A three-dimensional wave surface visual intelligent reconstruction and evaluation system

Abstract

This invention discloses a three-dimensional wave surface visual intelligent reconstruction and evaluation system, belonging to the field of marine engineering monitoring. The system includes an image acquisition module, an image stereo vision processing module, an attention-enhanced reconstruction neural network module, a camera attitude evaluation module, a visualization and output module, and various hydrodynamic parameter analysis modules. Images are acquired through a fixed-baseline binocular camera system. After preprocessing, a neural network incorporating multi-scale attention mechanisms is used to reconstruct the three-dimensional wave structure. Coordinate system transformation is achieved through self-supervised attitude evaluation, and finally, significant hydrodynamic parameters such as wave height and three-dimensional velocity field are extracted and visualized. This system requires no explicit calibration or large-scale data, possesses high automation, real-time performance, and strong environmental adaptability, and can be deployed on various platforms, significantly improving the accuracy and engineering applicability of non-contact wave observation.

Description

Technical Field

[0001] This invention belongs to the field of marine engineering monitoring, specifically relating to a three-dimensional wave surface visual intelligent reconstruction and evaluation system. Background Technology

[0002] With the continuous advancement of marine resource development and marine engineering construction, obtaining high-precision, timely three-dimensional wave structures and their hydrodynamic parameters (such as wave height, dominant frequency, propagation direction, and velocity field) has become an important issue in marine environmental monitoring. Current mainstream wave three-dimensional reconstruction and assessment technologies have significant shortcomings in practical applications, as detailed below:

[0003] 1. Traditional measurement methods: Contact sensors (such as buoys and pressure gauges) need to be deployed on the sea surface, which has limited spatial coverage, low resolution, and is easily damaged in harsh sea conditions; Remote sensing technology (such as satellite imagery and radar) has wide-area observation capabilities, but the update frequency and resolution are limited, making it difficult to achieve real-time reconstruction of small-scale, highly dynamic three-dimensional wave structures, especially with insufficient response to nonlinear characteristics such as breaking waves.

[0004] 2. Computer stereo vision methods: Traditional stereo vision relies on explicit calibration and feature matching, which is prone to failure when faced with drastic changes in wavefront and illumination interference; traditional stereo algorithms rely on single-point matching, which involves a large amount of computation and is difficult to achieve real-time results.

[0005] 3. Machine learning-based methods: Convolutional neural networks (CNN) and recurrent neural networks (RNN) are used to model wave fields, but they are highly dependent on training data and their accuracy decreases in small samples or new environments. In order to capture multi-scale dynamic features, deep networks are often constructed, resulting in lengthy models, low running efficiency, high deployment difficulty, and weak generalization ability.

[0006] In summary, existing technologies suffer from problems such as reliance on explicit calibration and complex modeling processes, poor adaptability to complex sea conditions, inability to recover three-dimensional velocity fields and dynamic spectrum features in real time, weak generalization ability with small samples, and large computational load and poor real-time performance, making it difficult to meet engineering requirements. Summary of the Invention

[0007] To address the aforementioned shortcomings, this invention provides a three-dimensional wave surface visual intelligent reconstruction and evaluation system to meet the needs of marine engineering, intelligent monitoring and environmental simulation fields for high precision, high timeliness and strong generalization ability.

[0008] The technical solution adopted in this invention is as follows: a three-dimensional wave surface visual intelligent reconstruction and evaluation system, comprising an image acquisition module, an image stereo vision processing module, an attention-enhanced reconstruction neural network module, a camera attitude evaluation module, a visualization and output module, and various hydrodynamic parameter analysis modules.

[0009] The image acquisition module is used to acquire visual information of the wave surface area. It uses a fixed baseline binocular camera to acquire image sequences of the target sea area in real time. It ensures spatiotemporal consistency through synchronous triggering and performs distortion correction and epipolar correction preprocessing. The acquired data includes grayscale image sequences from the left and right perspectives and their timestamps, and synchronously records the internal and external parameters of the camera system.

