A ship near-field iceberg fusion recognition method and system

Abstract

The application provides a ship near-field iceberg fusion recognition method and system, relates to the technical field of image processing, and particularly relates to iceberg detection and risk assessment in a multi-ice region environment. The ship near-field iceberg fusion recognition method specifically comprises the following steps: acquiring multi-view image data of an overwater part of a target iceberg; based on the multi-view image data, reconstructing a three-dimensional surface model of the overwater part of the target iceberg and determining a waterline where the iceberg and the water body meet; analyzing image information on the three-dimensional surface model and identifying the spatial distribution of preset risk features; based on the spatial distribution of the risk features, generating a three-dimensional probability field of an underwater part of the target iceberg; acquiring current draft data of at least one ship; and combining the three-dimensional probability field and the current draft data to generate an underwater three-dimensional danger weight map, wherein the weight map at least identifies an underwater region that has a collision risk relative to the current draft.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for near-field iceberg fusion recognition of ships, particularly to iceberg detection and risk assessment in multi-ice zone environments, and to a method for generating a three-dimensional probability distribution map of underwater danger zones by analyzing image features of the surface portion of the iceberg. Background Technology

[0002] With the evolution of the global economic landscape, the strategic and economic value of high-latitude waterways is becoming increasingly prominent, attracting more and more frequent commercial transportation, resource development, and scientific research activities. However, the extreme environment faced by navigation in such icy waters poses a huge challenge to ship safety, among which the ubiquitous icebergs are the most significant and direct threat to navigation safety.

[0003] Unlike the mature shipping systems in temperate or tropical regions, waters with iceberg risks have two significant characteristics that make near-field iceberg avoidance an unavoidable and extremely challenging problem.

[0004] First, these high-latitude waterways, especially those leading to ports, loading / unloading points, or specific operational areas, often have inadequate navigational aids due to potential seasonal ice cover. These waterways are frequently narrow, winding, and lack advanced navigational aid systems. More importantly, these limited navigable spaces are often densely packed with icebergs and ice floes of varying sizes. This means that vessels operating in close proximity to icebergs cannot maintain the same safe distance as they would in the open ocean. In many cases, vessels are forced to navigate high-risk passages between icebergs in narrow channels, making close encounters with icebergs the norm rather than an isolated incident. In such forced close-proximity scenarios, the required level of precision in understanding the iceberg threat has been raised to an unprecedented level.

[0005] Secondly, the complexity of near-field operations in ice-covered areas is also reflected in the relativity and dynamism of risks. A typical scenario is that large cargo ships, when approaching a port or work site, are usually guided by one or more shallow-draft pilots or icebreakers. At this time, the same iceberg poses drastically different levels of danger to different vessels. A huge underwater ice keel hidden 5 meters below the surface may pose no threat to a pilot ship with a draft of only 3 meters, but to a mother ship with a draft of 15 meters, it is a deadly obstacle that can lead to catastrophic consequences.

[0006] Similarly, the risks to the same vessel are not static. A vessel's draft changes significantly between its fully loaded state before unloading and its unloaded state after unloading. This means that the danger posed by an iceberg to the vessel dynamically changes as operations progress. This stark reality, characterized by forced approach and relative risk, poses a challenge to existing iceberg detection and risk assessment technologies. Current technologies provide risk information that is largely two-dimensional and absolute, failing to meet the needs of three-dimensional, relative, and sophisticated decision-making.

[0007] Traditional ice forecasting methods, such as satellite remote sensing and ice maps published by international ice patrol teams (IIP), form the basis for ships' macro-level route planning. These methods can provide the distribution of ice zones over a wide area of ​​the sea and the macro-level location of large icebergs. However, their drawbacks include low spatial resolution and long time delays. Satellite imagery cannot distinguish between small and medium-sized icebergs or ice floes, and the data update cycle is long. For icebergs that are drifting rapidly in a specific sea area, the information is severely delayed, making it impossible to provide ships with near-field collision avoidance decision support. More importantly, these methods cannot provide any information about the underwater morphology of icebergs.

[0008] Even cutting-edge forward-looking sonar (FLS) has inherent limitations. FLS provides only a narrow, real-time two-dimensional slice of the underwater environment ahead. When pilot ships and mother ships need to coordinate operations, the pilot ship's sonar cannot provide effective safety assurance for the deeper-draft mother ship. The mother ship's own sonar can only provide a point-to-surface view, requiring operators to rely on highly specialized skills and experience to infer the complete three-dimensional underwater hazard pattern. This information presentation method is highly prone to misjudgment during high-pressure near-field operations. Current technology lacks a method to transform the detected information into an intuitive, holistic three-dimensional hazard model that can be correlated with the ship's specific draft.

[0009] Therefore, there is an urgent need in this field for a novel technical solution that generates a three-dimensional underwater hazard probability distribution map through in-depth analysis of readily available iceberg surface information. Such a weighted map would allow any vessel, whether a pilot ship or a massive vessel, to determine its own safe water layer and danger zone based on its real-time draft, thereby achieving truly refined and personalized risk avoidance and revolutionizing the safety of navigation in ice-covered areas. Summary of the Invention

[0010] This invention provides a method for near-field iceberg fusion identification of ships, which specifically includes the following steps:

[0011] Acquire multi-view image data of the above-water portion of the target iceberg;

[0012] Based on multi-view image data, a three-dimensional surface model of the above-water portion of the target iceberg is reconstructed, and the waterline at the boundary between the iceberg and the water body is determined.

[0013] Analyze image information on a 3D surface model to identify the spatial distribution of pre-defined risk features;

[0014] Based on the spatial distribution of risk characteristics, a three-dimensional probability field is generated for the underwater portion of the target iceberg.

[0015] Obtain the current draft depth data of at least one vessel;

[0016] By combining the three-dimensional probability field with the current draft depth data, an underwater three-dimensional hazard weight map is generated. The weight map at least identifies underwater areas with collision risk relative to the current draft depth.

