Three-dimensional wave field measurement and analysis method and system based on multi-view camera

Abstract

The invention provides a three-dimensional wave field measurement analysis method and system based on a multi-view camera, and relates to the technical field of data measurement, and the method comprises the steps: collecting a wave field multi-angle image through the multi-view camera, building a space projection relation, obtaining a feature point set, and reconstructing an initial three-dimensional point cloud; extracting time-varying features by adopting a dynamic feature learning network; constructing a deformation constraint rule for correction according to a fluid mechanics equation; and finally calculating the three-dimensional characteristic parameters of the wave field. According to the invention, high-precision three-dimensional reconstruction of the wave field can be realized, the measurement accuracy is improved, and reliable technical support is provided for hydraulic engineering and marine environment monitoring.

Description

Technical Field

[0001] This invention relates to data measurement technology, and more particularly to a method and system for measuring and analyzing three-dimensional wave fields based on multi-view cameras. Background Technology

[0002] Three-dimensional measurement and analysis of ocean wave fields have significant application value in marine engineering, coastal zone management, ship design, and marine resource development. Accurate acquisition of three-dimensional wave field data is crucial for predicting wave behavior, assessing coastal erosion, optimizing marine structure design, and ensuring the safety of offshore operations. Traditional wave measurements primarily rely on point or line-based measurement equipment such as buoys and acoustic Doppler current profilers, which cannot obtain complete spatial distribution information of the wave field. With the development of computer vision and deep learning technologies, image-based wave field measurement methods have gradually emerged, providing a new technical approach for large-scale, non-contact wave field measurement.

[0003] Existing camera systems have limited coverage and cannot fully capture the spatial variation characteristics of large-scale wave fields. Especially under complex sea conditions, the multi-scale characteristics and irregular distribution of waves are difficult to record completely, resulting in incomplete spatial representation in the reconstruction results.

[0004] Existing wave field reconstruction methods mostly focus on static three-dimensional reconstruction, ignoring the temporal dynamic characteristics of the wave field. They cannot effectively extract and express the time-varying features of wave propagation and deformation, and the reconstruction results lack temporal consistency, making it difficult to accurately reflect the evolution process of the wave field.

[0005] Existing wave field reconstruction results often lack physical constraints and fail to fully consider the principles of fluid dynamics, resulting in deviations between the reconstructed three-dimensional data and the actual physical characteristics of waves. In particular, the accuracy is insufficient when calculating key parameters such as wave height, wavelength, and wave direction, which limits the reliability of the measurement results in engineering applications. Summary of the Invention

[0006] The present invention provides a method and system for measuring and analyzing three-dimensional wave fields based on multi-view cameras, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for measuring and analyzing three-dimensional wave fields based on a multi-view camera, comprising: Multi-angle image sequences of the wave field are acquired by a multi-view camera, the multi-view camera is calibrated, the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera are obtained, and the spatial projection relationship of the wave field surface points between different cameras is calculated based on the projection mapping network. Based on the spatial projection relationship, a viewpoint mapping association of feature points is established in the multi-angle image sequence to obtain the feature point set of the wave field. The spatial location distribution of the feature point set is calculated by the reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. A wave field dynamic feature learning network is used to analyze the temporal variation law of the initial three-dimensional point cloud. The time-varying features of the wave field are extracted through a temporal attention mechanism. The initial three-dimensional point cloud is then dynamically enhanced by an optimized network to generate a wave field three-dimensional point cloud with temporal consistency. Deformation constraint rules for the wave field are constructed based on fluid dynamics equations. The control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. The three-dimensional point cloud of the wave field is corrected based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to fluid dynamics characteristics. The geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field are analyzed using feature networks to calculate the three-dimensional feature parameters of the wave field.

[0008] The multi-view camera is calibrated, and its intrinsic and extrinsic parameter matrices are obtained. The spatial projection relationship of wave field surface points between different cameras is calculated based on a projection mapping network, including: The intrinsic parameter matrix and the extrinsic parameter matrix are input into the calibration parameter diffusion model. The calibration parameter diffusion model uses a Gaussian distribution to describe the temporal evolution of the parameters. A noise scheduling strategy is set based on the dynamic change characteristics of the wave field. The noise scale in the diffusion process is adjusted according to the noise scheduling strategy to generate the temporal probability distribution of the calibration parameters. Calculate parameter gradient information based on the time-series probability distribution, determine the sampling direction according to the parameter gradient information, perform directional sampling on the time-series probability distribution, and obtain multiple sets of calibration parameter samples; The spatial mapping relationship of wave field surface points between different cameras is calculated based on the multiple sets of calibration parameter samples. The spatial mapping relationship is then subjected to probability integration to obtain the projection transformation probability of wave field surface points. The posterior probabilities of the multiple sets of calibration parameter samples are calculated based on the correspondence of feature points on the wave field surface. The posterior probabilities are then input into the calibration parameter diffusion model for parameter optimization to obtain the optimal projection mapping parameters and establish the accurate projection relationship between wave field surface points and different cameras.

[0009] Based on the multiple sets of calibration parameter samples, the spatial mapping relationship of wave field surface points between different cameras is calculated. Probability integration is then performed on the spatial mapping relationship to obtain the projection transformation probability of the wave field surface points, including: Based on the multiple sets of calibration parameter samples, the projection transformation relationship of wave field surface points under a single camera is established, and the uncertainty distribution of wave field surface points during the camera view transformation process is analyzed according to the projection transformation relationship. Based on the uncertainty distribution, a mapping probability density function is constructed. The mapping probability density function uses the Dirac function to characterize the projection change law of the wave field surface point when it is transformed from the initial camera view to the target camera view. The Dirac function and the mapping probability density function are used to perform statistical analysis on the multiple sets of calibration parameter samples to obtain the initial projection probability of the wave field surface point. The initial projection probability and the mapping probability density function are integrated in the target camera imaging plane region. The inter-view mapping probability of the wave field surface points is obtained based on the result of the integration. Spatial consistency constraints between camera viewpoint pairs are constructed based on the inter-view mapping probability and the Dirac function. The spatial consistency constraint is minimized and optimized. Based on the result of the minimization optimization and the mapping probability density function, the optimal mapping probability that satisfies spatial consistency is obtained. The projection transformation probability between adjacent camera pairs is calculated based on the optimal mapping probability and the initial projection probability.

[0010] Based on the aforementioned spatial projection relationship, a viewpoint mapping association for feature points is established in the multi-angle image sequence to obtain a feature point set of the wave field. The spatial distribution of the feature point set is then calculated using a reconstruction network to reconstruct the initial three-dimensional point cloud of the wave field, including: A viewpoint mapping matrix is ​​constructed based on spatial projection relationship. Two-dimensional coordinate information and feature description information of wave field surface feature points are extracted from multi-angle image sequences. The viewpoint mapping matrix is ​​used to calculate the projection deviation value of feature points from the current viewpoint to adjacent viewpoints. Projection geometric constraints between feature points are established based on the projection deviation value. A matching score function is constructed, which integrates the projection geometric constraints and the feature description information. The matching score function calculates the correspondence score of feature points in adjacent viewpoints, and establishes the viewpoint mapping association of feature points based on the correspondence score. The feature point correspondences in the viewpoint mapping association are filtered, and the feature point correspondences whose matching degree scoring function values ​​are less than a preset scoring threshold are retained. A wave field feature point set is formed based on the retained feature point correspondences. The multi-view projection relationship of feature points in the wave field feature point set is analyzed by using a reconstruction network. The three-dimensional spatial position distribution of feature points is optimized by iteratively minimizing the matching degree scoring function. The optimal three-dimensional coordinates of the feature points are determined based on the three-dimensional spatial position distribution. The initial three-dimensional point cloud of the wave field surface is reconstructed based on the optimal three-dimensional coordinates.

[0011] A wave field dynamic feature learning network is used to analyze the temporal variation pattern of the initial 3D point cloud. The time-varying features of the wave field are extracted through a temporal attention mechanism. The initial 3D point cloud is then dynamically enhanced using an optimized network to generate a wave field 3D point cloud with temporal consistency, including: A wave field dynamic feature learning network is used to perform temporal sampling on the initial three-dimensional point cloud, extract the point cloud sequence within a fixed time window, and construct a temporal feature vector from the point cloud sequence; A temporal attention mechanism is constructed, wherein the temporal attention mechanism uses forward and reverse dual channels to calculate the dynamic correlation of the temporal feature vector at adjacent time points. The forward channel analyzes the motion trend of the wave field from the past to the present, and the reverse channel analyzes the motion law of the wave field from the present to the past, and establishes an attention weight matrix. The attention weight matrix is ​​used to perform multi-scale adaptive enhancement on the temporal features of wave field surface points. The multi-scale adaptive enhancement extracts the temporal correlation features of wave field surface points in the local neighborhood and the global range respectively, thereby obtaining the time-varying features of the wave field. The time-varying features are input into the optimization network, and a consistency constraint based on the temporal gradient is constructed based on the optimization network. The consistency constraint based on the temporal gradient ensures the continuity of motion of the wave field surface points at adjacent time points. The feature offset of the wave field surface points is calculated through iterative optimization. Dynamic feature enhancement is performed on the initial three-dimensional point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points and generate a wave field three-dimensional point cloud with temporal consistency.

