector Existing in two-dimensional image space, it must be converted into a velocity vector in a three-dimensional world coordinate system. This process relies on the inverse perspective transformation (IPM) model established in this invention.

[0054] Let the intrinsic parameter matrix of the UAV camera be... Obtained from camera calibration: in, Focal length The primary coordinates are used. The drone's onboard navigation system provides its real-time position in the world coordinate system. And the pose, represented as a rotation matrix from the world coordinate system to the camera coordinate system. .

[0055] For any feature point Starting pixel coordinates in the image and termination pixel coordinates Its three-dimensional position in the world coordinate system This can be obtained by solving for the intersection points of the light rays and the horizontal plane in the camera imaging model. Calculate the three-dimensional orientation vector of this pixel in the camera coordinate system. : The light comes from the camera position Starting from, its parametric equation is: The port's water surface is in a global coordinate system. Plane, by letting When the component is 0, the parameters can be calculated. : in and They are and The Z component. Substituting back into the parametric equations, we can obtain the world coordinates corresponding to that pixel. .

[0056] right and Perform the above calculations to obtain the world coordinates. and This feature point in time Real-world velocity vector for: For ROI Repeat this process for all successfully tracked feature points to obtain a set of original world velocity vectors. ,in It represents the number of feature points that were successfully tracked.

[0057] S2.1.4: Velocity Estimation Based on RANSAC Due to factors such as water surface reflection, wave breaking, or feature matching errors, the set Inevitably, the measurements will contain erroneous values. To obtain a robust estimate of the background current in this area, this embodiment employs a random sampling consensus algorithm to obtain a set of discrete background current velocity vector observation points distributed throughout the waterway. ,in It is a navigation mark Its geographical location.

[0058] Obtain discrete velocity point sets Subsequently, to obtain a continuously distributed background current field throughout the entire waterway, the Kriging interpolation algorithm was employed. This involved defining any spatial point within the waterway. Background water flow velocity vector at the location for: in, For the first Each navigational beacon to a spatial point Spatial weighting coefficients. Considering that port channels are typically long and narrow, the correlation of water flow along the channel axis (longitudinal) is significantly stronger than that perpendicular to the channel. Therefore, this embodiment constructs a weighting function based on a Gaussian kernel, specifically in the following form: In the formula, Point to be sought coordinates navigation mark The coordinates; This is the longitudinal correlation length threshold along the waterway direction. This is a threshold for the lateral correlation length perpendicular to the waterway direction. In this embodiment, it is set as follows: and This ensures that the reconstructed flow field conforms to the physical characteristics of a flow velocity that changes gently along the channel direction but decays rapidly along the transverse direction due to the influence of the shoreline boundary.

[0059] S1.3 Extract information on the impact of the preceding vessel on the water field. To quantify the degree of disturbance that ships cause to the water, for the collection Any navigation mark in It is located on a certain ship Influence domain Inside. Simultaneously, the theoretical background velocity at that location is queried within the background water flow field generated in S2.2. .

[0060] Define navigation marks Hydrodynamic disturbance vector at the location The vector difference between the two: Characterizes the ship The intensity and direction of the wave or wake generated at this location.

[0061] In order to transform discrete disturbance observations into observations of the ship An overall evaluation was conducted, and a descriptor for the impact of ships on the water field was constructed. This descriptor applies to elements falling within its domain of influence. Weighted aggregation of all beacon disturbance vectors: in, The number of navigation marks that fall into the affected area. The preferred setting for distance attenuation weights is: .

[0062] Descriptor The output scalar value represents the ship ahead. The disturbance energy to the surrounding waters. If this value exceeds a preset threshold, it indicates that the vessel is generating severe wave-making or a powerful propeller jet.

[0063] Next, this embodiment further elaborates on the specific implementation process of interference field extraction and calibration. Unmanned aerial vehicles (UAVs) can acquire a global overhead view covering the channel cross-section. However, in high-density port traffic scenarios, visual perception faces two major challenges: Firstly, the target scale varies greatly; distant ships may occupy only a few pixels in the image, while nearby ships occupy most of the frame, making it difficult for a unified feature extraction algorithm to handle both. Secondly, there is severe mutual occlusion, especially when multiple ships are parallel or passing each other; a large container ship may completely obscure a small pilot boat behind it.