[0017] The steps for obtaining multi-view image data of the above-water portion of a target iceberg include:

[0018] A multi-view image data S = { D_t | t ∈ [T_start, T_start + ΔT]} consisting of a data frame D_t is synchronously acquired at a preset frequency through a data acquisition system that integrates at least one visible light image sensor, one thermal infrared image sensor and one pose sensor. T_start and ΔT are the acquisition start time and the observation time window, respectively.

[0019] Each data frame D_t includes: a set of visible light images I_vis,t ={I_C1,t, I_C2,t, ..., I_Cn,t} acquired at time t, a thermal infrared image I_ir,t, and external parameters E_t, three-dimensional position vector T_t, and three-dimensional attitude rotation matrix R_t characterizing the sensor's six-degree-of-freedom spatial pose at time t; I_C1,t-I_Cn,t represent the images acquired by each visible light sensor at time t.

[0020] The reconstruction of the 3D surface model of the above-water portion of the target iceberg based on multi-view image data specifically includes: generating a set {D_m | m=1, ..., N} consisting of N independent dense depth maps. The generation process specifically includes: selecting multiple key moments from the multi-view image data, and specifying at least one camera viewpoint as a reference image at each key moment; for each reference image, determining an optimal depth value and generating the corresponding dense depth map D_m.

[0021] The steps for reconstructing the three-dimensional surface model of the above-water portion of the target iceberg also include:

[0022] The dense depth maps are projected onto a common coordinate system, and all generated 3D point clouds are filtered and fused based on multi-view geometric consistency verification to form a global 3D point cloud PC_final.

[0023] The surface normal vector is estimated based on the global 3D point cloud PC_final, and the isosurface is extracted to generate a 3D surface model M by solving the implicit surface function of the fitted normal vector field.

[0024] The specific steps for determining the waterline at the boundary between an iceberg and a body of water include:

[0025] Select the optimal thermal infrared reference image from the thermal infrared images contained in the multi-view image data;

[0026] Image segmentation is performed on the optimal thermal infrared reference image to distinguish between the first region representing the iceberg and the second region representing the water body, and a binary segmentation mask is generated.

[0027] Using the camera pose corresponding to the optimal thermal infrared reference image, the vertices of the 3D surface model are projected onto a binary segmentation mask, and the vertices are assigned a first region label or a second region label according to their projection position.

[0028] The waterline is ultimately determined by searching for edges connecting vertices with different region labels on the 3D surface model.

[0029] Analyzing image information on a 3D surface model to identify the spatial distribution of preset risk features includes: determining the waterline erosion risk distribution based on the geometric relationship between the 3D surface model and the waterline. This determination process includes: establishing a local coordinate system for each vertex on the waterline, and defining at least one horizontal normal vector pointing outward from the iceberg in each local coordinate system.

[0030] For each vertex p_q, a ray is projected from a predetermined height above it in the opposite direction to the horizontal normal vector, and the intersection distance λ_int,q between the ray and the three-dimensional surface model is calculated to quantify the horizontal indentation depth at the vertex. Based on the intersection distance λ_int,q, a corresponding erosion risk score S_u(p_q) is generated for each vertex p_q, resulting in a one-dimensional risk map distributed along the waterline.

[0031] Analyzing image information on a 3D surface model to identify the spatial distribution of preset risk features also includes: determining the structural crack risk distribution based on image texture information on the 3D surface model; creating a global texture map T_map for the 3D surface model based on multi-view image data; analyzing the global texture map T_map to identify two-dimensional crack regions, and projecting the two-dimensional crack regions onto the 3D surface model; reconstructing N_c 3D spatial crack paths through clustering and path fitting; and generating a corresponding structural risk score S_c(C_g) for each 3D spatial crack path C_g by normalizing and weighting its length, width, and minimum spatial distance from the waterline.

[0032] Based on the spatial distribution of risk characteristics, the generation of a three-dimensional probability field for the underwater portion of the target iceberg includes: calculating an initial probability value P_initial(v); defining a depth decay function f_d(z_v) that decays with increasing underwater depth, and applying this function to the initial probability value P_initial(v) to generate a three-dimensional probability field P_adjusted(v).

[0033] This specification also proposes a near-field iceberg fusion identification system for ships, which includes:

[0034] Acquisition module: Acquires multi-view image data of the above-water portion of the target iceberg; First return signal window period determination module: Predicts the first return signal window period based on the current weather image sequence.

[0035] Image analysis module: Based on multi-view image data, reconstruct a three-dimensional surface model of the above-water portion of the target iceberg and determine the waterline at the boundary between the iceberg and the water body;

[0036] Risk Feature Recognition Module: Analyzes image information on a 3D surface model to identify the spatial distribution of preset risk features;

[0037] Risk probability generation module: Based on the spatial distribution of risk characteristics, it generates a three-dimensional probability field of the underwater part of the target iceberg;

[0038] 3D Hazard Weight Map Generation Module: Combines the 3D probability field with the current draft depth data to generate an underwater 3D hazard weight map, in which the weight map at least identifies underwater areas with collision risk relative to the current draft depth.

[0039] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for near-field iceberg fusion identification of ships.

[0040] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for near-field iceberg fusion identification of ships.

[0041] Compared to existing technologies, this invention provides a method and system for near-field iceberg identification. Instead of directly measuring the invisible underwater portion, it establishes a three-dimensional surface model of the iceberg's above-water portion and proposes a method for evaluating this model. This involves probabilistically inferring the hazardous morphology of the underwater portion by quantitatively analyzing the geometric and textural features of the waterline and surface. The resulting hazard weight map is highly personalized and dynamic, addressing the pain point of relative risk in multi-ice zone navigation. Real-time, accurate draft data is combined with this parameter to generate a specific hazard zone for the ship at that particular moment. This means that a pilot ship with a draft of 3 meters and a mother ship with a draft of 15 meters will see their own and the other's restricted area maps on their respective navigation screens when facing the same iceberg, allowing the pilot ship to scientifically consider the risks to itself and the mother ship when planning its route. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of the near-field iceberg fusion identification of ships according to the present invention. Detailed Implementation

[0044] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0046] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0047] This specification presents an embodiment of a near-field iceberg fusion identification method for ships, which specifically includes the following steps:

[0048] Acquire multi-view image data of the above-water portion of the target iceberg;

[0049] Based on multi-view image data, a three-dimensional surface model of the above-water portion of the target iceberg is reconstructed, and the waterline at the boundary between the iceberg and the water body is determined.

