Golf club head three-dimensional spatial pose reconstruction method based on explicit boundary guidance and cross-view cross attention

Abstract

This invention relates to the interdisciplinary fields of computer vision, artificial intelligence, and sports technology, and particularly to a method for reconstructing the 3D spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention. The method includes: acquiring a multi-view synchronous image sequence; predicting an adaptive region of interest using a physically guided neural network; extracting a motion-blurred boundary probability field using an explicit boundary guidance module to obtain a boundary prior spatial feature map; inputting the multi-view feature map into a cross-view cross-attention module to achieve epipolar-level feature fusion based on epipolar geometric constraints; and decoding and outputting the club head's 3D pose parameters using a golf double pendulum kinematic prior model. This application effectively overcomes the problems of severe motion blur and multi-view self-occlusion in high-speed swings, achieving high-precision and highly robust 3D pose reconstruction.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, artificial intelligence and sports technology, and in particular to a method for reconstructing the three-dimensional spatial pose of golf club heads based on explicit boundary guidance and cross-viewpoint cross-attention. Background Technology

[0002] With the deep integration of computer vision and artificial intelligence technologies, machine vision-based indoor sports simulation equipment is gradually replacing traditional radar or infrared grating solutions. In indoor golf simulation systems, the three-dimensional spatial pose of the golf clubhead at the moment of impact—including parameters such as three-dimensional coordinates, clubface closure angle, clubhead trajectory, and swing plane—is the core basis for evaluating swing quality and calculating the initial launch conditions of the ball (ball speed, launch angle, sidespin, backspin, etc.).

[0003] However, existing 3D reconstruction technologies for golf club heads based on binocular or multi-view vision still face significant technical bottlenecks, mainly in the following aspects: First, extreme motion blur causes feature extraction to fail. Professional golfers can achieve clubhead linear speeds exceeding 50 m / s at the moment of impact. Even with high-speed cameras boasting global shutter speeds exceeding 1000 fps, the clubhead will still produce tens of pixels of motion blur on the imaging plane within a few hundred microseconds of exposure time. Traditional feature point extraction algorithms (such as SIFT, SURF, or deep learning-based keypoint detection) struggle to accurately locate the geometric apex of the clubhead in such severely blurred images. This inaccurate feature extraction directly leads to a significant decrease in the accuracy of 3D reconstruction.

[0004] Second, there is the issue of post-processing independence in multi-view feature fusion. Existing systems generally employ a "late fusion" strategy, where each viewpoint independently completes 2D keypoint or bounding box detection, and then calculates 3D coordinates through triangulation using calibration parameters. In this approach, if any viewpoint fails to detect 2D objects due to sudden changes in illumination, body occlusion, or motion blur, the error will be nonlinearly amplified during the triangulation stage, potentially even causing pose calculation failure. Current technologies lack an effective mechanism for achieving cross-view information interaction during the depth feature extraction stage.

[0005] Third, there is a lack of prior constraints in physical kinematics. The golf swing is a typical rigid body motion, and the human arm and club system can be highly approximated as a "double pendulum model". However, current AI vision models usually treat image sequences as discrete frames without physical constraints and perform brute-force regression, ignoring the strict velocity, acceleration and moment of inertia constraints between consecutive frames, which leads to jitter or abrupt changes in the predicted pose sequence on the time axis that violate physical laws.

[0006] In summary, there is an urgent need for a technological innovation that delves into the underlying algorithms to overcome extreme motion blur, achieve deep fusion of features at the multi-view level, and combine prior constraints of golf kinematics to output high-precision and highly robust three-dimensional spatial pose of the golf clubhead. Summary of the Invention

[0007] The purpose of this application is to provide a method, layout and system for reconstructing the three-dimensional spatial pose of golf clubheads based on explicit boundary guidance and cross-view cross-attention. It aims to solve the technical problems of low accuracy and poor robustness of clubhead three-dimensional reconstruction caused by high-speed motion blur, multi-view occlusion and lack of physical kinematic constraints in the prior art, and achieve high-precision, anti-blur and deep fusion pose reconstruction technology.