[0064] Therefore, the three observation intervals—head-on, tail-to-tail, and lateral—divided in this embodiment are for each identified vessel within the field of view. Dynamic logical partitioning. The purpose is to determine the relationship between the drone and a specific target ship. The system adaptively selects the clearest, most reliable, and least obstructed visual features from the current viewpoint based on the relative azimuth angle between the UAV and the target vessel. When the UAV is directly in front of a ship, the Kelvin wave system features of the ship's bow are most significant and unobstructed; in this case, the system should prioritize activating the wave system feature extraction algorithm. However, for another ship that is simultaneously in the field of view but located behind the target vessel, the system may choose to activate the wake feature extraction, which is visible to the target vessel, because its bow is obstructed. This adaptive logic ensures that the most effective local hydrodynamic features can be extracted for each target of interest, regardless of the UAV's global position, greatly improving the robustness of the algorithm in complex obstructed environments.

[0065] S2.1 Adaptive Feature Extraction Based on Viewpoint Geometric Constraints The shipborne main control subsystem calculates the unmanned aerial vehicle (UAV) relative to the ship. The observation azimuth is used to determine the activation region for subsequent feature extraction algorithms. This is based on the ship's position in S1.1.4. With drone location (Projected onto the horizontal plane), define the observation azimuth vector. Calculate this vector and the ship's heading vector. relative azimuth : according to The observation state is divided into three intervals, and different feature extraction logics are activated for each interval.

[0066] Heading-on observation interval ( The main field of view covers the bow, activating Kelvin wave system feature extraction. Tail-chasing observation range ( The main field of view covers the stern, activating propeller wake feature extraction. Lateral observation range ( The main field of view covers the midships and sides of the ship.

[0067] S2.1.1 Kelvin Wave System Feature Extraction In bow-to-bow observation mode, the system uses the Canny edge detection operator to process images of the ship's bow region and identify the diverging wave crest line generated by the bow. In shallow port waters, the angle between the wave crest line and the heading is constrained by both water depth and ship speed. The system extracts the observed half-diffusion angle of the wave crest line relative to the ship's longitudinal axis. Simultaneously, the average pixel gradient energy within the wave crest region is calculated as the observed wave intensity. : in, The set of detected peak line pixels, This is the image grayscale gradient vector.

[0068] S2.1.2 Propeller wake feature extraction In tail-chase observation mode, the high-brightness turbulent region at the stern of the ship, i.e., the white foam band, is identified. The wake region is segmented using Lab color space transformation in S1.1.2. Calculate the observed wake length in the region in the opposite direction to the ship's heading. and the observed wake width perpendicular to the heading .

[0069] S2.2 Feature Correction Based on Background Flow Field In port navigation, the core of ship collision avoidance decision-making is not only predicting the collision risk of geometric trajectories, but more importantly, anticipating the hydrodynamic interactions during close encounters. When a ship encounters an oncoming vessel, the huge waves generated by the other vessel can cause the ship to roll violently or even lose control; when a ship is overtaking a vessel, if it gets too close to its wake, the propeller will suffer cavitation erosion due to the intake of a large number of air bubbles, resulting in a sharp drop in thrust.

[0070] The intensity of these hydrodynamic disturbances is not determined by the ship's speed relative to the ground, but by its relative speed to the surrounding water. A ship traveling against a strong current at 3 knots with a speed of only 5 knots relative to the ground has a relative current speed of up to 8 knots, generating far more wave-making and wake energy than a ship traveling with a current at 3 knots with the ground at the same speed of 5 knots (relative current speed of only 2 knots). If collision avoidance systems rely solely on ground speed or raw visual characteristics—that is, the appearance of small waves in an image—the danger posed by ships traveling against the current will be severely underestimated, leading to insufficient safety distance settings.