[0010] The image stereo vision processing module processes binocular images, generates dense disparity maps through SGBM stereo matching, calculates 3D point clouds by combining camera intrinsic parameters, performs depth restoration and 3D point cloud computing on image pixels, and forms an original point cloud dataset describing the geometric shape of ocean waves and liquid surfaces.

[0011] The attention-enhanced reconstruction neural network module: constructs a deep neural network that integrates multi-scale optical flow estimation and spatial-channel attention mechanism to model the three-dimensional geometric structure of waves. The encoder extracts multi-scale wave graphic features, the attention mechanism enhances the feature response of key areas of wave crests and troughs, the decoder outputs dense disparity map and reconstructed point cloud, and introduces implicit geometric learning mechanism to achieve high-fidelity three-dimensional reconstruction based on disparity without external parameters.

[0012] The camera attitude evaluation module: Based on planar constraints, a self-supervised attitude estimation network is constructed to fit the principal plane. Combined with affine transformation, the rotation and translation information of the camera are separated. The camera attitude is estimated by relying only on the geometric consistency of the wavefront, realizing the geometric transformation from the camera coordinate system to the liquid surface coordinate system.

[0013] The visualization and output module generates a three-dimensional visualization of the ocean wave surface based on the disparity map and camera attitude estimation results. It converts the reconstructed point cloud into a real space waveform map and combines color mapping and animation sequence to show the wave evolution process, and outputs the visualization of the reconstructed wave morphology.

[0014] The hydrodynamic parameter analysis module extracts key hydrodynamic parameters, including significant wave height, wavelength, wave velocity, dominant frequency, energy spectrum distribution, and three-dimensional velocity field, based on the reconstructed continuous three-dimensional wave surface data. The three-dimensional velocity vector field is obtained through continuous inter-frame surface displacement inversion, supporting a comprehensive evaluation of the dynamic characteristics of nonlinear, non-stationary, and multi-scale wave morphologies.

[0015] Furthermore, the method for generating a 3D point cloud by the image stereo vision processing module is as follows: performing SGBM stereo matching on the binocular images, assuming that the images captured by the left and right cameras are respectively... and Wave 3D Points The projection points in the image are respectively the left image point. With right image point Then the horizontal disparity of the left image is defined as:

[0016]

[0017] The disparity matrix is ​​solved using the SGBM algorithm for grayscale image matching.

[0018] Pixel coordinates and disparity values ​​can form the following homogeneous vector:

[0019]

[0020] via reprojection matrix Backproject the pixel into homogeneous coordinates of a 3D point:

[0021]

[0022] in, It is a homogeneous factor, a reprojection matrix. The format is:

[0023]

[0024] in, The focal length of the camera; The baseline lengths of the left and right cameras; and These are the x and y coordinates of the main point in the left figure; It is the x-coordinate of the main point in the right figure;

[0025] Finally, by dividing by the homogeneous factor Obtain the Euclidean three-dimensional coordinates:

[0026] .

[0027] Furthermore, the structure of the attention-enhanced reconstructed neural network module follows an "encode-match-decode" paradigm, including:

[0028] The encoding stage employs multi-scale feature extraction: the left and right images are normalized to unify the brightness scale, and then dense feature maps at scales of 1 / 8, 1 / 16, and 1 / 32 are extracted through a residual structure encoder. Meanwhile, the left image is used as a reference image and encoded separately to guide feature alignment during the matching process.

[0029] In the matching stage, a one-dimensional relevance pyramid is constructed: similarity is calculated by dot product or normalized inner product to form a cost volume along the scan line direction, and the matching similarity is normalized by the Softmax function to obtain the sub-pixel level disparity estimate of each pixel; a self-attention mechanism is introduced to model the feature map globally, and the weights are dynamically adjusted in a context-aware manner to suppress noise and fill disparity holes.