[0008] Firstly, this application provides a method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention, including: Acquire a multi-view synchronous continuous image sequence of the golf hitting area and combine it with a Physically Guided Neural Network (PGNN) to predict adaptive regions of interest (ROIs). Feature extraction is performed on the image within the adaptive ROI, and a pole head boundary probability field is constructed through the explicit boundary guidance (EBG) module to obtain a multi-view spatial feature map with explicit boundary prior information. The multi-view spatial feature map is input into the cross-view cross attention (CVCA) module. Using the epipolar geometry constraints of the multi-view camera system, cross attention calculation based on epipolar lines is performed between the feature maps of the multi-view to obtain cross-view fused features. The cross-view fusion features are input into the pose decoder, and a loss function is constructed in conjunction with the golf double pendulum kinematic prior model for network optimization, outputting the three-dimensional spatial pose data of the golf clubhead at the current moment.

[0009] In the above implementation process, PGNN combined with uncertainty covariance is used to generate adaptive ROIs, removing a large number of redundant background pixels and ensuring the concentration of computational resources. Secondly, to address the difficulty of high-speed motion blur, an explicit boundary guidance (EBG) algorithm is proposed. This algorithm does not forcibly predict points in the blurred image, but rather follows the physical laws of blurring, extracting the boundary probability field of the blurred envelope, which is used as a hard prior attention mask to cover the feature map, substantially solving the feature loss problem caused by motion blur. More importantly, this application introduces a cross-view cross-attention (CVCA) module, explicitly integrating the "epidial constraint" in multi-view geometry into the self-attention / cross-attention computation paradigm of the Transformer architecture. This allows blurred or occluded feature points in one view to directly initiate a "feature query" to the epipolar neighborhood of another view, completing multi-view information complementarity and depth calculation in the deep learning feature space, completely avoiding the error amplification caused by independent detection in a single view. Finally, prior constraints are applied using the kinematics of a golf pendulum to ensure the physical smoothness and absolute rationality of the temporal results.

[0010] Furthermore, the step of acquiring a multi-view synchronous continuous image sequence of the golf hitting area and combining it with a physically guided neural network to predict an adaptive region of interest includes: Based on multiple pre-calibrated high-speed cameras, acquire raw image sequences from multiple perspectives at the same time; then, from the previous moment... The three-dimensional motion state of the clubhead is input to the physical guidance neural network, which outputs the current moment based on the rigid body kinematics equations of the golf club. The predicted three-dimensional spatial coordinates and uncertainty covariance matrix; The predicted three-dimensional spatial coordinates are projected onto the two-dimensional pixel plane of each viewpoint through the intrinsic and extrinsic parameter matrix of the multi-view camera to obtain the two-dimensional predicted center point. Based on the projection characteristics of the uncertainty covariance matrix, spatial anisotropic scaling coefficients are calculated from each viewpoint. A dynamically changing two-dimensional adaptive ROI sequence is generated with the two-dimensional prediction center point as the center, and the ROI image is cropped from the original image sequence.

[0011] In the above implementation, the physics-guided neural network is not a simple Kalman filter. By outputting the uncertainty covariance matrix in three-dimensional space and then projecting it back onto the two-dimensional plane, the system can adaptively calculate a ROI (Region of Interest) box that is not fixed but has anisotropic scaling characteristics in length and width, based on the physical deformation (perspective distortion) of the same moving object observed from different perspectives. When the system lacks confidence in motion prediction (such as nonlinear deceleration during a collision at the moment of impact), the eigenvalues ​​of the covariance matrix increase, and the ROI automatically expands to prevent target loss; during the steady swing phase, the covariance matrix converges, and the ROI shrinks to reduce computational load.

[0012] Furthermore, the step of constructing the pole head boundary probability field through the explicit boundary guidance module to obtain a multi-view spatial feature map with explicit boundary prior information includes: The ROI image is input into a feature extraction backbone network to extract multi-scale intermediate layer features. An explicit boundary prediction branch is then connected after the high-level semantic features of the feature extraction backbone network. This explicit boundary prediction branch outputs a two-dimensional pole head boundary probability field. ; Using an adaptive threshold function to evaluate the probability field of the pole head boundary Binarization and Gaussian smoothing are performed to generate a boundary attention mask. ; The boundary attention mask The network features are element-wise multiplied with the multi-scale intermediate layer features and the residuals are added, so that the network features focus on the blurred boundary regions generated by high-speed motion in the spatial dimension, thus obtaining a multi-view spatial feature map. ,in Indicates the view index. The number of feature channels, and The height and width of the feature map.