[0071] Therefore, background flow field data must be introduced to restore the visual appearance to its hydrodynamic essence. By correcting the obtained intrinsic parameters, namely the intrinsic half-divergence angle and the intrinsic wake impact length, a stable physical quantity characterizing the ship itself is obtained that does not change with the environmental flow field. In subsequent collision avoidance decisions, the system will use these intrinsic parameters, combined with the real-time background flow field at the intersection point of the ship and the target ship, to accurately predict the actual wave-making intensity and wake range that the target ship will generate in the upcoming intersection scenario, thereby planning a truly safe avoidance path and attitude adjustment commands.

[0072] Based on the reconstructed background flow field Obtain the position of the ship's center of mass. Background velocity vector at the location Calculate the relative velocity vector of the ship. : Based on this relative velocity, the features extracted from the aforementioned single viewpoint are physically corrected, and the observed wave spread angle is... It will be distorted with flow velocity. This can be achieved using the relative velocity modulus. Correct it and calculate the intrinsic half-diffusion angle. : Revised It can accurately reflect the wave-making pattern of ships in still water, which can be used to accurately calculate the wave propagation path.

[0073] Observation wake length It is the result of the combined effects of turbulent diffusion and background water transport. Based on a modified model of convection effects, the intrinsic wake impact length is calculated. It eliminates the visual errors of downstream stretching or upstream compression, restoring the true effective range of the propeller jet.

[0074] in, The turbulence intensity attenuation index; The flow coupling coefficient; This is the unit vector in the direction of the stern.

[0075] Based on the local features captured from a single viewpoint, and combined with the background flow field, a set of intrinsic hydrodynamic parameters was decoupled, which is independent of the current observation angle and eliminates environmental flow interference. .

[0076] Based on the acquisition of the intrinsic hydrodynamic parameter set, this embodiment will now elaborate on the specific implementation process of navigation decision generation.

[0077] S3.1 Construction and Spatiotemporal Evolution of Dynamic Hydrodynamic Disturbance Field First, utilize the decoupled intrinsic hydrodynamic parameter set For every ship sailing within the field of view A parameterized disturbance field model is constructed to characterize the ship's potential disturbance capability to the surrounding waters. This model does not depend on the current background current, but rather serves as the ship's inherent disturbance fingerprint.

[0078] The parameterization of the Kelvin wave interference field is based on the intrinsic half-diffusion angle. With ships Center of mass For the vertex, construct a path along its heading The symmetrically unfolding V-shaped region is defined as the potential wave-making influence zone. Wave energy within the region Wave energy distribution function decreases exponentially with distance for: In the formula, For any point within the affected area; The initial wave energy and the Froude number of the ship Positive correlation; is the wave energy spatial attenuation coefficient.

[0079] The parameterization of the propeller wake disturbance field is based on the intrinsic wake impact length. On the ship A rectangular region extending in the opposite direction of the heading is constructed at the tail end and defined as the potential wake influence zone. The turbulence intensity in this region It also decreases with distance.

[0080] To predict the actual interference situation at future moments, we first need to consider the planned route of this ship and the target ship. Current motion vector Predicting the future period Track the trajectories of the two ships and calculate the nearest meeting point. and meeting time .

[0081] At the moment of meeting Target ship The predicted location is Query the generated background flow field to obtain the background velocity vector at the meeting point. .

[0082] Subsequently, the actual water speed of the target ship at the meeting point was calculated. Using this actual water velocity, the parameterized disturbance field is instantiated in real time to obtain the predicted actual disturbance field: the predicted actual wave intensity. and Proportional. Predicted actual wake range By Substituting into the formula for calculating the intrinsic wake impact length yields the result.

[0083] S3.2 Quantitative Assessment of Disturbance Risk to the Ship After obtaining the predicted actual interference field at the moment of encounter, the predicted trajectory of the ship is spatiotemporally overlapped with the interference field to quantify the hydrodynamic risks that the ship will encounter.

[0084] At the meeting point The predicted location is The system calculates the position vector of the ship's center of mass relative to the target ship. .

[0085] If the ship's position Upon entering the predicted actual wave-making influence zone of the target vessel, the system calculates the lateral wave impact moment that the vessel will experience. : in, This is the ship's roll damping coefficient; To predict the gradient of the wave-generating energy field; This is the lateral unit vector of the ship. When When the ship's stability safety threshold is exceeded, a high risk of rolling is determined.