[0050] Analyze image information on a 3D surface model to identify the spatial distribution of pre-defined risk features;

[0051] Based on the spatial distribution of risk characteristics, a three-dimensional probability field is generated for the underwater portion of the target iceberg.

[0052] Obtain the current draft depth data of at least one vessel;

[0053] By combining the three-dimensional probability field with the current draft depth data, an underwater three-dimensional hazard weight map is generated. The weight map at least identifies underwater areas with collision risk relative to the current draft depth.

[0054] The steps for obtaining multi-view image data of the above-water portion of a target iceberg include:

[0055] A multi-view image data S = { D_t | t ∈ [T_start, T_start + ΔT]} consisting of a data frame D_t is synchronously acquired at a preset frequency through a data acquisition system that integrates at least one visible light image sensor, one thermal infrared image sensor and one pose sensor. T_start and ΔT are the acquisition start time and the observation time window, respectively.

[0056] Each data frame D_t includes: a set of visible light images I_vis,t ={I_C1,t, I_C2,t, ..., I_Cn,t} acquired at time t, a thermal infrared image I_ir,t, and external parameters E_t, three-dimensional position vector T_t, and three-dimensional attitude rotation matrix R_t characterizing the sensor's six-degree-of-freedom spatial pose at time t; I_C1,t-I_Cn,t represent the images acquired by each visible light sensor at time t.

[0057] In this embodiment, the step of acquiring multi-view image data of the above-water portion of the target iceberg is completed through an integrated data acquisition system and process. This data acquisition system provides high-quality, multimodal, multi-view image data with high-precision pose information, which is spatiotemporally synchronized, for subsequent 3D reconstruction and waterline calibration.

[0058] In one specific implementation, the data acquisition system is constructed as a multi-view image acquisition assembly deployed on the ship's superstructure, preferably on the top of the bridge or the foremast. To address the challenges of low temperatures, high humidity, and hull rolling in icy environments, the assembly is mounted on an actively stabilized electronically controlled gimbal, and all core components are equipped with heated and highly protected sealed enclosures. The assembly integrates a multimodal camera array, a positioning and attitude determination unit, and a synchronization controller. The multimodal camera array includes not only two or more high-resolution industrial-grade color cameras for acquiring 3D structural and texture information, preferably one or more baseline-fixed binocular stereo camera groups to obtain rich depth information, but also at least one long-wave infrared (LWIR) thermal imaging camera for detecting the ice-water temperature boundary. The positioning and attitude determination unit is coupled to a high-frequency inertial measurement unit (IMU) and a Global Navigation Satellite System (GNSS) receiver supporting real-time dynamic differential (RTK) technology. The synchronization controller ensures that the data acquisition of all visible light cameras, thermal imaging cameras and positioning and attitude determination units in the multimodal camera array is strictly synchronized at the microsecond level through hardware triggering.

[0059] Before executing the data acquisition process, the system first performs a calibration. This calibration not only determines the intrinsic parameter matrix of each camera but also calculates the relative pose transformation relationships between all cameras (including between visible light cameras and thermal imaging cameras) and between the entire camera array and the IMU / GNSS. When approaching the target iceberg during navigation, the system continuously operates at a preset fixed frequency. At each acquisition time point t, the synchronization controller sends a global trigger signal, enabling the system to instantaneously capture and encapsulate a complete data frame D_t. The data frame D_t includes: a set of images captured by the multimodal camera array at time t, specifically a set of visible light images I_vis,t = {I_C1,t, I_C2,t, ...,I_Cn,t} and a thermal infrared image I_ir,t; and external parameters E_t, determined by the positioning and attitude determination unit at the same time t, characterizing the six-degree-of-freedom pose of the entire camera assembly in the world coordinate system. These external parameters include the three-dimensional position vector T_t determined by RTK-GNSS and the three-dimensional attitude rotation matrix R_t determined by IMU. Furthermore, the data frame also contains precise timestamp information.

[0060] As the ship sails, its position and viewing angle relative to the target iceberg constantly change. The data acquisition system operates continuously within a preset observation time window ΔT, acquiring multi-view image data S consisting of a series of time-sequentially arranged data frames D_t. This multi-view image data S = {D_t | t ∈ [T_start, T_start+ ΔT]} is the multi-view image data of the surface portion of the target iceberg used for subsequent processing. This sequence not only includes high-resolution visible light images of the iceberg from different viewing angles but also thermal infrared images that can distinguish the ice-water boundary. Furthermore, each frame of image data is strictly bound to its spatial pose information at the moment of capture.

[0061] Based on multi-view image data, the reconstruction of the 3D surface model of the above-water portion of the target iceberg specifically includes:

[0062] Generate a set {D_m | m=1, ..., N} consisting of N independent dense depth maps. The generation process specifically includes: selecting multiple key moments from multi-view image data, and specifying at least one camera viewpoint as a reference image at each key moment;

[0063] For each reference image, an optimal depth value is determined, and the corresponding dense depth map D_m is generated.

[0064] The steps for reconstructing the three-dimensional surface model of the above-water portion of the target iceberg also include:

[0065] The dense depth maps are projected onto a common coordinate system, and all generated 3D point clouds are filtered and fused based on multi-view geometric consistency verification to form a global 3D point cloud PC_final.

[0066] The surface normal vector is estimated based on the global 3D point cloud PC_final, and the isosurface is extracted to generate a 3D surface model M by solving the implicit surface function of the fitted normal vector field.