[0030] The decoding stage employs a multi-level upsampling fusion strategy to gradually restore the low-resolution disparity map to the scale of the original image. It fuses feature information at all levels and uses the residual enhancement module to refine the edge structure, ultimately outputting a floating-point disparity map of the same size as the original image.

[0031] Furthermore, the attention-enhanced reconstruction neural network module employs a weighted loss function combining L1 disparity loss and position awareness during training. The position awareness weighted loss function is as follows:

[0032] ,

[0033] in, The set of valid pixels; A weighted function for the vertical position of pixels; To predict disparity values; This represents the true disparity value.

[0034] 5. A three-dimensional wave surface visual intelligent reconstruction and evaluation system according to claim 1, characterized in that: the camera pose evaluation module estimates the camera pose by constructing a self-supervised pose estimation network based on planar constraints and combining it with the liquid surface plane constraints to perform principal plane fitting; wherein, the self-supervised loss based on point cloud geometric consistency constraints is defined as:

[0035]

[0036] in, The normal vector of the wave plane; It is the plane intercept;

[0037] The loss is calculated by summing the distances from each point in the point cloud to the fitted plane.

[0038] The equation for the wave horizontal plane in the camera coordinate system is expressed as:

[0039]

[0040] The transformation process uses the wave free surface normal as the reference, constructs a right-handed coordinate system through Gram-Schmidt orthogonalization, determines the translation vector based on the projection of the camera origin on the wave plane, and finally outputs the transformation matrix from the camera coordinate system to the wave coordinate system.

[0041] Furthermore, the visualization output module combines the reconstructed point cloud data with the camera pose information to generate a three-dimensional wave pattern in real space, and supports waveform animation and color mapping display.

[0042] Furthermore, the hydrodynamic parameter analysis module extracts a three-dimensional velocity vector field by analyzing the surface displacement between consecutive frames, and calculates the dynamic characteristics of wave energy spectrum, dominant frequency, and direction distribution by combining the Fourier transform method.

[0043] Furthermore, the hydrodynamic parameter analysis module uses a velocity field reconstruction model from the nonlinear shallow water wave theory to obtain the wave surface elevation field. Invert the three-dimensional velocity vector field, where the vertical velocity is reconstructed. With horizontal velocity component The calculation satisfies:

[0044]

[0045]

[0046]

[0047] Among them, vertical velocity The expression contains two key terms: one is the instantaneous rate of change of the wavefront. First, it reflects the rapid undulation behavior of the wave at that point; second, it is the tensor time integral term in the direction of the horizontal gradient of the wave surface. This expression is used to describe the fluid response caused by non-uniform propagation; the entire expression is... Weighting was applied to account for the correction of the vertical velocity by the wavefront slope; the horizontal velocity field... The calculation depends on The time integral of the tensor is then subtracted from the projection correction term of the vertical velocity onto the wavefront gradient. This reflects the local flow characteristics of water particles near the water surface; It is defined as a combination of gravitational potential energy and kinetic energy, with the first term being gravitational potential energy and the second term being the kinetic energy correction considering the wave surface velocity.

[0048] Using the theoretical model described above, numerical calculations were performed on the reconstructed time-series wavefront data. , , A three-dimensional velocity field for the time series was constructed, the spatial gradient was estimated using central difference, the time derivative was calculated using three-point backward difference, and the solution was obtained through time integration. The process ultimately reconstructs the complete three-dimensional velocity vector field for each frame. .