[0012] Dynamic feature enhancement is performed on the initial 3D point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points, generating a temporally consistent 3D point cloud of the wave field, including: A chaotic dynamic model of the wave field is constructed based on the characteristic offset. The chaotic dynamic model describes the evolution of the wave field state vector over time through a nonlinear dynamic function. The trajectory deviation is calculated in the phase space of the chaotic dynamic model based on the wave field state vector. The trajectory deviation is substituted into the logarithmic growth rate function to obtain the Lyapunov exponent sequence. The maximum Lyapunov exponent is extracted from the Lyapunov exponent sequence. Substituting the maximum Lyapunov exponent, the preset benchmark enhancement coefficient, and the decay factor into the exponential decay function, an adaptive weight factor for feature enhancement is obtained. The feature offset is then adjusted using the adaptive weight factor to obtain enhanced features. The enhanced features are superimposed on the initial three-dimensional point cloud to form an enhanced point cloud. The enhanced point cloud is then substituted into the nonlinear dynamic function for integration. The result of the integration operation represents the temporal evolution trajectory of the points on the wave field surface. Based on the temporal evolution trajectory, the enhanced point cloud is continuously reconstructed to generate a three-dimensional point cloud of the wave field with temporal consistency.

[0013] The control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. Based on the optimized deformation constraint rules, the three-dimensional point cloud of the wave field is corrected to generate high-precision three-dimensional wave field data that conforms to hydrodynamic characteristics, including: Deformation constraint rules for a three-dimensional point cloud of a wave field are constructed. The deformation constraint rules include three types of hydrodynamic constraint terms. Each type of hydrodynamic constraint term corresponds to a constraint weight coefficient. A constraint weight vector is constructed based on the constraint weight coefficient. The constraint weight vector is mapped to a position vector in the particle swarm optimization space, and an initial velocity vector of the particle swarm is randomly generated based on the position vector; the constraint weight vector is substituted into the deformation constraint rule to calculate the constraint residual, and the constraint residual and the spatial curvature of the three-dimensional point cloud of the wave field are used to construct the optimization objective function. The optimization objective function is used to evaluate the quality of each particle's position in the particle swarm. The historical best position is taken as the individual best position, and the best position among all particles is taken as the global best position. The velocity component is calculated based on the individual best position and the global best position. The velocity components are superimposed to form the velocity increment of the particle swarm. The position vector of the particle swarm is iteratively updated using the velocity increment until the optimal constraint weight vector is obtained. The optimal constraint weight vector is substituted into the deformation constraint rule to correct the three-dimensional point cloud of the wave field, thereby obtaining a high-precision three-dimensional point cloud of the wave field that satisfies the hydrodynamic characteristics.

[0014] A second aspect of the present invention provides a three-dimensional wave field measurement and analysis system based on a multi-view camera, comprising: The first unit is used to acquire multi-angle image sequences of the wave field through a multi-view camera, calibrate the multi-view camera, obtain the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera, and calculate the spatial projection relationship of the wave field surface points between different cameras based on the projection mapping network. The second unit is used to combine the spatial projection relationship to establish the viewpoint mapping association of feature points in the multi-angle image sequence, obtain the feature point set of the wave field, and calculate the spatial position distribution of the feature point set through reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. The third unit is used to analyze the temporal variation law of the initial three-dimensional point cloud using a wave field dynamic feature learning network, extract the time-varying features of the wave field through a temporal attention mechanism, and perform dynamic feature enhancement on the initial three-dimensional point cloud through an optimization network to generate a wave field three-dimensional point cloud with temporal consistency. The fourth unit is used to construct deformation constraint rules for the wave field based on the fluid dynamics equations, iteratively optimize the control parameters of the deformation constraint rules through the particle swarm optimization algorithm, and correct the three-dimensional point cloud of the wave field based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to the fluid dynamics characteristics. The fifth unit is used to analyze the geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field using feature networks, and to calculate the three-dimensional feature parameters of the wave field.

[0015] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0017] The beneficial effects of this application are as follows: This invention achieves high-precision three-dimensional reconstruction and dynamic feature extraction of wave fields in natural waters through a multi-camera acquisition system and a multi-scale wave field modeling method. It overcomes the technical difficulties of limited measurement accuracy of traditional monocular cameras and the difficulty in capturing transient changes in wave fields, and improves the spatial resolution and temporal continuity of wave field measurement.

[0018] This invention combines fluid dynamics constraint rules and particle swarm optimization algorithms to correct the physical properties of three-dimensional point clouds of wave fields, making the reconstruction results more consistent with fluid dynamics characteristics. This effectively reduces measurement noise and reconstruction errors, enhances the reliability and stability of three-dimensional wave field data, and provides more accurate wave field parameters for marine engineering.

[0019] This invention achieves accurate extraction and analysis of time-varying features of wave fields through a wave field dynamic feature learning network and a temporal attention mechanism. It can effectively capture the dynamic evolution process of wave fields and provide important technical support and data foundation for fields such as marine environmental monitoring, marine engineering structure design and marine resource development. Attached Figure Description

[0020] Figure 1This is a flowchart illustrating the three-dimensional wave field measurement and analysis method based on a multi-view camera according to an embodiment of the present invention. Figure 2 This is a flowchart of the wave field dynamic feature learning network processing according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the optimization of the deformation constraint rules of a three-dimensional point cloud in a wave field based on the particle swarm optimization algorithm, as described in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1 This is a flowchart illustrating the three-dimensional wave field measurement and analysis method based on a multi-view camera according to an embodiment of the present invention. Figure 1 As shown, the method includes: Multi-angle image sequences of the wave field are acquired by a multi-view camera, the multi-view camera is calibrated, the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera are obtained, and the spatial projection relationship of the wave field surface points between different cameras is calculated based on the projection mapping network. Based on the spatial projection relationship, a viewpoint mapping association of feature points is established in the multi-angle image sequence to obtain the feature point set of the wave field. The spatial location distribution of the feature point set is calculated by the reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. A wave field dynamic feature learning network is used to analyze the temporal variation law of the initial three-dimensional point cloud. The time-varying features of the wave field are extracted through a temporal attention mechanism. The initial three-dimensional point cloud is then dynamically enhanced by an optimized network to generate a wave field three-dimensional point cloud with temporal consistency. Deformation constraint rules for the wave field are constructed based on fluid dynamics equations. The control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. The three-dimensional point cloud of the wave field is corrected based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to fluid dynamics characteristics. The geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field are analyzed using feature networks to calculate the three-dimensional feature parameters of the wave field.

[0024] In one optional implementation, the multi-view camera is calibrated to obtain its intrinsic and extrinsic parameter matrices. The spatial projection relationship of wave field surface points between different cameras is calculated based on a projection mapping network, including: The intrinsic parameter matrix and the extrinsic parameter matrix are input into the calibration parameter diffusion model. The calibration parameter diffusion model uses a Gaussian distribution to describe the temporal evolution of the parameters. A noise scheduling strategy is set based on the dynamic change characteristics of the wave field. The noise scale in the diffusion process is adjusted according to the noise scheduling strategy to generate the temporal probability distribution of the calibration parameters. Calculate parameter gradient information based on the time-series probability distribution, determine the sampling direction according to the parameter gradient information, perform directional sampling on the time-series probability distribution, and obtain multiple sets of calibration parameter samples; The spatial mapping relationship of wave field surface points between different cameras is calculated based on the multiple sets of calibration parameter samples. The spatial mapping relationship is then subjected to probability integration to obtain the projection transformation probability of wave field surface points. The posterior probabilities of the multiple sets of calibration parameter samples are calculated based on the correspondence of feature points on the wave field surface. The posterior probabilities are then input into the calibration parameter diffusion model for parameter optimization to obtain the optimal projection mapping parameters and establish the accurate projection relationship between wave field surface points and different cameras.

[0025] The calibration of the multi-view cameras employs Zhang's calibration method, which involves placing a calibration board around the wave field and capturing images of the board from different angles. Corner points on the calibration board are extracted using a sub-pixel detection algorithm, achieving an extraction accuracy of 0.1 pixels. Taking a tri-view camera system as an example, the camera's intrinsic parameter matrix includes focal length, principal point, and distortion coefficients. The intrinsic parameter matrix for camera 1 is as follows: focal length fx = 1024.5 pixels, fy = 1024.3 pixels, principal point coordinates cx = 640.2 pixels, cy = 512.1 pixels, radial distortion coefficients k1 = -0.08, k2 = 0.002. The intrinsic parameters for cameras 2 and 3 have a similar structure, but their values ​​differ slightly.

[0026] The extrinsic parameter matrix describes the positional relationship between cameras, including rotation matrices and translation vectors. Camera 2's rotation relative to camera 1 is rx = 0.02 radians, ry = 0.35 radians, rz = 0.01 radians, and its translation vectors are tx = 0.5 meters, ty = 0.02 meters, and tz = 0.01 meters. Camera 3 has similar parameters relative to camera 1, but with different values.