[0067] Specifically, multiple reference images are selected from the multi-view image data S to generate the depth map. This selection process consists of two steps: The first step is key moment selection. According to a preset strategy, N_t key moments {t_1, t_2, ..., t_Nt} are selected from the sequence S. The preset strategy can be at fixed time intervals. The second step is reference viewpoint selection. For each selected key moment t_j, one or more camera views are specified as reference views from the data frame D_tj at that moment. For example, for a stereo system with two cameras, at time t_j, the system can choose the image I_L,tj from the left camera as a reference image, and can also choose the image I_R,tj from the right camera as another independent reference image. Therefore, the total number N of the final dense depth map generated will be determined by the number of key moments N_t and the number of reference views selected at each moment.

[0068] For any selected reference image I_ref, which corresponds to a specific camera viewpoint k at a specific key moment t_j, when processing I_ref, the system selects N_s images as source images {I_src,i | i=1, ..., N_s} from its temporally and spatially neighboring data frames. Here, N_s is a preset parameter, for example, it can be set to 5, to balance reconstruction accuracy and computational efficiency.

[0069] For each pixel p = [u, v] in the reference image I_ref T The depth d(p) is determined through dense matching based on planar scanning. That is, within the frustum space of the reference camera, a series of discrete parallel planes Π_l are virtually set along the depth direction, each plane representing a depth hypothesis d_l. For each depth hypothesis d_l, the 3D coordinates P_ref(p, d_l) = d_l · K of the pixel p on that plane are calculated through the inverse projection of the camera intrinsic matrix K. - ¹ · [u, v, 1] T .

[0070] By utilizing the relative pose transformation between cameras, the 3D point P_ref is projected onto each I_src,i in N_s source images to find its corresponding pixel position.

[0071] The photometric consistency cost function C(p, d_l) between the neighborhood of pixel p in the reference image and the neighborhood of the corresponding projection point in all source images is calculated. Preferably, the cost function uses zero-mean normalized cross-correlation as the metric.

[0072] The depth assumption d_l that yields the optimal value of the cost function C is determined as the final depth value d*(p) of pixel p. This calculation is repeated for all pixels in the reference image I_ref to generate a dense depth map for that viewpoint, which we denote as D_ref.

[0073] Dense depth maps are computed for all selected reference images, resulting in a set of N independent dense depth maps, namely {D_m | m=1, ..., N}.

[0074] For each depth map D_m, an initial 3D point cloud PC_m is generated by projecting its corresponding absolute camera pose E_m in the world coordinate system. A geometric consistency check is performed on the set of all initial point clouds. A 3D spatial point is considered a valid high-confidence point only if the depth measurement differences are within a preset threshold by at least N_min different viewpoints. Finally, a single, dense, and accurate global 3D point cloud PC_final = {P_k | k =1, ...,M_p} is formed, where M_p is the total number of valid 3D points in the point cloud, and each P_k = (x_k, y_k, z_k) represents a 3D spatial point with defined world coordinates.

[0075] This embodiment performs surface reconstruction on the dense point cloud PC_final, preferably using Poisson surface reconstruction. Based on the local neighborhood information of each point P_k, its surface normal vector n_k is estimated. By solving a Poisson equation with the point cloud normal vector field as the gradient, a scalar field that best fits the entire point cloud manifold is calculated. The isosurface of this scalar field is extracted using the moving cube algorithm to generate a triangular mesh model M.

[0076] The specific steps for determining the waterline at the boundary between an iceberg and a body of water include:

[0077] Select the optimal thermal infrared reference image from the thermal infrared images contained in the multi-view image data;

[0078] Image segmentation is performed on the optimal thermal infrared reference image to distinguish between the first region representing the iceberg and the second region representing the water body, and a binary segmentation mask is generated.

[0079] Using the camera pose corresponding to the optimal thermal infrared reference image, the vertices of the 3D surface model are projected onto a binary segmentation mask, and the vertices are assigned a first region label or a second region label according to their projection position.

[0080] The waterline is ultimately determined by searching for edges connecting vertices with different region labels on the 3D surface model.

[0081] In a specific implementation, the waterline determination process involves evaluating all thermal infrared images contained in the multi-view image data S to select an optimal reference image for the final waterline calculation.

[0082] The system iterates through each thermal infrared image I_ir,t in the sequence and scores it according to a preset optimization strategy. This optimization strategy includes multiple factors, such as image sharpness and camera angle at the time of capture. Through this evaluation process, the system ultimately determines the image with the highest score as the unique optimal thermal infrared reference image, denoted as I_ir,opt.

[0083] After determining the optimal thermal infrared reference image I_ir,opt, image segmentation is performed to distinguish the iceberg region from the water region. Since the physical temperature difference between ice and water manifests as a clear difference in pixel intensity in long-wave infrared images, a graph-cut-based segmentation algorithm is preferably used. This algorithm comprehensively considers pixel grayscale information and spatial neighborhood information, thereby producing a smooth and accurate segmentation result at the ice-water boundary. The image segmentation output is a two-dimensional binary segmentation mask B_opt, where pixels belonging to the iceberg region are assigned label 1, and pixels belonging to the water region are assigned label 0.

[0084] Using the camera pose E_opt, which is acquired synchronously with the optimal reference image I_ir,opt, and the pre-calibrated camera intrinsic parameters K, projection calculations are performed on each vertex v on the 3D surface model M.

[0085] Specifically, the world coordinates of each vertex v are transformed into the coordinate system of the reference camera through the inverse transformation of the camera pose E_opt. Then, through the projection transformation of the camera intrinsic parameter K, the corresponding pixel coordinates p_v on the 2D image I_ir,opt are calculated. The label value of the pixel coordinate p_v on the binary segmentation mask B_opt is queried and this label is assigned to the 3D vertex v, denoted as L(v). By traversing all vertices on the model M, this process accurately renders the 2D thermal infrared segmentation results onto the 3D geometric surface.