[0049] Advantages and benefits of this invention: This invention does not rely on external structured light or calibration plates, nor does it require large-scale manual annotation. It possesses high automation and practicality, and is particularly suitable for sea state areas where traditional buoys and underwater sensors are difficult to deploy. The system can be deployed on shore-based observation platforms, unmanned vessel-borne systems, or near-shore floating platforms, and has real-time acquisition and online evaluation capabilities. Combining multi-frame image input and GPU-accelerated inference mechanisms, the evaluation method can achieve real-time wave 3D structure and velocity field estimation at a second-level image processing speed, meeting the urgent needs of current marine engineering for rapid response and high spatial resolution wave field perception. This invention significantly improves the accuracy, efficiency, and engineering adaptability of non-contact wave observation. Attached Figure Description

[0050] Figure 1 This is a flowchart of the intelligent wave field reconstruction method provided in the embodiments of the present invention;

[0051] Figure 2 This is a schematic diagram of training labels provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of a three-dimensional neural network with attention enhancement provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the network output point cloud results provided in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the actual wave 3D reconstruction result provided by an embodiment of the present invention;

[0055] Figure 6 This is a waveform height parameter reconstruction error diagram provided in an embodiment of the present invention;

[0056] Figure 7 This is a time-series evaluation effect diagram of wave height parameters provided in an embodiment of the present invention;

[0057] Figure 8 This is a power spectrum evaluation effect diagram provided by an embodiment of the present invention;

[0058] Figure 9 This is a wave three-dimensional transient velocity evaluation effect diagram provided by the embodiment of the present invention, wherein (a) is a histogram of the angle proportion of the (Ux,Uy) vector, and (b) is a vertical velocity W accuracy comparison diagram. Detailed Implementation

[0059] This invention proposes a three-dimensional wave surface visual intelligent reconstruction and evaluation technology for real marine environments. The core of this technology lies in using deep neural networks to achieve accurate three-dimensional modeling of the wave field and automatic extraction of hydrodynamic parameters, significantly improving the accuracy, efficiency, and engineering adaptability of non-contact wave observation. To achieve these goals, the specific implementation combines disparity estimation and optical flow matching mechanisms from deep learning, designing and implementing a wave field reconstruction network architecture that integrates multi-scale attention mechanisms and implicit geometric modeling. This is supplemented by self-supervised camera attitude evaluation and point cloud conversion modules, ultimately constructing a complete three-dimensional wave surface reconstruction and hydrodynamic analysis system. Figure 1 The diagram illustrates the technical flow of each module of the present invention. The following will, in conjunction with the technical flow of the present invention, provide a detailed explanation of the overall system architecture, core module functions, and their specific implementations using examples.

[0060] Example 1

[0061] To acquire high-quality wave image data for model training and testing, a stereo image dataset oriented towards real-world, complex sea conditions was constructed during the implementation process. This dataset originated from binocular cameras actually deployed along the Black Sea coast, with acquisition spanning multiple day-night cycles, covering typical wave conditions from wind waves and small waves to more complex breaking waves. The binocular cameras were mounted on a fixed platform to ensure baseline stability and coverage of the wave propagation area in the nearshore waters. The acquired data includes grayscale image sequences from left and right perspectives along with their timestamps, and simultaneously recorded the camera system's intrinsic and extrinsic parameters to support subsequent epipolar correction and stereo matching. Each pair of images was initially processed using the SGBM algorithm to generate a dense disparity map, forming training samples with ground truth values, and a wave point cloud under stereo vision was generated as a subsequent comparison reference. Specifically:

[0062] Let the images captured by the left and right cameras be respectively and Wave 3D Points The projection points in the image are respectively the left image point. With right image point Then the horizontal disparity of the left image is defined as:

[0063]

[0064] The disparity matrix is ​​solved by matching grayscale images using the SGBM algorithm.

[0065] Pixel coordinates and disparity values ​​can form the following homogeneous vector:

[0066]

[0067] via reprojection matrix Backproject the pixel into homogeneous coordinates of a 3D point:

[0068]

[0069] in It is a homogeneous factor, a reprojection matrix. The format is:

[0070]

[0071] in, The focal length of the camera; The baseline lengths of the left and right cameras; and These are the x and y coordinates of the main point (usually the center of the image) in the left figure; It is the x-coordinate of the main point in the right figure.