[0013] In the above implementation process, traditional CNN feature extraction networks tend to have their receptive fields dispersed into the background region when dealing with motion-blurred targets. The explicit boundary prediction branch forces the network to specifically learn the gradient features of the "motion-blurred edges." The generated boundary attention mask... This is equivalent to providing a "focusing searchlight" for subsequent networks, after element-wise multiplication and residual addition. It effectively enhances the feature representation of blurred edges and suppresses interference from complex backgrounds such as grass and jerseys.

[0014] Further, the step of inputting the multi-view spatial feature map into the cross-view cross-attention module, and using the epipolar geometry constraints of the multi-view camera system to perform epipolar-based cross-attention calculation between the feature maps of the multiple views, includes: for the source view... Feature pixels in Utilizing the source perspective From the perspective of the target The fundamental matrix between Calculate the corresponding polar lines ,in For feature pixels Homogeneous coordinates; from the target's perspective Multi-view spatial feature map In, along the polar line Sampling is performed on the feature region where the query vector is located to generate a set of candidate key-value pairs; based on the query vector Key vector Sum value vector The calculation formula for performing cross-attention fusion is as follows: in, Spatial feature map from the source perspective At pixel The features at that location are generated through a linear transformation. and From the target perspective in the polar neighborhood The features of the sampling points within are generated through a linear transformation. For the feature vector dimension, The sampling band is dynamically adjusted to tolerate epipolar error. By concatenating the cross-attention output features from each perspective with the original perspective features and performing dimensionality reduction using a multilayer perceptron, cross-perspective fusion features are obtained.

[0015] In the above implementation process, this application ingeniously integrates multi-view solid geometry theory (epidural geometry) in computer vision with the cutting-edge Transformer architecture in artificial intelligence. Standard global attention mechanisms require computation... The complexity of the previous method is limited, and it cannot perceive the spatial positional relationships of the camera. In contrast, the CVCA algorithm in this application explicitly guides the network: if it wants to confirm a point in viewpoint A... The three-dimensional information can only be obtained from the polar line of viewpoint B. The system searches for corresponding features in the vicinity. This not only drastically reduces computational complexity from global search to linear band search, but also substantially uses the camera's physical intrinsic and extrinsic parameters as a guide for the network structure, giving the network a natural ability to resolve 3D space. This breaks through the "black box" limitation of traditional feature fusion, which only involves channel stacking or fully connected mapping.

[0016] Furthermore, the step of constructing a loss function based on the prior kinematic model of a golf double pendulum for network optimization includes: Construct a loss that includes data items Boundary-guided loss and kinematic prior loss The joint loss function is expressed as: in, For dynamic balancing weights; The kinematic prior loss The formula is expressed as: in, Let Lie algebras be the rotations at the current time and the previous two time points, respectively. Let the rod head inertia tensor be... The equivalent joint torque is derived based on the double pendulum model. Angular velocity, This is the inter-frame time interval. For positive kinematic functions, The parameter is the prior parameter for the rod length.

[0017] In the aforementioned implementation process, data-driven AI models are prone to jitter due to environmental noise. This application addresses this by designing a loss function that incorporates the rigid body Euler equations (including external torque and gyroscopic torque terms) as strong constraints into the loss landscape. (Li algebra) The use of Euler angles avoids the gimbal lock problem and the unit norm constraint problem of quaternions, enabling the network to have perfect continuous differentiability during backpropagation. This requires the network not only to "see" accurately on a single frame of image, but also to predict a three-dimensional trajectory that perfectly conforms to the laws of Newtonian classical mechanics and golf ball dynamics, eliminating the irregular oscillations in the reconstruction results.

[0018] Secondly, this application provides a layout for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-viewpoint cross-attention, including: Synchronous trigger control unit, used to generate a global synchronous trigger level signal with microsecond-level precision; A multi-view high-speed camera system includes at least two high-speed industrial cameras with global shutters, distributed in orthogonal or oblique positions in the golf hitting area. The high-speed industrial cameras receive the global synchronization trigger level signal to perform synchronous exposure acquisition. The high-frequency strobe lighting system is physically linked with the global synchronization trigger level signal to output a high-intensity narrow pulse light source within the camera exposure window; The spatial calibration component employs a three-dimensional calibration target, which is positioned at the center of the impact area to extract the external parameter matrix, internal parameter matrix, and distortion coefficients of each high-speed industrial camera in the world coordinate system.