[0086] After all vertices of the 3D model M are labeled as either ice or water, the final waterline is determined by searching the model boundary.

[0087] Traverse every edge in the model's triangular mesh. If the two endpoints of an edge, v_1 ​​and v_2, have different labels (i.e., L(v_1) ≠ L(v_2), then that edge is considered part of the waterline. All edges identified as waterline segments are grouped together, and the resulting 3D polyline is determined as the 3D waterline W_L of the target iceberg.

[0088] Analyzing image information on a 3D surface model to identify the spatial distribution of preset risk features includes: determining the waterline erosion risk distribution based on the geometric relationship between the 3D surface model and the waterline. This determination process includes: establishing a local coordinate system for each vertex on the waterline, and defining at least one horizontal normal vector pointing outward from the iceberg in each local coordinate system.

[0089] For each vertex p_q, a ray is projected from a predetermined height above it in the opposite direction to the horizontal normal vector, and the intersection distance λ_int,q between the ray and the three-dimensional surface model is calculated to quantify the horizontal indentation depth at the vertex. Based on the intersection distance λ_int,q, a corresponding erosion risk score S_u(p_q) is generated for each vertex p_q, resulting in a one-dimensional risk map distributed along the waterline.

[0090] Specifically, a preliminary estimate of the underwater extent of the glacier is made: this process is based on a three-dimensional waterline W_L, which is a three-dimensional polyline composed of N_w ordered vertices, i.e., W_L = {p_q | q=1, ..., N_w}.

[0091] To perform standardized geometric measurements at each waterline point, a local coordinate system needs to be established for each vertex p_q on the waterline. This local coordinate system consists of three mutually orthogonal unit vectors (v_fwd,q, v_out,q, v_up,q). Here, v_up,q represents the global vertical upward direction, which is the positive Z-axis direction of the world coordinate system, and v_up,q = [0,0, 1]. Tv_fwd,q represents the tangent direction of the waterline at that point. The tangent direction is approximated by the positions of its neighboring points, i.e., v_fwd,q = normalize(p_q+1 - p_q-1), where normalize(·) represents the normalization operation. v_out,q represents the horizontal normal direction pointing from the inside of the iceberg to the outside, obtained by the cross product of the first two vectors using the right-hand rule: v_out,q = normalize(v_up,q × v_fwd,q). The local coordinate system defines a sampling plane with p_q as the origin, perpendicular to the tangent direction of the waterline.

[0092] The degree of indentation on the ice surface is measured using ray projection. For each vertex p_q on the waterline, a ray emission point p_s,q is defined at a preset sampling height h_s directly above it along the v_up,q direction. This sampling height h_s is a configurable parameter, which can be set from 1.0 to 3.0 meters depending on sea state or iceberg type. The three-dimensional coordinates of this emission point are calculated using the following formula:

[0093] p_s,q=p_q + h_s (0,0,1)

[0094] A virtual ray R_q is emitted from the ray emission point p_s,q in a horizontal direction (0, -v_out,q, 0) opposite to the direction of the iceberg's exterior, pointing towards the interior of the iceberg. The first intersection of this ray R_q with the 3D surface model M is calculated. If the ray intersects the model M, the test returns an intersection distance λ_int,q. This distance λ_int,q quantifies the depth of the horizontal indentation of the ice surface relative to the vertical line of the waterline at the sampling height h_s.

[0095] Based on the calculated intersection distance λ_int,q, each vertex p_q on the waterline is assigned a quantified erosion risk score Su(p_q). The erosion risk score can be defined as:

[0096] S_u(p_q)=max(0,λ_int,q)

[0097] The formula states that if the ice surface is convex or vertical (resulting in negative or zero λ_int,q), its risk of indentation is zero. Only when the surface is concave (λ_int,q>0) is the risk score positive, and the larger the value, the more severe the indentation and the higher the probability of an underwater ice foot. By repeating the above process of establishing a local coordinate system, ray projection, and score calculation for all vertices p_q along the entire waterline, a one-dimensional risk map {S_u(p_q) |q=1, ..., N_w} is generated.

[0098] Iceberg melting occurs in three zones: the surface (aerial zone) and the deep water (underwater zone). The surface zone is primarily influenced by air temperature and solar radiation. In some seasons in ice-rich areas, temperatures are low or even below freezing, and sunlight intensity is limited. Therefore, the melting and sublimation of the surface zone is very slow. The deep water zone is mainly affected by heat conduction from the deep sea. The water flow is relatively gentle, so the melting rate in the deep water zone is also relatively slow and uniform. The waterline (ice-water interface) is the area where erosion and melting are most intense and fastest, a convergence point of multiple destructive forces. The surface seawater is slightly warmer, and wave turbulence constantly brings in new, relatively warm seawater, resulting in a much higher heat exchange efficiency than the calm deep water. The continuous pounding, scouring, and impacting of waves mechanically grinds and breaks down the ice at the waterline. This dynamic action greatly accelerates the melting of the ice.

[0099] Over time, the rate of retreat of an iceberg at the waterline is far faster than that above (in the air) and below (in deep water). The upper portion melts slowly and remains largely unchanged; the lower portion also melts slowly and similarly retains its massive original shape. This differential erosion rate directly leads to the evolution of the iceberg's sidewall morphology, ultimately forming a vertical profile that is narrow in the middle and wide at both ends. This narrow middle section is the waterline depression we can observe from the outside. This differential erosion process means that when the waterline has already contracted inward by 1 meter, the massive, slowly melting ice below it, which was originally on the same vertical plane, may have only contracted inward by 10 centimeters. Therefore, this waterline depression that has contracted inward by 1 meter must correspond to an underwater platform below it, extending outward by about 90 centimeters (1 meter - 10 centimeters), which we cannot see. This platform is the underwater ice foot. The waterline depression is like the shadow cast by the underwater ice foot on the water's surface.