[0086] If the ship's position The system calculates the inflow bubble content of the propeller when the target ship falls into the predicted actual wake influence zone. This content is directly proportional to the turbulence intensity in the wake field. When When the critical value is exceeded, a propeller cavitation risk alarm will be triggered, indicating a risk of stall or sudden drop in thrust.

[0087] S3.3 Navigation Decision Generation and Path Optimization Based on the above risk quantification results, the system generates navigation recommendations that balance geometric collision avoidance and hydrodynamic safety. This invention is for the target vessel. Construct an asymmetric, V-shaped dynamic safety domain that varies dynamically with ship speed and flow field. The boundary of this safety zone not only takes into account the physical dimensions of the vessel, but also includes the predicted wave-making spread range and wake influence distance.

[0088] This invention adopts RRTThe collision avoidance path planning algorithm searches within the constraints of this dynamic safety domain, ensuring that the planned path not only maintains a safe distance from the target vessel's physical hull but also avoids its high-risk hydrodynamic interference core area. Simultaneously, it generates predictive attitude adjustment commands. When unavoidable passage through the wave-making edge area of ​​the target vessel is required, the necessary anti-roll rudder angle is calculated in advance. : in, This is the rudder effectiveness coefficient of the ship. This refers to the ship's speed relative to the water. For example, the system might suggest applying 5 degrees of left rudder within the next 30 seconds to actively counteract the impending wave-making impact on the starboard side and ensure course stability.

[0089] Ultimately, the shipboard main control subsystem will convert the generated navigation decisions into a structured sequence of navigation commands. The output is in the form of a sequence of instructions. This sequence of instructions contains not only discrete path points but also a set of continuous control functions with time-varying parameters, within a future time window. The specific form inside is: in, For the desired geographic location trajectory planned by the system, To achieve the desired water velocity, For the desired course, The calculated anti-disturbance pressure rudder angle.

[0090] Next, based on the previous decisions, this embodiment further performs the monitoring and parameter correction process.

[0091] S4.1 Real-time monitoring of navigation command execution status The shipboard main control subsystem outputs navigation command sequences. Subsequently, the ship's motion status is continuously monitored in real time through its integrated navigation system, and the deviation between the expected and actual status is calculated.

[0092] The actual position obtained by the ship's GNSS The expected position in the instruction sequence Compare the results and calculate the trajectory tracking residual vector. : The actual heading obtained With the desired course in the instruction sequence Compare and calculate the heading response residuals. : S4.2 Online parameter correction When the system continuously observes significant, directionally stable track tracking residuals over a relatively long period of time This indicates a discrepancy between the actual environmental forces experienced by the ship and the model predictions. The most direct and fundamental source of this discrepancy is the background flow field reconstructed based on discrete beacon interpolation. There is a difference between the flow field and the actual background flow field at the location of the ship.

[0093] The calibration process consists of a trigger condition and a calibration step. First, the calibration trigger condition is defined: it is only valid if the modulus of the track tracking residual is... Continuously exceeding a preset distance threshold And this state lasts for more than a time threshold. The system only activates the online calibration procedure at this time. This design effectively filters out high-frequency, short-term track jitter caused by gusts, waves, or transient responses of the control system, ensuring that the calibration targets persistent, systematic deviations caused by background flow field errors. Next, the flow field calibration steps are performed: once the triggering conditions are met, the system will perform the following calculations: Obtain the ground velocity vector : Directly read the ship's ground speed and ground heading. These two scalars together constitute the ship's actual ground speed vector. This vector represents the absolute motion of the ship relative to the Earth's fixed coordinate system.

[0094] Constructing the water velocity vector : Read the speed of this ship relative to the water Read the actual bow direction of this ship The bow direction is the angle between the keel and true north, representing the direction the ship points in the water. Based on this, the actual velocity vector relative to the water is constructed. : This vector represents the motion of the ship relative to the surrounding water.