[0027] The intrinsic and extrinsic parameter matrices obtained from calibration are input into the calibration parameter diffusion model, which uses a Gaussian distribution to describe the temporal evolution of the parameters. This model treats the calibration parameters as random variables evolving over time, with the initial time t=0 set as a pure state and the final time t=T set as a completely noisy state. During the temporal evolution, the model gradually adds noise to the parameters according to a Gaussian process, with the noise level controlled by the variance parameter.

[0028] Considering the dynamic characteristics of the wave field, a special noise scheduling strategy is required. The wave field surface changes rapidly; therefore, a smaller noise scale (0.001) is used in the early stage of the time series, a medium noise scale (0.01) is used in the middle stage, and a larger noise scale (0.1) is used in the later stage. This noise scheduling strategy can adapt to the nonlinear changing characteristics of the wave field surface. Through a time-series diffusion process, a time-series probability distribution of the calibration parameters is generated, containing 1000 time steps, each time step corresponding to a set of parameter distributions.

[0029] Based on the generated temporal probability distribution, parameter gradient information is calculated. For each time step t, the rate of change of the parameter between adjacent time steps t+1 and t-1 is calculated to obtain the trend of parameter change. The parameter gradient information is represented as a vector, with the vector direction indicating the direction of the fastest parameter change and the vector magnitude representing the rate of change. Taking the focal length fx of camera 1 as an example, the gradient at time t=500 is 0.05, indicating that the parameter has a slight upward trend at this time.

[0030] Based on the calculated gradient information, the sampling direction is determined, prioritizing sampling along directions with larger gradients to capture the main parameter variation patterns. A Markov chain Monte Carlo method is used to perform directional sampling of the temporal probability distribution, with 100 samples per set, each containing the intrinsic and extrinsic parameters of all cameras. The sampling results form multiple sets of calibration parameter samples, each representing a single camera parameter configuration.

[0031] Based on multiple sets of calibration parameter samples, the spatial mapping relationship of points on the wave field surface between different cameras is calculated. For a point P(X,Y,Z) on the wave field surface, it is projected onto the image plane of camera 1 using the camera projection model, resulting in point p1(u1,v1). Similarly, point P is projected onto the image planes of camera 2 and camera 3, resulting in points p2(u2,v2) and p3(u3,v3), respectively. For each set of calibration parameter samples, a different set of projection results is obtained.

[0032] Taking point P(1.5, 2.0, 0.2) as an example, using the first set of parameter samples, its projection in camera 1 is p1(512.3, 480.5), and its projection in camera 2 is p2(623.1, 490.8); using the second set of parameter samples, the projection results are slightly different. Statistical analysis is performed on the projection results of all samples to construct the projection transformation probability distribution of point P from camera 1 to camera 2. Integral operation is performed on this probability distribution to obtain the projection transformation probability of point P. This process is repeated to calculate the projection transformation probability between all camera pairs.

[0033] To verify the accuracy of the calibration parameters, feature points need to be extracted from the wave field surface. The Harris corner detection algorithm is used to extract corner points from images captured by different cameras, and significant feature points are retained after filtering. For feature point p1i in camera 1, corresponding points p2i and p3i in cameras 2 and 3 are found based on the established projection relationship. The Euclidean distance between the actual and predicted corresponding points is calculated; a smaller distance indicates higher parameter accuracy. The posterior probability of each calibration parameter sample is calculated based on the accuracy of the matching point pairs.

[0034] The posterior probability calculation considers two factors: feature point matching accuracy and prior parameter probability, with a weighting ratio of 7:3. The calculated posterior probability is input into the calibration parameter diffusion model for a back-diffusion process, and the optimal projection mapping parameters are obtained through iterative optimization. The feature point matching accuracy corresponding to the optimal parameters can reach within 2 pixels, establishing a precise projection relationship between wave field surface points across different cameras. Experimental verification shows that this method achieves a 3D reconstruction accuracy of 5 millimeters under wave height conditions ranging from 0.2 meters to 0.8 meters, meeting the accuracy requirements for dynamic wave field monitoring.

[0035] In one optional implementation, the spatial mapping relationship of wave field surface points between different cameras is calculated based on the multiple sets of calibration parameter samples, and the projection transformation probability of the wave field surface points is obtained by performing a probability integral operation on the spatial mapping relationship. Based on the multiple sets of calibration parameter samples, the projection transformation relationship of wave field surface points under a single camera is established, and the uncertainty distribution of wave field surface points during the camera view transformation process is analyzed according to the projection transformation relationship. Based on the uncertainty distribution, a mapping probability density function is constructed. The mapping probability density function uses the Dirac function to characterize the projection change law of the wave field surface point when it is transformed from the initial camera view to the target camera view. The Dirac function and the mapping probability density function are used to perform statistical analysis on the multiple sets of calibration parameter samples to obtain the initial projection probability of the wave field surface point. The initial projection probability and the mapping probability density function are integrated in the target camera imaging plane region. The inter-view mapping probability of the wave field surface points is obtained based on the result of the integration. Spatial consistency constraints between camera viewpoint pairs are constructed based on the inter-view mapping probability and the Dirac function. The spatial consistency constraint is minimized and optimized. Based on the result of the minimization optimization and the mapping probability density function, the optimal mapping probability that satisfies spatial consistency is obtained. The projection transformation probability between adjacent camera pairs is calculated based on the optimal mapping probability and the initial projection probability.

[0036] The projection transformation relationship of wave field surface points under a single camera is constructed based on multiple sets of calibration parameter samples. These samples are obtained by calibrating the calibration board multiple times at different positions and directions. Each set of calibration parameters includes the camera's intrinsic parameter matrix K and extrinsic parameter matrix [R|t]. The intrinsic parameter matrix K includes focal lengths fx and fy and principal point coordinates cx and cy; the extrinsic parameter matrix includes the rotation matrix R and the translation vector t. For a multi-view system consisting of N cameras, the calibration process obtains M sets of calibration parameter samples, each containing the intrinsic and extrinsic parameters of the N cameras. The value of M is typically between 20 and 30; the more samples, the more accurate the subsequent probability estimation. In practical applications, a 9×7 calibration board is used to obtain 25 sets of calibration parameter samples at different distances (0.5m to 2m) and angles (-30° to +30°).

[0037] The uncertainty distribution analysis of wave field surface points during camera viewpoint transformation is based on projection error statistics. For a wave field surface point P in three-dimensional space with coordinates (X,Y,Z), it is projected onto the imaging plane of the i-th camera using the j-th set of calibration parameter samples, resulting in two-dimensional pixel coordinates pij=(uij,vij). Ideally, the pixel coordinates obtained using different sets of calibration parameters should be completely consistent, but due to calibration errors and noise, the actual projection results differ.

[0038] The projection results {pi1,pi2,...piM} obtained from M sets of calibration parameters are statistically analyzed, and their mean μi = (μui,μvi) and covariance matrix Σi are calculated. The mean μi represents the projection position, and the covariance matrix Σi characterizes the uncertainty distribution of the projection position. The uncertainty distribution is usually represented as a two-dimensional Gaussian distribution on the imaging plane, and its principal axis direction and magnitude are determined by the eigenvectors and eigenvalues ​​of the covariance matrix. In practical applications, the largest eigenvalue of the covariance matrix is ​​usually less than 5 pixels squared, indicating that the standard deviation of the projection position is within 2.2 pixels.

[0039] The mapping probability density function is constructed using a two-dimensional Gaussian distribution model to describe the projection change of points on the wave field surface as they transition from the initial camera's viewpoint to the target camera's viewpoint. For the initial camera i and the target camera j, the mapping probability density function fij(p|P) represents the probability density of the pixel coordinate p projected from the 3D point P onto the target camera j. The Dirac function δij(p,P) is used to precisely locate the projection point, defined as having a value of 1 if and only if p is the exact projection position of P under camera j, otherwise it is 0.

[0040] Combining the uncertainty distribution, the mapping probability density function is represented as a two-dimensional Gaussian distribution. Its mean is given by the accurate projection position determined by the Dirac function, and the covariance matrix is ​​obtained from the aforementioned uncertainty analysis. For each set of calibration parameter samples k (k=1,2...M), the point p under the initial camera i is projected into 3D space to obtain P, and then projected onto the target camera j to obtain p'. The {p'1,p'2...p'M} obtained from M sets of samples are statistically analyzed, and the mean and covariance matrix are calculated to construct the mapping probability density function. In practical applications, for images with a resolution of 1920×1080, uniformly distributed 100×100 grid points are selected for the mapping probability density function calculation, forming a mapping probability lookup table to improve the efficiency of subsequent calculations.

[0041] For a pixel p in the initial camera i, the initial projection probability of its corresponding 3D point P in the target camera j is denoted as Pi0(p'). The calculation method is to perform statistical analysis on M sets of calibration parameter samples, and calculate the position distribution of point p projected onto the target camera j through each set of calibration parameters. Specifically, for each set of calibration parameters k, calculate the intersection point Pk of the 3D ray corresponding to point p and the calibration plate plane, and then project Pk onto the target camera j to obtain p'k.