[0072] Finally, by dividing by the homogeneous factor Obtain the Euclidean three-dimensional coordinates:

[0073]

[0074] like Figure 2 The image shows an example of training labels. Based on this, different datasets were constructed by combining various lighting and angle changes, and divided into training, validation, and testing subsets for network training and performance validation.

[0075] Compared to traditional methods that require cost calculation and global search in a two-dimensional image plane, this invention performs relevance matching only in the horizontal direction, significantly reducing computational complexity and improving training efficiency and convergence speed. This design limits the disparity between the left and right images to an offset along the scan line direction and sets the vertical optical flow component to zero, thereby achieving directional constraints on the matching task and significantly improving the stability and accuracy of disparity estimation in dynamic ocean wave scenes.

[0076] The input images are corrected binocular RGB image pairs, with a uniform size of 1600×1200 pixels, and pixel values ​​normalized to the range [0,1] to suit neural network training. The model training uses the Adam optimizer with an initial learning rate of 0.0004, employing a cosine annealing scheduling strategy. Training lasts for 200 epochs, with each epoch containing 3600 image pairs, and a batch size of 8. An early stopping mechanism is implemented when the validation set EPE (Endpoint Error) decreases by less than 1% for 15 consecutive epochs, terminating training.

[0077] During the encoding stage, the left and right images are respectively input into a residual feature extraction network with shared parameters. This module is built based on the ResNet structure and outputs feature maps at three scales, corresponding to 1 / 8, 1 / 16, and 1 / 32 resolutions, respectively. The left image is additionally used by a position encoding module to generate a guiding mask, the output dimension of which is aligned with the feature maps at each scale, and all are stored with float32 precision.

[0078] like Figure 3 As shown, the attention-enhanced reconstruction neural network module constructed in this invention follows an "encoding-matching-decoding" design paradigm. First, the input left and right images are normalized to unify the brightness scale and improve the model's robustness to different lighting conditions. Then, the images are fed into a feature extractor with a residual structure to extract dense feature maps at three scales: 1 / 8, 1 / 16, and 1 / 32. Simultaneously, the left image is used as a reference image and encoded separately to guide feature alignment during the matching process.

[0079] The network then constructs a one-dimensional relevance pyramid at various scales, calculates similarity using dot product or normalized inner product, forms a cost volume along the scan line direction, and normalizes the matching similarity using the Softmax function to obtain a sub-pixel level disparity estimate for each pixel. To enhance the matching stability of weak textures and occluded regions, a self-attention mechanism is introduced to globally model the feature map, dynamically adjusting weights through context awareness to suppress noise and fill disparity holes.

[0080] During the decoding stage, the network employs a multi-level upsampling fusion strategy to gradually restore the low-resolution disparity map to the original image size. Simultaneously, it fuses feature information from various scales and introduces a residual enhancement module to refine the edge structure, improving the clarity and continuity of the disparity map. The output disparity map is a floating-point tensor of the same size as the original image, possessing precise sub-pixel resolution capabilities, ensuring complete reconstruction of the three-dimensional wave structure even under complex sea conditions.

[0081] This invention not only employs the standard L1 disparity loss during training but also designs a position-aware weighted loss function to explicitly characterize the physical laws governing the error tolerance of different vertical positions in the image. In the vertical direction of the image, waves closer to the camera are more prone to errors due to their larger disparity. Therefore, higher tolerance is assigned to nearby areas in the error calculation, while stronger constraints are applied to distant waves, thereby improving the overall reconstruction stability and physical consistency.

[0082]

[0083] in, The set of valid pixels; It is a weighted function for the vertical position of pixels, reflecting the physical law of how the disparity value changes with the vertical coordinate.

[0084] Experimental results show that the average endpoint error (EPE) of the disparity estimation module on the test set is 1.12 pixels, and the average spatial error (MAE) of 3D point cloud reconstruction reaches millimeter-level accuracy in most areas.