[0100] Directly measuring the geometry at the waterline is extremely difficult. The waterline region is dynamically changing, often obscured by waves and mist, and the visual boundary between wet ice and water is blurred. This invention does not attempt to directly measure this blurred boundary, but instead measures the outline of the ice wall at a fixed height directly above the waterline. This is because if the waterline is severely eroded, the ice wall directly above it will inevitably contract inward relative to the waterline point. Since icebergs have arbitrary shapes and waterlines are arbitrary curves, there is no globally unified inside-out direction. The first step of this invention is to establish an independent local coordinate system with that point as the origin for each vertex p_q to be measured on the waterline. The calculation of the horizontal normal vector v_out,q provides an objective definition of the inside-out direction of that point. After defining the measurement direction, a virtual observation reference point p_s,q is determined at a height h_s directly above the waterline point p_q. Then, starting from this reference point p_s,q, a virtual measurement ray R_q is emitted in the opposite direction to the outside of the iceberg (-v_out,q). The distance between the intersection of this ray and the 3D model M represents the horizontal depth of the ice surface relative to the waterline at a height h_s. The larger this value, the more severe the indentation.

[0101] The final result of this invention is not a vague qualitative description, but a one-dimensional, continuous, quantitative risk map. This map clearly indicates which sections along the entire waterline are high-risk disaster areas and which sections are relatively safe.

[0102] Analyzing image information on a 3D surface model to identify the spatial distribution of preset risk features also includes: determining the structural crack risk distribution based on image texture information on the 3D surface model; creating a global texture map T_map for the 3D surface model based on multi-view image data; analyzing the global texture map T_map to identify two-dimensional crack regions, and projecting the two-dimensional crack regions onto the 3D surface model; reconstructing N_c 3D spatial crack paths through clustering and path fitting; and generating a corresponding structural risk score S_c(C_g) for each 3D spatial crack path C_g by normalizing and weighting its length, width, and minimum spatial distance from the waterline.

[0103] Specifically, a complete global texture map T_map is created for the 3D model M based on multi-view image data. The 3D model M is then UV unwrapped and tiled onto a 2D plane. Each 3D vertex v is assigned a unique 2D texture coordinate (x, y).

[0104] For any triangular facet f on the 3D model M, it may be visible in multiple images in the multi-view image data. The source image with the smallest angle between the camera viewpoint and the normal vector of the triangular facet f at the time of shooting is selected as the best source image for the triangular facet f.

[0105] After selecting the optimal source image for all patches, the corresponding image blocks are acquired and stitched together in a two-dimensional texture map space to obtain a global texture map T_map. Preferably, multi-band fusion or Poisson image editing is used to smooth the boundaries of adjacent image blocks.

[0106] After obtaining a high-resolution texture map T_map that represents the complete surface appearance of the iceberg, the system can analyze it to identify cracks. The subsequent process is consistent with the aforementioned scheme, but the logical foundation is now solid:

[0107] A pre-trained deep learning convolutional neural network model based on the U-Net architecture is used as input, and a binary crack mask B_c is output.

[0108] Based on the UV mapping relationship, all crack pixels with a value of 1 in the mask B_c are projected onto the surface of the 3D model M to form a 3D crack point set PC_crack_raw. The DBSCAN clustering algorithm is used to group the 3D crack point set to distinguish N_c independent cracks, and path fitting is performed on each point cloud cluster Clust_g (g=1, ..., N_c) to reconstruct an ordered 3D polyline C_g representing each crack.

[0109] For each reconstructed 3D crack C_g, calculate the structural risk score S_c(C_g):

[0110]

[0111] Among them, w_L, w_W, and w_D are weighting coefficients that sum to 1, representing the importance of the three factors of length, width, and waterline distance in the overall risk assessment.

[0112] The length risk factor F_L(C_g) characterizes the relative size of the crack. It is quantified by dividing the absolute length of the crack L_g by the maximum visible height H_ice of the iceberg's above-water portion.

[0113] The width risk factor F_W(C_g) characterizes the severity of the crack, which is quantified by dividing the average crack width W_g by a preset critical width W_crit.

[0114] The waterline distance risk factor F_D(C_g) characterizes the vulnerability of a crack, i.e., how close it is to the waterline of the most vulnerable area. An exponential decay function is used to model it to reflect the nonlinear relationship that the closer the distance, the more rapidly the risk increases.

[0115] D_g,min is the minimum spatial distance between the crack C_g and the three-dimensional waterline W_L, and ε is a small positive constant to prevent the denominator from being zero.

[0116] Based on the spatial distribution of risk characteristics, the generation of a three-dimensional probability field for the underwater portion of the target iceberg includes: calculating an initial probability value P_initial(v); defining a depth decay function f_d(z_v) that decays with increasing underwater depth, and applying this function to the initial probability value P_initial(v) to generate a three-dimensional probability field P_adjusted(v).

[0117] Specifically, a three-dimensional voxel space V is created underwater. For any voxel v in the space, its initial probability P_initial(v) is the sum of contributions from two risk sources: waterline erosion and structural crack risk.

[0118] The probability P_undercut(v) of waterline erosion risk is given by the fact that each vertex p_q on the waterline becomes a probability source based on its risk score S_u(p_q). For any voxel v in space, its probability contribution from the waterline erosion risk is the sum of the influences of all waterline points on it:

[0119] Here, dist() represents distance calculation. The closer a voxel is to a high-risk waterline depression, the higher the probability that it initially contains ice.

[0120] The probability P_cracks(v) of structural crack risk is given by each vertex v_crack on each three-dimensional crack C_g, which becomes the probability source based on the risk score S_c(C_g) of the crack to which it belongs.

[0121] ;

[0122] Where β_u and β_c are preset risk coefficients. and The spatial attenuation parameter used to control the radiation range of risk characteristics in three-dimensional space determines the voxel The rate at which the probability of risk decreases as one moves away from the source of risk (waterline depression or crack).