[0042] We statistically analyze the M sets of samples {p'1, p'2, ..., p'M} and calculate the frequency of point p' appearing at each position on the imaging plane of the target camera j, forming the initial projection probability distribution Pi0(p'). In practical applications, to reduce computational load, feature points are selected in the image of the initial camera i for calculation. Feature points are usually selected from regions with large image gradients, such as the texture boundaries of wave fields. Approximately 500 to 1000 feature points are selected for each image.

[0043] The initial projection probability Pi0(p') and the mapping probability density function fij(p'|P) ​​are integrated over the imaging plane Ωj of the target camera j to obtain the interview mapping probability Pij(p') of point p from camera i to camera j. The integration region Ωj is the effective imaging region of the target camera j, typically the entire image plane. The integration process uses a discrete approximation method, dividing the imaging plane into a grid, calculating the probability value of each grid point, and summing them. The grid size is typically set to 4×4 pixels, forming 480×270 grid points on a 1920×1080 resolution image. The interview mapping probability Pij(p') reflects the uncertainty of point p's transformation between camera i and camera j, providing a basis for subsequent spatial consistency constraints.

[0044] Spatial consistency constraints are constructed based on the inter-view mapping probability and the Dirac function. For a multi-view system consisting of N cameras, spatial consistency should be satisfied among any three cameras i, j, and k. That is, the mapping result of point p from camera i through camera j to camera k should be consistent with the mapping result directly from camera i to camera k. The spatial consistency constraint function C is defined as the sum of squares of the differences between the two mapping path results. For each point p, its spatial consistency constraint value in all three-camera combinations is calculated to form a global spatial consistency metric. In practical applications, for a system with 5 cameras, there are 10 three-camera combinations that require spatial consistency constraints to be calculated.

[0045] Using the spatial consistency constraint function C as the optimization objective, the parameters (mean and covariance matrix) of the mapping probability density function are adjusted to minimize the value of C. The initial values ​​of the optimization process are given by the mapping probability density function parameters obtained from the aforementioned statistical analysis. The learning rate is set to 0.01, and the maximum number of iterations is 100. In each iteration, the gradient of C with respect to each parameter is calculated, and the parameters are updated in the opposite direction of the gradient. The optimization terminates when the change in the value of C over five consecutive iterations is less than a threshold (usually set to 0.001) or when the maximum number of iterations is reached. After optimization, the optimal mapping probability density function f*ij(p'|P) ​​that satisfies spatial consistency is obtained.

[0046] The projection transformation probability is calculated based on the optimal mapping probability and the initial projection probability. For an adjacent camera pair (i,j), the projection transformation probability Tij(p,p') of point p from camera i to camera j is defined as the conditional probability of the 3D point P corresponding to p under camera i projecting to p' under camera j. Combining the optimal mapping probability density function f*ij(p'|P) ​​and the initial projection probability Pi0(p'), the projection transformation probability Tij(p,p') is calculated using Bayes' theorem. In practical applications, to improve computational efficiency, a projection transformation probability lookup table is pre-calculated. The table has dimensions Ni×Nj, where Ni and Nj are the effective pixel counts of camera i and camera j, respectively (usually downsampled to 1 / 4 of the original image, i.e., 480×270).

[0047] In a practical application, a multi-view system consisting of five industrial cameras with a resolution of 1920×1080 was used to acquire 25 sets of calibration parameter samples. For feature points on the wave field surface, the average projection error between adjacent camera pairs was 1.2 pixels, and the maximum projection error was 3.5 pixels. Applying the projection transformation probability calculated by this method, the matching accuracy between adjacent camera pairs improved from 78% to 92%, especially in the edge region of the wave field, where the accuracy improvement was more significant (from 65% to 87%). Compared with traditional mapping methods based on the fundamental matrix, this method has stronger robustness in handling calibration uncertainties and can effectively support subsequent feature point view mapping association and 3D point cloud reconstruction.

[0048] In one optional implementation, combining the spatial projection relationship, a viewpoint mapping association of feature points is established in the multi-angle image sequence to obtain a feature point set of the wave field. The spatial location distribution of the feature point set is calculated by reconstruction network analysis, and the initial three-dimensional point cloud of the wave field is reconstructed, including: A viewpoint mapping matrix is ​​constructed based on spatial projection relationship. Two-dimensional coordinate information and feature description information of wave field surface feature points are extracted from multi-angle image sequences. The viewpoint mapping matrix is ​​used to calculate the projection deviation value of feature points from the current viewpoint to adjacent viewpoints. Projection geometric constraints between feature points are established based on the projection deviation value. A matching score function is constructed, which integrates the projection geometric constraints and the feature description information. The matching score function calculates the correspondence score of feature points in adjacent viewpoints, and establishes the viewpoint mapping association of feature points based on the correspondence score. The feature point correspondences in the viewpoint mapping association are filtered, and the feature point correspondences whose matching degree scoring function values ​​are less than a preset scoring threshold are retained. A wave field feature point set is formed based on the retained feature point correspondences. The multi-view projection relationship of feature points in the wave field feature point set is analyzed by using a reconstruction network. The three-dimensional spatial position distribution of feature points is optimized by iteratively minimizing the matching degree scoring function. The optimal three-dimensional coordinates of the feature points are determined based on the three-dimensional spatial position distribution. The initial three-dimensional point cloud of the wave field surface is reconstructed based on the optimal three-dimensional coordinates.

[0049] The viewpoint mapping matrix is ​​constructed based on the aforementioned intrinsic and extrinsic parameter matrices of the multi-camera system. For a multi-camera system consisting of N cameras, a viewpoint mapping matrix M is established between every two cameras. ij , representing the projection transformation relationship from the viewpoint of the i-th camera to the viewpoint of the j-th camera. ijThe matrix has a dimension of 3×4 and contains rotation and translation components, where the rotation component is determined by the rotation matrix R in the extrinsic parameter matrix. i and R j The calculation shows that the translation component is determined by the translation vector t in the extrinsic parameter matrix. i and t j The calculations show that for a multi-view system with 5 cameras, a total of 20 view mapping matrices need to be constructed. The accuracy of the view mapping matrices directly affects the accuracy of subsequent feature point correspondences; therefore, during camera calibration, the reprojection error is controlled within 0.5 pixels.

[0050] The SIFT (Scale-Invariant Feature Transform) algorithm was used to extract surface feature points from the wave field. For each image sequence acquired by the camera, keyframes were selected for feature point extraction, with a keyframe interval of 3 to 5 frames. In each keyframe image, the SIFT algorithm extracted feature points with scale and rotation invariance. Each feature point contained two-dimensional coordinate information (u, v) and a 128-dimensional feature descriptor vector D. The feature descriptor vector D contained the gradient information of the feature point's neighborhood, which was used for subsequent feature matching. For images with a resolution of 1920×1080, 800 to 1200 feature points were typically extracted. Regions with high feature point density were mainly concentrated in the textured areas of the wave field, such as the boundaries between wave crests and troughs.

[0051] The calculation of projection deviation is the basis for establishing projection geometric constraints. For feature point p under the i-th camera viewpoint... i The coordinates are (u i ,v i ), using the view mapping matrix M ij Projecting it onto the viewpoint of the j-th camera, we obtain the projection point p. i ', coordinates are (u i ',v i '). If p i The corresponding feature point p is in the j-th camera view. j The coordinates are (u j ,v j If the projection deviation value is p, then the projection deviation value is defined as p. i 'and p j The Euclidean distance between them. Ideally, the projection deviation of perfectly corresponding feature points should be zero, but due to factors such as camera calibration errors, feature extraction errors, and surface reflections from wave fields, the actual projection deviation is usually not zero. The smaller the projection deviation, the better the geometric consistency of the feature point correspondence. Projection geometric constraint C geom Defined as the square of the projection deviation value, it is used for subsequent matching score.

[0052] The matching score function S comprehensively considers the projection geometric constraints C.geom Similarity C to feature description information desc， Similarity C of feature description information desc Defined as two feature point descriptor vectors D i and D j The Euclidean distance between them. The matching score function S is C. geom and C desc The weighted sum: S = w1 × C geom +w2×C desc The weighting coefficients w1 and w2 control the importance of geometric constraints and descriptor similarity, respectively. Typically, w1 is set to 0.7 and w2 to 0.3. This configuration emphasizes geometric consistency and helps reduce false matches. The smaller the matching score function S value, the more reliable the correspondence between feature points.

[0053] The feature point correspondence is established using a bidirectional matching strategy. For each feature point p in the i-th camera viewpoint... i Find the feature point p with the smallest matching score function S value under the j-th camera view. j As candidate corresponding points. Meanwhile, for p j Find the feature point p with the smallest S value from the i-th camera viewpoint. i '. If p i 'With the original feature point p i If they are the same, then p is considered to be... i and p j This establishes a valid correspondence. This bidirectional matching strategy effectively reduces fuzzy matching and multiple matching problems. For a system with five camera views, the above matching process is performed on each pair of camera views to obtain a preliminary feature point view mapping association.