[0085] The disparity tensor data output from the front-end network processing module is used as the primary intermediate data format and sent to the camera pose evaluation module through a unified data interface. Specifically, after image preprocessing, feature matching, and disparity calculation, a dense disparity map is generated and passed to the pose evaluation module via memory sharing or a high-speed caching mechanism. Based on the input disparity map and combined with camera intrinsic parameters, the pose evaluation module further reconstructs the 3D point cloud and estimates the camera's pose changes during observation, thereby providing spatial reference information for the subsequent reconstruction network and visualization modules.

[0086] Based on the reprojection formula in the solid geometric model, the disparity value and pixel position are jointly constructed into homogeneous coordinates, and mapped to a 3D point set in the camera coordinate system through the reprojection matrix. This point cloud data can intuitively reflect the morphology of surface fluctuations, but because it depends on the camera coordinates and lacks a unified marine reference frame, it is difficult to use for engineering analysis. Therefore, this invention further introduces a self-supervised attitude estimation network based on point cloud geometric consistency. Without manual calibration, it automatically fits the principal plane equation of the wave free surface and constructs the transformation matrix from the camera coordinate system to the wave coordinate system. The corresponding self-supervised loss based on point cloud geometric consistency constraints is defined as follows:

[0087]

[0088] in, The normal vector of the wave plane; It is the plane intercept;

[0089] Therefore, the equation for the wave surface in the camera coordinate system can be expressed as:

[0090]

[0091] The transformation process uses the wave free surface normal as a reference, constructs a right-handed coordinate system through Gram-Schmidt orthogonalization, and determines the translation vector based on the projection of the camera origin onto the wave plane. After the attitude transformation is completed, all 3D points are uniformly projected onto the wave coordinate system to construct a relatively stable and physically meaningful 3D wave morphology representation; the specific steps are as follows:

[0092] Select the origin of the camera coordinate system. Projecting onto the horizontal plane of the waves, determine the translation vector. :

[0093]

[0094] Defining the wave coordinate system Axis and plane normal vector Parallel, therefore, The unit vector of the axis is:

[0095]

[0096] Choose the initial vector and Non-parallel initial vectors The new coordinate system is determined using Gram-Schmidt orthogonalization. Axial unit vector:

[0097]

[0098] Then determine the new coordinate system according to the right-hand rule. axis:

[0099]

[0100] Final rotation matrix It consists of three orthogonal unit vectors:

[0101]

[0102] For any point in the original camera coordinate system Points transformed into wave coordinate system for:

[0103]

[0104] This completes the full transformation of 3D point cloud data from the camera coordinate system to the wave coordinate system, enabling effective extraction and analysis of wave surface information. It is worth noting that the positive directions of the coordinate axes can be adjusted according to actual needs by changing the rotation matrix. The sign of the elements in a vector allows for flexible control over the definition of each axis direction, adapting to different measurement scenarios or analysis objectives. For example... Figure 4 The image shown is a schematic diagram of the point cloud output of the neural network. Figure 5 The image shown is a 3D reconstruction result of the point cloud results mapped onto the real wave.

[0105] like Figure 6 As shown, the wave height parameter reconstruction error diagram provided by the embodiment of the present invention shows that the wave height program output by the neural network is highly consistent with the actual reference measurement data at most times.

[0106] Based on the completed 3D point cloud construction, the wavefront elevation sequence is further transformed into a 3D velocity vector field using a velocity field reconstruction model from nonlinear shallow water wave theory. The following formula provides the equation from the wavefront elevation field. Departure, Reconstruct Vertical Velocity With horizontal velocity component Theoretical approach:

[0107]

[0108]

[0109]

[0110] Among them, vertical velocity The expression contains two key terms: one is the instantaneous rate of change of the wavefront. First, it reflects the rapid undulation behavior of the wave at that point; second, it is the tensor time integral term in the direction of the horizontal gradient of the wave surface. This expression is used to describe the fluid response caused by non-uniform propagation. Weighting is applied to account for the correction of vertical velocity by wavefront slope. Horizontal velocity field The calculation depends on The time integral of the tensor is then subtracted from the projection correction term of the vertical velocity onto the wavefront gradient. This reflects the local flow characteristics of water particles near the water surface. It is defined as a combination of gravitational potential energy and kinetic energy, with the first term being gravitational potential energy and the second term being a kinetic energy correction considering the wave surface velocity.