[0123] In this embodiment, and The value of is related to the resolution or physical scale of the 3D voxel mesh, and ranges from 0.5 to 3.0. Specifically, larger coefficients (such as...) This means that risk assessment is highly concentrated near the feature point, suitable for icebergs with extremely fragmented structures and drastic local changes; while smaller coefficients (such as...) This indicates that the risk has a longer spatial correlation. Even if there is a certain distance from the feature point, the system will still indicate the potential risk, thus leaving a larger safety margin in path planning.

[0124] The initial probability field P_initial(v) is the sum of these two contributions: P_initial(v) = P_undercut(v) + P_cracks(v).

[0125] A depth decay function f_d(z_v) is introduced, where z_v is the depth of voxel v relative to the waterline in the vertical direction. This function causes the probability to decrease with increasing depth. Preferably, a normalized exponential decay function is used:

[0126] Where γ is the depth attenuation coefficient, H_ice is the maximum visible height of the iceberg above water, and the attenuation function is applied to the initial probability field to obtain the depth-adjusted three-dimensional probability field: P_adjusted(v) = P_initial(v) · f_d(z_v);

[0127] Depth attenuation coefficient A dimensionless parameter used to adjust the conservatism of underwater volume estimation for icebergs. Different types of icebergs have different underwater / surface volume ratios. The value is set based on the statistical form of historical iceberg observation data in this sea area, with a preferred value range of 0.2 to 1.5. When smaller values ​​are used (e.g., 0.2 to 0.5), the decay is slower, simulating table-shaped icebergs or icebergs whose thickness does not change much with depth. The resulting probability field remains high in deep water, suitable for situations with extremely high risk aversion requirements; when... When larger values ​​are used (e.g., 0.8 to 1.5), the decay is faster, simulating the shape of a sharp-bottomed iceberg that contracts rapidly at the bottom. This can be improved by adjusting... The system can adapt to risk assessment needs under different sea conditions.

[0128] Obtain the current draft depth data of at least one vessel; combine the three-dimensional probability field with the current draft depth data to generate an underwater three-dimensional hazard weight map.

[0129] In one specific embodiment, data interface communication is established with the ship's load and draft monitoring system or integrated bridge system. These shipboard systems can read data in real time from draft sensors deployed around the hull and calculate the ship's average draft under its current condition, as well as any possible trim and list.

[0130] This system acquires at least one key draft parameter, d_ship. In scenarios involving multiple vessels working together, such as when a pilot vessel guides a mother vessel near a port, it simultaneously acquires the draft of the pilot vessel, d_pilot, and the draft of the mother vessel, d_main.

[0131] For a specific ship, the real danger lies not only in the presence of ice at some point underwater, but also in whether that ice is at a depth capable of colliding with the hull. Therefore, a risk correction function f_draft based on draft is introduced to adjust the original probability value of each voxel.

[0132] For any voxel v in voxel space, its depth relative to the waterline in the vertical direction is z_v. The risk correction function f_draft is defined as follows:

[0133]

[0134] Where: d_ship is the current real-time draft of the vessel. d_safety is a preset safety redundancy depth used to account for uncertainties such as vessel roll caused by waves and measurement errors.

[0135] `decay_func(z_v)` is a decay function whose risk correction value decreases rapidly when the voxel depth exceeds the absolute danger zone (i.e., draft + safety redundancy). Preferably, the decay function is a simple linear or exponential decay function, allowing its value to smoothly transition from 1 to 0.

[0136] The system calculates the final hazard weight W_final(v) for each voxel v for a specific ship:

[0137] The W_final(v) value incorporates the probability that ice exists at that location.

[0138] This specification also proposes a near-field iceberg fusion identification system for ships, which includes:

[0139] Acquisition module: Acquires multi-view image data of the above-water portion of the target iceberg; First return signal window period determination module: Predicts the first return signal window period based on the current weather image sequence.

[0140] Image analysis module: Based on multi-view image data, reconstruct a three-dimensional surface model of the above-water portion of the target iceberg and determine the waterline at the boundary between the iceberg and the water body;

[0141] Risk Feature Recognition Module: Analyzes image information on a 3D surface model to identify the spatial distribution of preset risk features;

[0142] Risk probability generation module: Based on the spatial distribution of risk characteristics, it generates a three-dimensional probability field of the underwater part of the target iceberg;

[0143] 3D Hazard Weight Map Generation Module: Combines the 3D probability field with the current draft depth data to generate an underwater 3D hazard weight map, in which the weight map at least identifies underwater areas with collision risk relative to the current draft depth.

[0144] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for near-field iceberg fusion identification of ships.

[0145] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for near-field iceberg fusion identification of ships.

[0146] 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. When executed, the computer program 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 RAMbus dynamic RAM (RDRAM), etc.

[0147] 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.

[0148] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the 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 method for near-field iceberg fusion identification of ships, characterized in that, The method includes: Acquire multi-view image data of the above-water portion of the target iceberg; Based on multi-view image data, a three-dimensional surface model of the above-water portion of the target iceberg is reconstructed, and the waterline at the boundary between the iceberg and the water body is determined. Analyze image information on a 3D surface model to identify the spatial distribution of pre-defined risk features; Based on the spatial distribution of risk characteristics, a three-dimensional probability field is generated for the underwater portion of the target iceberg. Obtain the current draft depth data of at least one vessel; By combining the three-dimensional probability field with the current draft depth data, an underwater three-dimensional hazard weight map is generated. The weight map at least identifies underwater areas with collision risk relative to the current draft depth. Based on the spatial distribution of risk characteristics, the generation of a three-dimensional probability field for the underwater portion of the target iceberg includes: creating a three-dimensional voxel space V underwater; for any voxel v in the space, its initial probability P_initial(v) is the superposition of contributions from two risk sources: waterline erosion and structural crack risk; the probability P_undercut(v) of waterline erosion risk is given by the sum of the contributions from all waterline points. Each vertex p_q on the waterline becomes a probability source based on its risk score S_u(p_q); for any voxel v in the space, its probability contribution from waterline erosion risk is the sum of the influence of all waterline points on it. ； Where dist() represents distance calculation; The probability P_cracks(v) of structural crack risk is given by each vertex v_crack on each three-dimensional crack C_g, which becomes the probability source based on the risk score S_c(C_g) of the crack to which it belongs. ； Where β_u and β_c are preset risk coefficients; N_w is the number of ordered vertices in the three-dimensional waterline W_L, and N_c is the number of independent cracks; P_initial(v) = P_undercut(v) + P_cracks(v); Define a depth decay function f_d(z_v) that decays with increasing underwater depth, where z_v is the depth of voxel v relative to the waterline in the vertical direction; Where γ is the depth attenuation rate coefficient, and H_ice is the maximum visible height of the iceberg above water. Applying the attenuation function to the initial probability field, we obtain the depth-adjusted three-dimensional probability field: P_adjusted(v) = P_initial(v) · f_d(z_v).