[0054] For the established feature point correspondence, a matching score threshold T is set. score T score The value should be between 10 and 15. Only retain scores with a matching score S value less than T. score The correspondence between feature points is established. Furthermore, a local consistency check mechanism is introduced: for each feature point p... i Within its spatial neighborhood (typically with a radius of 5% of the image width), select K nearest neighbor feature points (K=10), and check whether the correspondence between these neighborhood feature points matches p. i The correspondences are consistent in spatial distribution. The local consistency metric L is calculated as the proportion of feature points in the neighborhood that meet the consistency condition, and only the correspondences of feature points with an L value greater than 0.6 are retained. After screening, a wave field feature point set P is formed, containing M feature points, and each feature point has a correspondence in at least two camera views.

[0055] The reconstruction network employs a multi-view reconstruction model based on tensor decomposition. For each feature point in the wave field feature point set P, an observation matrix O is constructed, with matrix elements O ij This represents the j-th coordinate component (u or v) of the feature point from the i-th camera viewpoint. Missing values ​​in the observation matrix O correspond to cases where the feature point is not observed from certain viewpoints. The observation matrix O is decomposed using tensor decomposition techniques to obtain the initial 3D spatial position X of the feature point. init .

[0056] The initial value X of the three-dimensional spatial position of the feature point init Starting with the initial step, iterative optimization is performed by minimizing the reprojection error function E. The reprojection error function E is defined as the sum of the squared differences between the feature point's 3D coordinates projected onto each camera's viewpoint and its actual observed coordinates. The Levenberg-Marquardt algorithm is used for optimization, with an initial iteration step size of 0.01, dynamically adjusted during iterations, and a maximum of 100 iterations. During iteration, if the reprojection error changes by less than 0.01 pixels over five consecutive iterations, the iteration is terminated early. The optimal 3D coordinates X of the feature point are obtained after optimization. opt .

[0057] Reconstruction of the initial 3D point cloud of the wave field based on the optimal 3D coordinates X of the feature points opt， For dense reconstruction, the PMVS (Patch-based Multi-view Stereo) algorithm is used, with feature points as seed points, expanding outwards to generate a dense 3D point cloud. During the expansion process, for each candidate point, its normalized cross-correlation (NCC) value under multiple camera views is calculated, with an NCC threshold set to 0.7, retaining only points with NCC values ​​greater than the threshold. To improve reconstruction quality, surface smoothing constraints are introduced to ensure smooth changes in the normal vectors of adjacent points. The resulting initial 3D point cloud contains geometric information of the wave field surface, with a point cloud density of 50 to 100 points per square centimeter.

[0058] In a practical application, a multi-view system consisting of five industrial cameras with a resolution of 1920×1080 was used. The cameras were arranged in a hemispherical pattern, covering a wave field of 1.5m×1.5m. In a marine engineering pool experiment, under regular wave conditions with a wave height of 10cm and a period of 1.2s, approximately 900 feature points were extracted from each camera's viewpoint. After viewpoint mapping, association, and filtering, approximately 3500 effective feature points were formed, resulting in an initial reconstructed 3D point cloud containing approximately 150,000 spatial points. Compared to traditional stereo vision methods, this method improves reconstruction accuracy by approximately 35%, particularly in the wave crest region, where the height error is reduced from ±3mm to ±1.2mm, laying a solid foundation for subsequent dynamic feature analysis of the wave field.

[0059] In one optional implementation, a wave field dynamic feature learning network is used to analyze the temporal variation pattern of the initial three-dimensional point cloud, and the time-varying features of the wave field are extracted through a temporal attention mechanism. The initial three-dimensional point cloud is then dynamically enhanced using an optimized network to generate a wave field three-dimensional point cloud with temporal consistency, including: A wave field dynamic feature learning network is used to perform temporal sampling on the initial three-dimensional point cloud, extract the point cloud sequence within a fixed time window, and construct a temporal feature vector from the point cloud sequence; A temporal attention mechanism is constructed, wherein the temporal attention mechanism uses forward and reverse dual channels to calculate the dynamic correlation of the temporal feature vector at adjacent time points. The forward channel analyzes the motion trend of the wave field from the past to the present, and the reverse channel analyzes the motion law of the wave field from the present to the past, and establishes an attention weight matrix. The attention weight matrix is ​​used to perform multi-scale adaptive enhancement on the temporal features of wave field surface points. The multi-scale adaptive enhancement extracts the temporal correlation features of wave field surface points in the local neighborhood and the global range respectively, thereby obtaining the time-varying features of the wave field. The time-varying features are input into the optimization network, and a consistency constraint based on the temporal gradient is constructed based on the optimization network. The consistency constraint based on the temporal gradient ensures the continuity of motion of the wave field surface points at adjacent time points. The feature offset of the wave field surface points is calculated through iterative optimization. Dynamic feature enhancement is performed on the initial three-dimensional point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points and generate a wave field three-dimensional point cloud with temporal consistency.

[0060] like Figure 2 As shown, the method includes: When the wave field dynamic feature learning network performs temporal sampling of the initial 3D point cloud, a sliding window strategy is adopted. The window length T is set to 5 to 15 frames, and in this embodiment, it is set to 8 frames, with a window sliding step size of 2 frames. For the point cloud sequence {P1, P2, ..., P8} within each time window, each point cloud frame P t Contains N three-dimensional points {p1 t p2 t ,...,p n t}, where p i t Represents the three-dimensional coordinates (x, y) of the i-th point in the t-th frame. i t ,y i t ,z i t To construct the temporal feature vector, for each point p... it Extract the local geometric feature g i t , which includes curvature, normal vector and height information. Specifically, calculate the local curvature of point p i t using the K-nearest neighbor method with K = 10, and obtain the mean curvature and Gaussian curvature by fitting the local quadratic surface; calculate the normal vector by the principal component analysis method; the height information is the z coordinate value. The temporal feature vector F i t is composed of the local geometric feature g i t and the differential features of the adjacent frames before and after. The differential features capture the change information at adjacent times.

[0061] The temporal attention mechanism uses a forward and backward dual-channel structure to calculate the dynamic correlation of the temporal feature vector at adjacent times. The forward channel analyzes the movement trend of the wave field from the past to the present and constructs the forward correlation matrix A forward . For each pair of times (t1, t2) within the time window, where t1 < t2, calculate the correlation degree between the feature vectors F i t1 and F i t2 . The correlation degree is obtained by dividing the inner product of the feature vectors by the vector norm, and the value range is [-1, 1]. Similarly, the backward channel analyzes the movement law of the wave field from the present to the past and constructs the backward correlation matrix A backward . For each pair of times (t1, t2) within the time window, where t1 > t2, calculate the correlation degree between the feature vectors F i t1 and F i t2 . After the correlation matrices of the forward channel and the backward channel are normalized, they are weighted and fused by the weight factors α and β to form the attention weight matrix W attention . In practical applications, α = 0.6 and β = 0.4. This configuration can pay more attention to the forward movement trend of the wave field while considering the supplementary role of the backward information.

[0062] The multi-scale adaptive enhancement uses the attention weight matrix to enhance the temporal features of the surface points of the wave field. At the local neighborhood scale, for each point p i t select the M nearest points in terms of spatial distance (M = 15) to form the local neighborhood set N local (p i t ). Calculate the feature similarity between this point and each point in the neighborhood under the guidance of the attention weight matrix W attention . The higher the similarity, the greater the feature contribution of the neighborhood point to the central point. The local enhanced feature Flocali t The weighted average is obtained, with the weights proportional to the feature similarity. On a global scale, Q points (Q=100) are randomly selected from the entire point cloud to form a global sampling set N. global (P t ), calculate the temporal correlation between the center point and the global sampling points, and extract the global enhancement feature F. globali t Local and global features are separated by adaptive weight γ. i t To perform fusion, γ i t Based on dynamic adjustment of local density and curvature at points, regions with higher curvature and density are assigned higher weights to local features. The final time-varying feature F is obtained. enhancedi t It takes into account both the local details and global structural changes of the wave field.

[0063] The optimized network constructs consistency constraints for temporal gradients based on time-varying characteristics to ensure the continuity of motion of points on the wave field surface at adjacent time steps. The temporal gradient is defined as the rate of change of a point's position between adjacent time steps, i.e., (p... i t -p i (t-1) The temporal gradient consistency constraint requires that gradient changes between adjacent time points be smooth, and the constraint function J is given by J / Δt, where Δt is the time interval between adjacent frames. consistency The network is constructed by calculating the second-order difference of the temporal gradient. The optimized network employs a four-layer fully connected structure, with the number of input layer nodes equal to the dimension of the time-varying feature (typically 64), hidden layer nodes of 128, 256, and 128 respectively, and output layer nodes of 3, corresponding to the 3D coordinate offset of the point. Network training uses a batch size of 32, a learning rate of 0.001, and 50 training epochs. The feature offset Δp is calculated through iterative optimization. i t It reflects the optimal adjustment amount of the point in the time sequence.

[0064] Dynamic feature enhancement adjusts the initial 3D point cloud based on feature offsets. For each point p i t Its enhanced position is p i t enhanced =p i t +λ·Δp i tλ is a smoothing factor that controls the intensity of feature enhancement, with a value range of [0,1], and is set to 0.7 in this embodiment. To ensure the physical rationality of the wave field surface, physical constraints are introduced during the enhancement process, including volume conservation constraints and surface tension constraints. If the enhanced point violates the physical constraints, constraint projection is performed, projecting the point to the nearest position that satisfies the constraints.