[0095] Calculating the true background current velocity: According to the basic principle of vector composition, the ship's velocity relative to the ground is the vector sum of its velocity relative to the water and the current velocity. Therefore, the true background current velocity vector at the ship's location is... The result can be obtained directly through simple vector subtraction: Calculate model residuals and update the flow field: from the constructed background flow field map In the middle, check the current location of this ship. The model predicts the background flow velocity at that location. Calculate the residual velocity vector of the model. : Finally, the calculated true background flow rate As a new virtual velocity observation point with the highest confidence, the ground truth data is fused into the existing background flow field map using the same anisotropic spatial weighted interpolation algorithm as in S1.3. This allows for online updating and calibration of the local flow field.

[0096] When this vessel is in contact with the target vessel During close encounters, if the system observes and predicts the lateral wave impact moment... Simultaneous and dramatic heading response residuals That is, even if the ship anticipates the need to reduce the rudder angle... Afterwards, the course was still pushed off course, which proved that the target ship... The actual intensity of the hydrodynamic disturbance exceeded the predictions of previous models.

[0097] This invention attributes the heading response residual to the target ship. initial wave energy Or intrinsic wake impact length The intrinsic parameters are underestimated. Gradient descent is used to update these intrinsic parameters. The loss function is defined. To predict the squared difference between the yaw and the actual yaw, the parameter update rule is as follows: In the formula, The corrected initial wave energy; The learning rate. Gradient. The heading residual can be transferred back to the lateral moment model through the chain rule, and ultimately correlated with the initial wave energy.

[0098] When the system repeatedly observes similar prediction deviations when encountering similar vessels, the parameter update mechanism gradually adjusts the general intrinsic hydrodynamic parameter template for that vessel type. This allows the system to learn from case experience, forming an accurate profile of the disturbance capabilities of different vessel types, thereby making more precise risk predictions in future collision avoidance decisions.

[0099] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0100] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A ship collision avoidance method based on multi-source perception fusion, characterized in that, The method includes: By using drones to collect video images of ships and navigation marks in the waterway, the background water flow velocity in the waterway is calculated, and the static background water flow field covering the waterway is reconstructed. Based on video images, hydrodynamic features around the ship are extracted; the extracted hydrodynamic features are physically corrected using the static background water flow field to obtain the intrinsic hydrodynamic parameter set. Based on the intrinsic hydrodynamic parameter set and the real-time background flow field, the actual dynamic hydrodynamic disturbance field of the target ship is constructed and predicted, and a navigation decision containing the desired trajectory and predictive attitude adjustment commands is generated. Monitor the residual between the ship's actual motion state and the navigation decision; when the residual meets the preset trigger conditions, update the static background water flow field and make online corrections to the intrinsic hydrodynamic parameter set.

2. The ship collision avoidance method based on multi-source perception fusion according to claim 1, characterized in that, The steps for using drones to collect video images of ships and navigation marks within a waterway include: using ship detection boxes extracted based on a feature pyramid network, mapping the endpoints of the detection boxes from the image coordinate system to the world geographic coordinate system to obtain the ship... First geographical coordinates With tail geographic coordinates ; Calculate ships length And the Froude number, which characterizes a ship's wave-making ability. : ； ； in, To track the ground velocity vector obtained by the displacement of the ship's center of mass within a continuous time window, This is the acceleration due to gravity.

3. The ship collision avoidance method based on multi-source perception fusion according to claim 2, characterized in that, In the step of extracting hydrodynamic features around a ship based on video images, in order to eliminate navigation marks that are interfered with by the ship, the ship length is used. and Froude's number Determine the vessel Dynamic influence domain Its longitudinal length threshold With horizontal width threshold The calculation formula is: ； ； in, The preset longitudinal hydrodynamic attenuation constant, It is the transverse diffusion constant; For any navigation mark Calculate the position vector from the ship to the navigation mark. ,in The geographical coordinates of the navigation beacon. The coordinates of the ship's center of mass; when the position vector Falling into the dynamic influence domain If the time limit is reached, the navigation mark is determined to be a disturbed navigation mark; otherwise, it is determined to be an isolated navigation mark.