[0019] Thirdly, this application provides a golf club head three-dimensional spatial pose reconstruction system based on explicit boundary guidance and cross-view cross-attention, including: The data acquisition module is used to acquire a multi-view synchronous continuous image sequence of the golf hitting area and combine it with a physics-guided neural network to predict adaptive regions of interest. The explicit boundary feature extraction module is used to extract features from the image within the adaptive region of interest. The explicit boundary guidance module constructs a pole head boundary probability field to obtain a multi-view spatial feature map with explicit boundary prior information. The cross-view fusion module is used to input the multi-view spatial feature map into the cross-view cross-attention module, and use the epipolar geometry constraint of the multi-view camera system to perform epipolar-based cross-attention calculation between the feature maps of the multi-view to obtain cross-view fused features. The pose decoding and optimization module is used to input the cross-view fusion features into the pose decoder, and combine the golf double pendulum kinematic prior model to construct a loss function for network optimization, and output the three-dimensional spatial pose data of the golf club head at the current moment.

[0020] Fourthly, this application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method described in any of the first aspects.

[0021] Fifthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in any of the first aspects.

[0022] The technical effects and advantages of this invention are as follows: Physics-guided adaptive ROI strategy By combining a physically guided neural network with an uncertainty covariance matrix, a dynamically adaptive region of interest (ROI) with spatially anisotropic scaling characteristics is generated. This method can automatically adjust the region size based on motion prediction confidence, reducing redundant background computation while maintaining target recall, thus helping to maintain the stability of target tracking in high-speed motion scenarios.

[0023] Explicit boundary-guided feature extraction mechanism To address the blurring problem in high-speed motion, an explicit boundary prediction branch outputs the pole head boundary probability field, which is then transformed into a spatial attention mask applied to multi-scale intermediate features. This mechanism guides the network to focus on the blurred boundary region, suppressing interference from complex backgrounds and providing a more discriminative feature representation for subsequent pose calculation.

[0024] Cross-view feature fusion based on polar geometry constraints The cross-view attention module embeds epipolar geometric constraints into the attention computation process, restricting feature matching search to the epipolar neighborhood. This method reduces the complexity of global attention computation and, when occlusion or feature blurring exists in a single viewpoint, achieves information complementarity through cross-view feature interaction, which helps improve the robustness of 3D pose reconstruction.

[0025] Temporal pose optimization incorporating kinematic priors Prior constraints based on a golf pendulum kinematic model are introduced into the loss function, and the pose of consecutive frames is smoothed using the rigid body rotation differential equation. This design helps to suppress temporal jitter caused by pure visual regression, making the output pose sequence more consistent with the physical continuity of rigid body motion. Attached Figure Description

[0026] Figure 1 A schematic diagram of the overall algorithm flow for a golf club head 3D spatial pose reconstruction method based on explicit boundary guidance and cross-view cross attention, provided for embodiments of this application; Figure 2 The internal network topology and feature flow diagram of the explicit boundary guidance (EBG) feature extraction module provided in this embodiment of the application; Figure 3 A schematic diagram of the distribution of multiple cameras and the intersection of the fields of view in the physical layout system for reconstructing the three-dimensional spatial pose of a golf club head provided in an embodiment of this application; Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Terminology Definition For ease of understanding, the key terms used in this application are explained below: (1) Physically Guided Neural Network (PGNN): A network structure that combines known rigid body kinematics equations with deep learning models to predict the motion state of a target and its uncertainties.

[0029] (2) Explicit Boundary Guidance (EBG): This technique involves outputting a motion-blurred boundary probability field through a branch network and converting it into a spatial attention mask to guide the feature extraction direction of the main network.

[0030] (3) Cross-view cross attention (CVCA): A computational method that embeds multi-view epipolar geometric constraints into the Transformer cross attention mechanism to perform feature matching and fusion in the epipolar neighborhood.

[0031] (4) Lie algebra so(3): The Lie algebra corresponding to the special orthogonal group SO(3) is represented by a three-dimensional vector to avoid the gimbal lock problem of Euler angles and has good continuous differentiability in network optimization.

[0032] Technical Problem Analysis Existing 3D pose reconstruction schemes for golf club heads mostly adopt a technical approach of "independent detection from a single viewpoint and late-stage triangulation." This approach has the following shortcomings: First, severe motion blur caused by high-speed movement significantly reduces the accuracy of 2D keypoint detection; second, independent processing of each viewpoint lacks information interaction at the feature level, and occlusion or detection failure at a single viewpoint will lead to nonlinear amplification of the overall reconstruction error; third, the pure data-driven regression method ignores the kinematic constraints between consecutive frames, resulting in jitter and abrupt changes in the reconstructed trajectory that do not conform to physical laws.