[0111] Using the theoretical model described above, numerical calculations were performed on the reconstructed time-series wavefront data. , , Various parameters are used to construct a three-dimensional velocity field for the time series. The spatial gradient is estimated using central difference, the time derivative is calculated using three-point backward difference, and the solution is obtained through time integration. The process involves reconstructing the complete three-dimensional velocity vector field for each frame. .

[0112] The model first extracts the wavefront height change rate and horizontal gradient tensor based on the central difference method and time derivative calculation, then reconstructs the latent function through time integration, and finally calculates the horizontal and vertical velocity vectors by combining kinetic and potential energy components. This model strictly adheres to physical constraints and can accurately reconstruct the near-surface velocity structure of waves without introducing fluid particle simulations. It is suitable for applications such as fluid coupling, vortex detection, and wave-current interaction analysis.

[0113] like Figure 7The image shown is a graph illustrating the wave height parameter evaluation effect provided in an embodiment of the present invention; as shown... Figure 8 The image shown is a power spectrum evaluation effect diagram provided by an embodiment of the present invention. Experiments show that the system can achieve millimeter-level three-dimensional reconstruction accuracy under multiple typical sea states, with the frequency main peak error controlled within 0.01Hz. Figure 9 The image shown is a three-dimensional transient velocity evaluation effect diagram of waves provided by an embodiment of the present invention. Its velocity vector field is highly consistent with the measured reference data. The network relying on the attention mechanism can still adaptively fill the parallax missing area, avoiding the interference of reconstruction holes on the wave field modeling accuracy.

[0114] This invention does not rely on external structured light or calibration boards, nor does it require large-scale manual annotation. It possesses high automation and practicality, making it particularly suitable for sea state areas where traditional buoys and underwater sensors are difficult to deploy. The system can be deployed on shore-based observation platforms, unmanned vessel systems, or near-shore floating platforms, providing real-time data acquisition and online evaluation capabilities. Combining multi-frame image input with GPU-accelerated inference, this method achieves real-time wave 3D structure and velocity field estimation at second-level image processing speeds, meeting the urgent needs of current marine engineering for rapid response and high spatial resolution wave field perception.

Claims

1. A three-dimensional wave surface visual intelligent reconstruction and evaluation system, characterized in that, It includes an image acquisition module, an image stereo vision processing module, an attention-enhanced reconstruction neural network module, a camera attitude evaluation module, a visualization and output module, and various hydrodynamic parameter analysis modules. The image acquisition module is used to acquire visual information of the wave surface area. It uses a fixed baseline binocular camera to acquire image sequences of the target sea area in real time. It ensures spatiotemporal consistency through synchronous triggering and performs distortion correction and epipolar correction preprocessing. The acquired data includes grayscale image sequences from the left and right perspectives and their timestamps, and synchronously records the internal and external parameters of the camera system. The image stereo vision processing module processes binocular images, generates dense disparity maps through SGBM stereo matching, calculates 3D point clouds by combining camera intrinsic parameters, performs depth restoration and 3D point cloud computing on image pixels, and forms an original point cloud dataset describing the geometric shape of ocean waves and liquid surfaces. The attention-enhanced reconstruction neural network module: constructs a deep neural network that integrates multi-scale optical flow estimation and spatial-channel attention mechanism to model the three-dimensional geometric structure of waves. The encoder extracts multi-scale wave graphic features, the attention mechanism enhances the feature response of key areas of wave crests and troughs, the decoder outputs dense disparity map and reconstructed point cloud, and introduces implicit geometric learning mechanism to achieve high-fidelity three-dimensional reconstruction based on disparity without external parameters. The camera attitude evaluation module: Based on planar constraints, a self-supervised attitude estimation network is constructed to fit the principal plane. Combined with affine transformation, the rotation and translation information of the camera are separated. The camera attitude is estimated by relying only on the geometric consistency of the wavefront, realizing the geometric transformation from the camera coordinate system to the liquid surface coordinate system. The visualization and output module generates a three-dimensional visualization of the ocean wave surface based on the disparity map and camera attitude estimation results. It converts the reconstructed point cloud into a real space waveform map and combines color mapping and animation sequence to show the wave evolution process, and outputs the visualization of the reconstructed wave morphology. The hydrodynamic parameter analysis module extracts key hydrodynamic parameters, including significant wave height, wavelength, wave velocity, dominant frequency, energy spectrum distribution, and three-dimensional velocity vector field, based on the reconstructed continuous three-dimensional wave surface data. The three-dimensional velocity vector field is obtained through continuous inter-frame surface displacement inversion, supporting a comprehensive evaluation of the dynamic characteristics of nonlinear, non-stationary, and multi-scale wave morphologies.