4. The ship collision avoidance method based on multi-source perception fusion according to claim 1, characterized in that, The steps for extracting hydrodynamic features around a ship based on video images also include: calculating the relationship between the UAV and the ship. The observation azimuth angle is used to extract visual hydrodynamic features in the head-on or tail-on observation interval. When the ship is in the bow observation range, extract the half-diffusion angle of the wave crests on both sides of the bow relative to the ship's longitudinal axis. And calculate the observed wave intensity : ； in, The set of detected peak line pixels, The number of pixels. The image grayscale gradient vector; When in the tail-chasing observation range, the wake region is segmented using color space transformation, and the observed wake length in the opposite direction of the ship's course is extracted. .

5. A ship collision avoidance method based on multi-source perception fusion according to claim 4, characterized in that, The steps for physically correcting the extracted visual hydrodynamic features to obtain the intrinsic hydrodynamic parameter set include: Obtaining ships from static background water flow field Background velocity vector at the centroid And calculate the relative velocity vector of the ship. ; Using relative velocity vectors to observe the half-diffusion angle After performing flow field correction, the intrinsic half-diffusion angle, which does not change with the flow field, is obtained. The calculation formula is as follows: 。 6. A ship collision avoidance method based on multi-source perception fusion according to claim 5, characterized in that, The step of physically correcting the extracted visual hydrodynamic features to obtain the intrinsic hydrodynamic parameter set also includes: Based on the convection effect correction model, the relative velocity vector is used to determine the observed wake length. By performing transport corrections to eliminate visual errors of downstream stretching or upstream compression, the intrinsic wake impact length is obtained. : ； in, The turbulence intensity attenuation index, The flow coupling coefficient is the coefficient of mass. This is a unit vector in the direction of the stern. Intrinsic half-diffusion angle and intrinsic wake impact length Together they constitute the intrinsic hydrodynamic parameter set.

7. A ship collision avoidance method based on multi-source perception fusion according to claim 3, characterized in that, The steps to calculate the background water velocity within the channel include: S1.2.1: Region of interest delineation for isolated navigation marks; S2.1.2: Feature point tracking and pixel displacement field generation; S2.1.3: Conversion of pixel displacement to world coordinate system velocity vector; S2.1.4: Flow velocity estimation based on RANSAC.

8. A ship collision avoidance method based on multi-source perception fusion according to claim 1, characterized in that, The steps for generating a navigation decision that includes the desired trajectory and predictive attitude adjustment commands include: calculating the lateral wave impact moment that the ship will experience when traversing the actual dynamic hydrodynamic disturbance field of the target vessel at the meeting point. ; Calculate the required predictive anti-disturbance pressure rudder angle As the attitude adjustment command for navigation decision-making, its calculation formula is: ； in, This is the rudder effectiveness coefficient of the ship. This is the ship's speed relative to the water.

9. A ship collision avoidance method based on multi-source perception fusion according to claim 1, characterized in that, When the residual meets the preset triggering conditions, the steps for updating the static background current field include: activating online parameter correction when the residual is a track tracking residual that continuously exceeds preset time and distance thresholds; and reading the ship's ground velocity vector. By combining the magnitude of the water velocity read by the ship's onboard log with the actual bow direction read by the compass, the actual water velocity vector of the ship is constructed. The true background current velocity vector at the ship's location is calculated using vector subtraction. The spatial weighted interpolation algorithm is used to integrate the real background velocity vector into the existing static background water flow field for online updating.

10. A ship collision avoidance system based on multi-source perception fusion, characterized in that, The system includes: Background water field reconstruction module: Uses drones to collect video images of ships and navigation marks in the waterway, calculates the background water flow velocity in the waterway, and reconstructs the static background water flow field covering the waterway. Parameter set reconstruction module: Based on video images, extract hydrodynamic features around the ship; use static background water flow field to physically correct the extracted hydrodynamic features to obtain the intrinsic hydrodynamic parameter set; The measurement generation module: Based on the intrinsic hydrodynamic parameter set and the real-time background flow field, it constructs and predicts the actual dynamic hydrodynamic disturbance field of the target ship, and generates a navigation decision containing the desired trajectory and predictive attitude adjustment commands. Online correction module: monitors the residual between the ship's actual motion state and the navigation decision; when the residual meets the preset trigger conditions, it updates the static background water flow field and corrects the intrinsic hydrodynamic parameter set online.

Images