2. The method for near-field iceberg fusion identification of ships according to claim 1, characterized in that: The steps for obtaining multi-view image data of the above-water portion of a target iceberg include: A multi-view image data S = {D_t | t ∈ [T_start, T_start + ΔT]} consisting of a data frame D_t is synchronously acquired at a preset frequency through a data acquisition system that integrates at least one visible light image sensor, one thermal infrared image sensor and one pose sensor. T_start and ΔT are the acquisition start time and the observation time window, respectively. Each data frame D_t includes: a set of visible light images I_vis,t = {I_C1,t, I_C2,t, ..., I_Cn,t} acquired at time t, a thermal infrared image I_ir,t, and external parameters E_t, three-dimensional position vector T_t, and three-dimensional attitude rotation matrix R_t characterizing the sensor's six-degree-of-freedom spatial pose at time t; I_C1,t-I_Cn,t represent the images acquired by each visible light sensor at time t.

3. The method for near-field iceberg fusion identification of ships according to claim 2, characterized in that, The reconstruction of the 3D surface model of the above-water portion of the target iceberg based on multi-view image data specifically includes: generating a set {D_m | m=1, ..., N} consisting of N independent dense depth maps. The generation process specifically includes: selecting multiple key moments from the multi-view image data, and specifying at least one camera viewpoint as a reference image at each key moment; for each reference image, determining an optimal depth value and generating the corresponding dense depth map D_m.

4. The near-field iceberg fusion identification method for ships according to claim 3, characterized in that: The steps for reconstructing the three-dimensional surface model of the above-water portion of the target iceberg also include: The dense depth maps are projected onto a common coordinate system, and all generated 3D point clouds are filtered and fused based on multi-view geometric consistency verification to form a global 3D point cloud PC_final. The surface normal vector is estimated based on the global 3D point cloud PC_final, and the isosurface is extracted to generate a 3D surface model M by solving the implicit surface function of the fitted normal vector field.

5. The near-field iceberg fusion identification method for ships according to claim 4, characterized in that: The specific steps for determining the waterline at the boundary between an iceberg and a body of water include: Select the optimal thermal infrared reference image from the thermal infrared images contained in the multi-view image data; Image segmentation is performed on the optimal thermal infrared reference image to distinguish between the first region representing the iceberg and the second region representing the water body, and a binary segmentation mask is generated. Using the camera pose corresponding to the optimal thermal infrared reference image, the vertices of the 3D surface model are projected onto a binary segmentation mask, and the vertices are assigned a first region label or a second region label according to their projection position. The waterline is ultimately determined by searching for edges connecting vertices with different region labels on the 3D surface model.

6. The near-field iceberg fusion identification method for ships according to claim 5, characterized in that: Analyzing the image information on the three-dimensional surface model and identifying the spatial distribution of pre-wind hazard features includes: determining the waterline erosion risk distribution based on the geometric relationship between the three-dimensional surface model and the waterline; establishing a local coordinate system for each vertex on the waterline, with each local coordinate system defining at least one horizontal normal vector pointing outwards from the iceberg. For each vertex p_q, a ray is projected from a preset height above it in the opposite direction to the horizontal normal vector, and the intersection distance λ_int,q between the ray and the three-dimensional surface model is calculated to quantify the horizontal indentation depth at the vertex. Based on the intersection distance λ_int,q, a corresponding erosion risk score S_u(p_q) is generated for each vertex p_q to obtain a one-dimensional risk map distributed along the waterline.

7. The near-field iceberg fusion identification method for ships according to claim 6, characterized in that: Analyzing image information on a 3D surface model to identify the spatial distribution of preset risk features also includes: determining the structural crack risk distribution based on image texture information on the 3D surface model. This determination process includes: creating a global texture map T_map for the 3D surface model based on multi-view image data; analyzing the global texture map T_map to identify two-dimensional crack regions, projecting the two-dimensional crack regions onto the 3D surface model, and reconstructing N_c 3D spatial crack paths through clustering and path fitting; and generating a corresponding structural risk score S_c(C_g) for each 3D spatial crack path C_g by normalizing and weighting its length, width, and minimum spatial distance from the waterline.

8. A near-field iceberg fusion identification system for ships, the system being used to execute a near-field iceberg fusion identification method for ships as described in any one of claims 1-7, characterized in that, The system includes: Acquisition module: Acquires multi-view image data of the above-water portion of the target iceberg; First return signal window period determination module: Predicts the first return signal window period based on the current weather image sequence. Image analysis module: Based on multi-view image data, reconstruct a three-dimensional surface model of the above-water portion of the target iceberg and determine the waterline at the boundary between the iceberg and the water body; Risk Feature Recognition Module: Analyzes image information on a 3D surface model to identify the spatial distribution of preset risk features; Risk probability generation module: Based on the spatial distribution of risk characteristics, it generates a three-dimensional probability field of the underwater part of the target iceberg; 3D Hazard Weight Map Generation Module: Combines the 3D probability field with the current draft depth data to generate an underwater 3D hazard weight map, in which the weight map at least identifies underwater areas with collision risk relative to the current draft depth.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, implements a near-field iceberg fusion identification method for ships as described in any one of claims 1-7.

Images