[0065] In a practical application, initial 3D point cloud data of a wave field containing 5,000 spatial points and 10 consecutive temporal frames was processed. Before processing, the average point position deviation between adjacent frames was 3.7 mm, and the temporal consistency metric was 0.65. After applying the above method, the average point position deviation between adjacent frames decreased to 0.9 mm, and the temporal consistency metric improved to 0.92. The improvement in feature consistency was particularly significant in the wave peak region, with the standard deviation of the temporal variation of local curvature decreasing from 0.082 to 0.025, indicating that the enhanced point cloud has higher accuracy and stability in capturing wave dynamic features. Compared with the traditional Kalman filter smoothing method, the proposed dynamic feature learning network improves temporal consistency by 37% and preserves wave detail features by 42%.

[0066] Through the aforementioned wave field dynamic feature learning network and temporal attention mechanism, the temporal consistency enhancement of the three-dimensional point cloud of the wave field was successfully achieved, laying a solid foundation for subsequent constraint optimization and feature parameter calculation based on fluid dynamics.

[0067] In one optional implementation, dynamic feature enhancement is performed on the initial three-dimensional point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points and generate a wave field three-dimensional point cloud with temporal consistency, including: A chaotic dynamic model of the wave field is constructed based on the characteristic offset. The chaotic dynamic model describes the evolution of the wave field state vector over time through a nonlinear dynamic function. The trajectory deviation is calculated in the phase space of the chaotic dynamic model based on the wave field state vector. The trajectory deviation is substituted into the logarithmic growth rate function to obtain the Lyapunov exponent sequence. The maximum Lyapunov exponent is extracted from the Lyapunov exponent sequence. Substituting the maximum Lyapunov exponent, the preset benchmark enhancement coefficient, and the decay factor into the exponential decay function, an adaptive weight factor for feature enhancement is obtained. The feature offset is then adjusted using the adaptive weight factor to obtain enhanced features. The enhanced features are superimposed on the initial three-dimensional point cloud to form an enhanced point cloud. The enhanced point cloud is then substituted into the nonlinear dynamic function for integration. The result of the integration operation represents the temporal evolution trajectory of the points on the wave field surface. Based on the temporal evolution trajectory, the enhanced point cloud is continuously reconstructed to generate a three-dimensional point cloud of the wave field with temporal consistency.

[0068] A chaotic dynamic model of the wave field is constructed based on the characteristic offset. This model describes the evolution of the wave field state vector over time using a nonlinear dynamic function. Specifically, a Lorentz system is used as the chaotic dynamic model of the wave field, and the state vector contains the position, velocity, and acceleration information of points on the wave field surface. For example, for a point P(x,y,z) in the acquired initial 3D point cloud, its characteristic offset is ΔP(Δx,Δy,Δz). The state vector of this point is set as S(t), which contains the position, velocity, and acceleration information of the point. The evolution of the state vector S(t) over time is described by a nonlinear dynamic function F.

[0069] The trajectory deviation in the phase space of a chaotic dynamics model is calculated based on the wave field state vector. Specifically, two trajectories with a very close initial distance are selected in the phase space, and their distance changes over time are calculated. For example, initial state vectors S1(0) and S2(0) are selected, where S2(0) = S1(0) + δ0, and δ0 is a very small initial perturbation, typically 0.00001. The two state vectors are allowed to evolve separately in the chaotic dynamics model, and the Euclidean distance δ(t) between the two trajectories is calculated at fixed time steps Δt (usually 0.01 seconds). The distance sequence δ(0), δ(Δt), δ(2Δt)...δ(T) from t=0 to t=T (e.g., T=100 seconds) is recorded.

[0070] Substituting the trajectory deviation into the logarithmic growth rate function yields the Lyapunov exponential sequence. Specifically, for each time point t, λ(t) = ln(δ(t) / δ0) / t is calculated, resulting in the Lyapunov exponential sequence {λ(Δt), λ(2Δt)...λ(T)}. The maximum Lyapunov exponent λ is then extracted from this sequence. ax The maximum Lyapunov exponent is typically obtained by averaging the latter half of the exponential sequence (e.g., t > 50 seconds). In practical applications, the typical value of the maximum Lyapunov exponent in a wave field is between 0.5 and 2.0, indicating the degree of chaos in the system.

[0071] Substituting the maximum Lyapunov exponent, a preset baseline enhancement coefficient, and a decay factor into the exponential decay function yields the adaptive weighting factor for feature enhancement. The preset baseline enhancement coefficient α is typically set to 1.5, and the decay factor β is typically set to 0.8. The adaptive weighting factor w is applied through the exponential decay function w = α·exp(-β·λ). ax ) is calculated. When λ ax When λ is large, it indicates a high degree of chaos in the system, and feature enhancement should be more conservative; in this case, the value of w is small. ax When λ is small, the system exhibits low chaos, allowing for more significant feature enhancement; in this case, the value of w is relatively large. For example, when λ... axWhen the value is 0.8, the calculated adaptive weighting factor w is approximately 0.72.

[0072] Enhanced features are obtained by adjusting the feature offset using an adaptive weighting factor. For each point's feature offset ΔP(Δx, Δy, Δz), the enhanced feature E(E) is calculated. X E Y E r =w·ΔP(Δx,Δy,Δz). For example, when the feature offset of a point is ΔP(0.05,0.03,-0.04) and the adaptive weight factor w=0.72, the calculated enhanced feature is E(0.036,0.0216,-0.0288).

[0073] The enhanced features are superimposed onto the initial 3D point cloud to form the enhanced point cloud. For each point P(x,y,z) in the initial point cloud, the enhanced point P'(x',y',z') is calculated as P(x,y,z) + E(E). X E Y E r For example, if the original coordinates of a point are P(2.345,1.678,0.892) and the enhanced feature is E(0.036,0.0216,-0.0288), then the coordinates of the enhanced point are P'(2.381,1.6996,0.8632).

[0074] The enhanced point cloud is substituted into the nonlinear dynamic function for integration. The fourth-order Runge-Kutta method is used for numerical integration, with a time step of 0.01 seconds. The total integration time is the required wave field simulation duration, such as 10 seconds. The point cloud state at each time step is recorded during integration, forming the temporal evolution trajectory of the wave field surface points.

[0075] The enhanced point cloud is continuously reconstructed based on the temporal evolution trajectory. The point cloud at each time step is triangulated to construct the topological structure of the wave field surface. Then, the wave field state at any time point is calculated using linear interpolation. Finally, the Laplace smoothing algorithm is applied to optimize the reconstructed point cloud, eliminating noise and discontinuities. The smoothing weight is typically set to 0.3, and the number of iterations is 3. Through these processes, a temporally consistent 3D point cloud of the wave field is generated, achieving continuous dynamic reconstruction of the wave field surface in the time dimension.

[0076] In practical applications, for initial point cloud data of wave fields with a collection frequency of 30Hz, the above method can achieve high-frequency reconstruction results of 120Hz, which effectively improves the temporal resolution and continuity of wave field surface reconstruction, making the reconstructed wave field surface visually smoother and more natural, and its physical properties more consistent with the actual wave motion laws.

[0077] In one optional implementation, the control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. Based on the optimized deformation constraint rules, the three-dimensional point cloud of the wave field is corrected to generate high-precision three-dimensional wave field data that conforms to hydrodynamic characteristics, including: Deformation constraint rules for a three-dimensional point cloud of a wave field are constructed. The deformation constraint rules include three types of hydrodynamic constraint terms. Each type of hydrodynamic constraint term corresponds to a constraint weight coefficient. A constraint weight vector is constructed based on the constraint weight coefficient. The constraint weight vector is mapped to a position vector in the particle swarm optimization space, and an initial velocity vector of the particle swarm is randomly generated based on the position vector; the constraint weight vector is substituted into the deformation constraint rule to calculate the constraint residual, and the constraint residual and the spatial curvature of the three-dimensional point cloud of the wave field are used to construct the optimization objective function. The optimization objective function is used to evaluate the quality of each particle's position in the particle swarm. The historical best position is taken as the individual best position, and the best position among all particles is taken as the global best position. The velocity component is calculated based on the individual best position and the global best position. The velocity components are superimposed to form the velocity increment of the particle swarm. The position vector of the particle swarm is iteratively updated using the velocity increment until the optimal constraint weight vector is obtained. The optimal constraint weight vector is substituted into the deformation constraint rule to correct the three-dimensional point cloud of the wave field, thereby obtaining a high-precision three-dimensional point cloud of the wave field that satisfies the hydrodynamic characteristics.

[0078] like Figure 3 As shown, the method includes: The deformation constraint rules are constructed using three types of fluid dynamics constraints: continuity constraints, momentum conservation constraints, and surface tension constraints. Continuity constraints ensure the conservation of fluid mass at the wave field surface, momentum conservation constraints guarantee that fluid motion follows Newton's second law, and surface tension constraints maintain the accuracy of the fluid surface tension effect. Each type of constraint corresponds to a constraint weight coefficient, denoted as α, β, and γ, respectively. These three weight coefficients constitute the constraint weight vector W=[α,β,γ]. In practical applications, the initial value of α is set to 0.45, the initial value of β is set to 0.35, and the initial value of γ is set to 0.2. These initial values ​​are empirical values ​​obtained from multiple experimental statistics.