2. The three-dimensional wave surface visual intelligent reconstruction and evaluation system according to claim 1, characterized in that, The method for generating a 3D point cloud by the image stereo vision processing module is as follows: SGBM stereo matching is performed on the binocular images, assuming the images captured by the left and right cameras are respectively... and Wave 3D Points The projection points in the image are respectively the left image point. With right image point Then the horizontal disparity of the left image is defined as: （1） The disparity matrix is ​​solved using the SGBM algorithm for grayscale image matching. Pixel coordinates and disparity values ​​can form the following homogeneous vector: (2) via reprojection matrix Backproject the pixel into homogeneous coordinates of a 3D point: （3） in, It is a homogeneous factor, a reprojection matrix. The format is: （4） in, The focal length of the camera; The baseline lengths of the left and right cameras; and These are the x and y coordinates of the main point in the left figure; It is the x-coordinate of the main point in the right figure; Finally, by dividing by the homogeneous factor Obtain the Euclidean three-dimensional coordinates: （5）。 3. The three-dimensional wave surface visual intelligent reconstruction and evaluation system according to claim 1, characterized in that, The structure of the attention-enhanced reconstructed neural network module follows the "encode-match-decode" paradigm, including: The encoding stage employs multi-scale feature extraction: the left and right images are normalized to unify the brightness scale, and then dense feature maps at scales of 1 / 8, 1 / 16, and 1 / 32 are extracted through a residual structure encoder. Meanwhile, the left image is used as a reference image and encoded separately to guide feature alignment during the matching process. In the matching stage, a one-dimensional relevance pyramid is constructed: similarity is calculated by dot product or normalized inner product to form a cost volume along the scan line direction, and the matching similarity is normalized by the Softmax function to obtain the sub-pixel level disparity estimate of each pixel; a self-attention mechanism is introduced to model the feature map globally, and the weights are dynamically adjusted in a context-aware manner to suppress noise and fill disparity holes. The decoding stage employs a multi-level upsampling fusion strategy to gradually restore the low-resolution disparity map to the scale of the original image. It fuses feature information at all levels and uses the residual enhancement module to refine the edge structure, ultimately outputting a floating-point disparity map of the same size as the original image.

4. The three-dimensional wave surface visual intelligent reconstruction and evaluation system according to claim 1, characterized in that: The visualization output module combines the reconstructed point cloud data with the camera pose information to generate a three-dimensional wave pattern in real space, and supports waveform animation and color mapping display.

5. The three-dimensional wave surface visual intelligent reconstruction and evaluation system according to claim 1, characterized in that: The hydrodynamic parameter analysis module extracts a three-dimensional velocity vector field by analyzing surface displacement between consecutive frames, and calculates the dynamic characteristics of wave energy spectrum, dominant frequency, and direction distribution by combining Fourier transform method.

Images