[0079] The mapping from the constraint weight vector to the particle swarm optimization space employs a linear transformation method. Assuming the particle swarm contains N particles, where N typically ranges from 20 to 50, and is 30 in this embodiment, each particle corresponds to a position vector Pi (i = 1, 2... N). The initial position vector is obtained through a linear mapping of the constraint weight vector, ensuring that the initial particle distribution is within the effective range of the constraint weights. For each particle i, the initial position vector Pi = [αi, βi, γi], where αi is randomly generated within the range [0.3, 0.6], βi is randomly generated within the range [0.2, 0.5], and γi is randomly generated within the range [0.1, 0.3].

[0080] The initial velocity vector Vi is generated using a uniform random distribution, with the velocity range controlled within the interval [-0.05, 0.05], i.e., Vi = [vα,i,vβ,i,vγ,i], where vα,i,vβ,i, and vγ,i are all randomly selected within the range [-0.05, 0.05]. A smaller initial velocity vector helps the algorithm perform a fine-grained search in the early stages.

[0081] The constraint residual calculation process involves substituting the constraint weight vector corresponding to each particle's position into the deformation constraint rule to evaluate the degree of constraint satisfaction. For a given 3D point cloud dataset D of a wave field, containing M spatial points {p1, p2... pM}, each point pj (j = 1, 2... M) has 3D coordinates (xj, yj, zj). For each particle i, the residual values ​​R1,i, R2,i, and R3,i corresponding to its position Pi are calculated. R1,i represents the continuity constraint residual, obtained by calculating the deviation between the fluid volume change rate and the expected value in a local region of the point cloud; R2,i represents the momentum conservation constraint residual, obtained by calculating the deviation between the local fluid momentum change and the external force; R3,i represents the surface tension constraint residual, obtained by calculating the inconsistency between surface curvature and surface tension effects.

[0082] The spatial curvature calculation employs a local fitting method. For each point pj in the three-dimensional point cloud of the wave field, its K nearest neighbors are selected (K is typically 8 to 12, and 10 is used in this embodiment). Based on these nearest neighbors, a local quadratic surface is fitted, and then the average curvature Cj and Gaussian curvature Gj of that point are calculated. The curvature characteristics reflect the geometric properties of the wave field surface and are an important indicator for evaluating the correction effect.

[0083] The objective function F is optimized by comprehensively considering both constraint residuals and spatial curvature: F = w1 × (R1,i + R2,i + R3,i) + w2 × E, where E represents the curvature energy term of the point cloud, which is calculated by weighting the average curvature and Gaussian curvature of all points; w1 and w2 are weighting factors that balance constraint residuals and curvature features. Experiments have shown that the best results are achieved when w1 = 0.7 and w2 = 0.3.

[0084] The evaluation of particle position quality is based on the objective function value F. The smaller the F value, the better the particle position. For each particle i, its historical best position Pi is recorded. best This refers to the position where the objective function value is minimum during the particle's historical iterations. The global optimal position is Pg. best This is the position where the objective function value is minimum among all particles.

[0085] The velocity component calculation uses the inertia weighting method, and the new velocity vector calculation formula is: Vi new =ω×Vi+c1×r1×(Pi best -Pi)+c2×r2×(Pg best -Pi), where ω is the inertia weight, usually set between 0.7 and 0.9, and 0.8 in this embodiment; c1 is the cognitive learning factor, which controls the tendency of the particle to move towards the individual optimal position, and is set to 1.5; c2 is the social learning factor, which controls the tendency of the particle to move towards the global optimal position, and is set to 1.5; r1 and r2 are random numbers in the interval [0,1], introducing randomness to avoid getting trapped in local optima.

[0086] Particle position updates use a position accumulation method: Pi new =Pi+Vi new The updated position needs to undergo boundary checks to ensure that the constraint weights are within the valid range: α∈[0.3,0.6], β∈[0.2,0.5], γ∈[0.1,0.3]. If a dimension exceeds the boundary, it is set as the boundary value, and the corresponding velocity component is set to zero.

[0087] The process terminates when the maximum number of iterations is reached (typically set to 100 to 200 times, 150 times in this embodiment) or the change in the global optimal position is less than a preset threshold of 0.001 for 10 consecutive iterations. After meeting the termination condition, the global optimal position Pg is set. best W, as the optimal constraint weight vector optimal =[α optimal ,β optimal ,γ optimal ].

[0088] The final wave field 3D point cloud correction process involves adjusting the optimal constraint weight vector W. optimal Substituting the deformation constraint rules, the point positions of the original 3D point cloud are adjusted. Specifically, for each point pj, correction vectors Δpj1, Δpj2, and Δpj3 for the three types of constraint terms are calculated, and the combined correction vector Δpj = α. optimal ×Δpj1+β optimal ×Δpj2+γ optimal ×Δpj3, then update point pj new =pj+Δpj.

[0089] In a practical application, the particle swarm optimization algorithm converged after 87 iterations to optimize a 3D point cloud dataset of a wave field containing 10,000 spatial points, yielding the optimal constraint weight vector W. optimal =[0.52,0.38,0.1]. This weight vector was applied to correct the 3D point cloud of the wave field. Before correction, the average constraint residual of the point cloud was 0.078, which decreased to 0.012 after correction. The average standard deviation of the curvature of the point cloud surface decreased from 0.065 to 0.022, indicating that the corrected point cloud better conforms to hydrodynamic characteristics and has a smoother, more continuous surface. Compared with the traditional manual parameter tuning method, the computational efficiency is improved by approximately 68%, and the accuracy is improved by approximately 43%.

[0090] The above-mentioned particle swarm optimization algorithm can effectively obtain the optimal constraint weight vector, realize high-precision correction of the three-dimensional point cloud of the wave field, and provide an accurate three-dimensional data foundation for subsequent wave field feature analysis and parameter calculation.

[0091] A second aspect of the present invention provides a three-dimensional wave field measurement and analysis system based on a multi-view camera, comprising: The first unit is used to acquire multi-angle image sequences of the wave field through a multi-view camera, calibrate the multi-view camera, obtain the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera, and calculate the spatial projection relationship of the wave field surface points between different cameras based on the projection mapping network. The second unit is used to combine the spatial projection relationship to establish the viewpoint mapping association of feature points in the multi-angle image sequence, obtain the feature point set of the wave field, and calculate the spatial position distribution of the feature point set through reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. The third unit is used to analyze the temporal variation law of the initial three-dimensional point cloud using a wave field dynamic feature learning network, extract the time-varying features of the wave field through a temporal attention mechanism, and perform dynamic feature enhancement on the initial three-dimensional point cloud through an optimization network to generate a wave field three-dimensional point cloud with temporal consistency. The fourth unit is used to construct deformation constraint rules for the wave field based on the fluid dynamics equations, iteratively optimize the control parameters of the deformation constraint rules through the particle swarm optimization algorithm, and correct the three-dimensional point cloud of the wave field based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to the fluid dynamics characteristics. The fifth unit is used to analyze the geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field using feature networks, and to calculate the three-dimensional feature parameters of the wave field.

[0092] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0093] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0094] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional wave field measurement and analysis method based on multi-view cameras, characterized in that, include: Multi-angle image sequences of the wave field are acquired by a multi-view camera, the multi-view camera is calibrated, the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera are obtained, and the spatial projection relationship of the wave field surface points between different cameras is calculated based on the projection mapping network. Based on the spatial projection relationship, a viewpoint mapping association of feature points is established in the multi-angle image sequence to obtain the feature point set of the wave field. The spatial location distribution of the feature point set is calculated by the reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. A wave field dynamic feature learning network is used to analyze the temporal variation law of the initial three-dimensional point cloud. The time-varying features of the wave field are extracted through a temporal attention mechanism. The initial three-dimensional point cloud is then dynamically enhanced by an optimized network to generate a wave field three-dimensional point cloud with temporal consistency. Deformation constraint rules for the wave field are constructed based on fluid dynamics equations. The control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. The three-dimensional point cloud of the wave field is corrected based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to fluid dynamics characteristics. The geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field are analyzed using feature networks to calculate the three-dimensional feature parameters of the wave field.

2. The method according to claim 1, characterized in that, The multi-view camera is calibrated, and its intrinsic and extrinsic parameter matrices are obtained. The spatial projection relationship of wave field surface points between different cameras is calculated based on a projection mapping network, including: The intrinsic parameter matrix and the extrinsic parameter matrix are input into the calibration parameter diffusion model. The calibration parameter diffusion model uses a Gaussian distribution to describe the temporal evolution of the parameters. A noise scheduling strategy is set based on the dynamic change characteristics of the wave field. The noise scale in the diffusion process is adjusted according to the noise scheduling strategy to generate the temporal probability distribution of the calibration parameters. Calculate parameter gradient information based on the time-series probability distribution, determine the sampling direction according to the parameter gradient information, perform directional sampling on the time-series probability distribution, and obtain multiple sets of calibration parameter samples; The spatial mapping relationship of wave field surface points between different cameras is calculated based on the multiple sets of calibration parameter samples. The spatial mapping relationship is then subjected to probability integration to obtain the projection transformation probability of wave field surface points. The posterior probabilities of the multiple sets of calibration parameter samples are calculated based on the correspondence of feature points on the wave field surface. The posterior probabilities are then input into the calibration parameter diffusion model for parameter optimization to obtain the optimal projection mapping parameters and establish the accurate projection relationship between wave field surface points and different cameras.

3. The method according to claim 2, characterized in that, Based on the multiple sets of calibration parameter samples, the spatial mapping relationship of wave field surface points between different cameras is calculated. Probability integration is then performed on the spatial mapping relationship to obtain the projection transformation probability of the wave field surface points, including: Based on the multiple sets of calibration parameter samples, the projection transformation relationship of wave field surface points under a single camera is established, and the uncertainty distribution of wave field surface points during the camera view transformation process is analyzed according to the projection transformation relationship. Based on the uncertainty distribution, a mapping probability density function is constructed. The mapping probability density function uses the Dirac function to characterize the projection change law of the wave field surface point when it is transformed from the initial camera view to the target camera view. The Dirac function and the mapping probability density function are used to perform statistical analysis on the multiple sets of calibration parameter samples to obtain the initial projection probability of the wave field surface point. The initial projection probability and the mapping probability density function are integrated in the target camera imaging plane region. The inter-view mapping probability of the wave field surface points is obtained based on the result of the integration. Spatial consistency constraints between camera viewpoint pairs are constructed based on the inter-view mapping probability and the Dirac function. The spatial consistency constraint is minimized and optimized. Based on the result of the minimization optimization and the mapping probability density function, the optimal mapping probability that satisfies spatial consistency is obtained. The projection transformation probability between adjacent camera pairs is calculated based on the optimal mapping probability and the initial projection probability.

4. The method according to claim 1, characterized in that, Based on the aforementioned spatial projection relationship, a viewpoint mapping association for feature points is established in the multi-angle image sequence to obtain a feature point set of the wave field. The spatial distribution of the feature point set is then calculated using a reconstruction network to reconstruct the initial three-dimensional point cloud of the wave field, including: A viewpoint mapping matrix is ​​constructed based on spatial projection relationship. Two-dimensional coordinate information and feature description information of wave field surface feature points are extracted from multi-angle image sequences. The viewpoint mapping matrix is ​​used to calculate the projection deviation value of feature points from the current viewpoint to adjacent viewpoints. Projection geometric constraints between feature points are established based on the projection deviation value. A matching score function is constructed, which integrates the projection geometric constraints and the feature description information. The matching score function calculates the correspondence score of feature points in adjacent viewpoints, and establishes the viewpoint mapping association of feature points based on the correspondence score. The feature point correspondences in the viewpoint mapping association are filtered, and the feature point correspondences whose matching degree scoring function values ​​are less than a preset scoring threshold are retained. A wave field feature point set is formed based on the retained feature point correspondences. The multi-view projection relationship of feature points in the wave field feature point set is analyzed by using a reconstruction network. The three-dimensional spatial position distribution of feature points is optimized by iteratively minimizing the matching degree scoring function. The optimal three-dimensional coordinates of the feature points are determined based on the three-dimensional spatial position distribution. The initial three-dimensional point cloud of the wave field surface is reconstructed based on the optimal three-dimensional coordinates.

5. The method according to claim 1, characterized in that, A wave field dynamic feature learning network is used to analyze the temporal variation pattern of the initial 3D point cloud. The time-varying features of the wave field are extracted through a temporal attention mechanism. The initial 3D point cloud is then dynamically enhanced using an optimized network to generate a wave field 3D point cloud with temporal consistency, including: A wave field dynamic feature learning network is used to perform temporal sampling on the initial three-dimensional point cloud, extract the point cloud sequence within a fixed time window, and construct a temporal feature vector from the point cloud sequence; A temporal attention mechanism is constructed, wherein the temporal attention mechanism uses forward and reverse dual channels to calculate the dynamic correlation of the temporal feature vector at adjacent time points. The forward channel analyzes the motion trend of the wave field from the past to the present, and the reverse channel analyzes the motion law of the wave field from the present to the past, and establishes an attention weight matrix. The attention weight matrix is ​​used to perform multi-scale adaptive enhancement on the temporal features of wave field surface points. The multi-scale adaptive enhancement extracts the temporal correlation features of wave field surface points in the local neighborhood and the global range respectively, thereby obtaining the time-varying features of the wave field. The time-varying features are input into the optimization network, and a consistency constraint based on the temporal gradient is constructed based on the optimization network. The consistency constraint based on the temporal gradient ensures the continuity of motion of the wave field surface points at adjacent time points. The feature offset of the wave field surface points is calculated through iterative optimization. Dynamic feature enhancement is performed on the initial three-dimensional point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points and generate a wave field three-dimensional point cloud with temporal consistency.

6. The method according to claim 5, characterized in that, Dynamic feature enhancement is performed on the initial 3D point cloud based on the feature offset to achieve temporal continuous reconstruction of wave field surface points, generating a temporally consistent 3D point cloud of the wave field, including: A chaotic dynamic model of the wave field is constructed based on the characteristic offset. The chaotic dynamic model describes the evolution of the wave field state vector over time through a nonlinear dynamic function. The trajectory deviation is calculated in the phase space of the chaotic dynamic model based on the wave field state vector. The trajectory deviation is substituted into the logarithmic growth rate function to obtain the Lyapunov exponent sequence. The maximum Lyapunov exponent is extracted from the Lyapunov exponent sequence. Substituting the maximum Lyapunov exponent, the preset benchmark enhancement coefficient, and the decay factor into the exponential decay function, an adaptive weight factor for feature enhancement is obtained. The feature offset is then adjusted using the adaptive weight factor to obtain enhanced features. The enhanced features are superimposed on the initial three-dimensional point cloud to form an enhanced point cloud. The enhanced point cloud is then substituted into the nonlinear dynamic function for integration. The result of the integration operation represents the temporal evolution trajectory of the points on the wave field surface. Based on the temporal evolution trajectory, the enhanced point cloud is continuously reconstructed to generate a three-dimensional point cloud of the wave field with temporal consistency.

7. The method according to claim 1, characterized in that, The control parameters of the deformation constraint rules are iteratively optimized using a particle swarm optimization algorithm. Based on the optimized deformation constraint rules, the three-dimensional point cloud of the wave field is corrected to generate high-precision three-dimensional wave field data that conforms to hydrodynamic characteristics, including: Deformation constraint rules for a three-dimensional point cloud of a wave field are constructed. The deformation constraint rules include three types of hydrodynamic constraint terms. Each type of hydrodynamic constraint term corresponds to a constraint weight coefficient. A constraint weight vector is constructed based on the constraint weight coefficient. The constraint weight vector is mapped to a position vector in the particle swarm optimization space, and an initial velocity vector of the particle swarm is randomly generated based on the position vector; the constraint weight vector is substituted into the deformation constraint rule to calculate the constraint residual, and the constraint residual and the spatial curvature of the three-dimensional point cloud of the wave field are used to construct the optimization objective function. The optimization objective function is used to evaluate the quality of each particle's position in the particle swarm. The historical best position is taken as the individual best position, and the best position among all particles is taken as the global best position. The velocity component is calculated based on the individual best position and the global best position. The velocity components are superimposed to form the velocity increment of the particle swarm. The position vector of the particle swarm is iteratively updated using the velocity increment until the optimal constraint weight vector is obtained. The optimal constraint weight vector is substituted into the deformation constraint rule to correct the three-dimensional point cloud of the wave field, thereby obtaining a high-precision three-dimensional point cloud of the wave field that satisfies the hydrodynamic characteristics.

8. A three-dimensional wave field measurement and analysis system based on a multi-view camera, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire multi-angle image sequences of the wave field through a multi-view camera, calibrate the multi-view camera, obtain the intrinsic parameter matrix and extrinsic parameter matrix of the multi-view camera, and calculate the spatial projection relationship of the wave field surface points between different cameras based on the projection mapping network. The second unit is used to combine the spatial projection relationship to establish the viewpoint mapping association of feature points in the multi-angle image sequence, obtain the feature point set of the wave field, and calculate the spatial position distribution of the feature point set through reconstruction network analysis to reconstruct the initial three-dimensional point cloud of the wave field. The third unit is used to analyze the temporal variation law of the initial three-dimensional point cloud using a wave field dynamic feature learning network, extract the time-varying features of the wave field through a temporal attention mechanism, and perform dynamic feature enhancement on the initial three-dimensional point cloud through an optimization network to generate a wave field three-dimensional point cloud with temporal consistency. The fourth unit is used to construct deformation constraint rules for the wave field based on the fluid dynamics equations, iteratively optimize the control parameters of the deformation constraint rules through the particle swarm optimization algorithm, and correct the three-dimensional point cloud of the wave field based on the optimized deformation constraint rules to generate high-precision three-dimensional wave field data that conforms to the fluid dynamics characteristics. The fifth unit is used to analyze the geometric structure and dynamic change information in the high-precision three-dimensional data of the wave field using feature networks, and to calculate the three-dimensional feature parameters of the wave field.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images