[0033] To address the aforementioned issues, this application proposes a 3D spatial pose reconstruction method based on explicit boundary guidance and cross-viewpoint cross-attention. Through a physical guidance ROI strategy, boundary attention feature extraction, feature-level fusion of epipolar geometric constraints, and joint optimization of kinematic priors, high-precision and robust pose reconstruction is achieved.

[0034] Please see Figure 1 , Figure 1 This is a schematic diagram of the overall process of a golf club head three-dimensional spatial pose reconstruction method based on explicit boundary guidance and cross-view cross attention, provided in an embodiment of this application.

[0035] S100: Physically Guided Adaptive Region of Interest (ROI) Extraction Strategy Golf camera images typically have extremely high resolution (e.g., 2000x2000 or higher), but the clubhead only occupies a few dozen pixels in the frame. Inputting the entire original image into a deep neural network would result in a massive computational burden, making the system unsuitable for real-time performance. Traditional methods often use a fixed-size rectangle extending from the previous frame's position as the search area for the current frame. However, due to the drastic deceleration at impact, the target can easily fly out of the fixed ROI.

[0036] To address this issue, this application introduces a Physically Guided Neural Network (PGNN) incorporating uncertainty. The core of PGNN lies in implicitly embedding the state transition logic of the Kalman filter into the memory gates of the RNN (Recurrent Neural Network).

[0037] Set the state vector as , respectively representing three-dimensional position, velocity, acceleration, Euler angles, and angular velocity.

[0038] PGNN forward prediction module output: Simultaneously, the network also outputs an uncertainty covariance matrix representing the prediction confidence level. .

[0039] Using the camera's projection matrix ( For internal reference purposes. (as an external parameter), project the three-dimensional prediction center onto the first... A pixel plane from a single viewpoint: Next, the anisotropic scaling factor is calculated on the two-dimensional plane. This is based on the projective covariance of the covariance matrix. (in (For the Jacobian determinant of the projection matrix), calculate its eigenvalues. And feature vectors. Along the direction of the feature vectors, using the corresponding feature values ​​as size scaling factors, a dynamically changing adaptive ROI that perfectly fits the current motion uncertainty is generated. This method ensures that the recall rate of the target reaches 100% while maximizing the utilization of computational resources.

[0040] S200: Explicit Boundary Guided (EBG) Feature Extraction Please see Figure 2 After ROI cropping, the image is fed into a feature extraction backbone network based on convolutional neural networks (CNN) or VisionTransformer (such as a modified version of ResNet-50).

[0041] Due to motion blur, conventional feature extraction easily loses the target texture. This application designs an explicit boundary prediction branch. This branch does not predict the exact point of the pole head, but rather predicts the "high-gradient physical boundary" generated by motion blur. Let the image at point... The gray value at that location is The fuzzy trajectory generated by the high-speed motion of the pole head can theoretically be modeled as an anisotropic Gaussian kernel function integral. The network, through supervised learning, is trained to output a representation of... The two-dimensional boundary probability field.

[0042] Specifically, to allow the backbone network to utilize this boundary information, a soft attention mechanism is employed. This involves a transformation using an adaptive threshold function and Gaussian smoothing. Mask this attention The output feature map of the backbone network at a certain layer (e.g., Layer 3) Perform element-wise multiplication and combine it with a residual connection: in, This is a learnable scalar parameter. This operation redistributes network features in the spatial dimension, forcing the network to focus its perceptual "attention" on regions with explicitly defined boundaries. It effectively masks interference from background light spots and grass textures, resulting in high-quality multi-view spatial feature maps. .

[0043] S300: Epipolar geometry constraint computation of the Cross-View Cross-Attention (CVCA) module This step integrates the underlying geometric principles of computer vision into the mathematical derivation of the AI ​​architecture. A left camera viewpoint (source viewpoint) is provided. ) and the right camera view (target view) ).exist Feature map A certain pixel Hope in Feature map The corresponding feature regions are searched to complete the depth information calculation.

[0044] In traditional Transformer cross-attention... It will go to all pixels in the target feature map ( Performing inner product matching calculations on (a number of points) is not only computationally intensive (with a complexity of quadratic), but more importantly, it can create the illusion of "pseudo-matching" in completely unrelated background areas.

[0045] This application utilizes the fundamental matrix obtained by calibration using multiple cameras. (Fundamental Matrix). For the homogeneous coordinates of the source feature points. Its projection onto the target's viewpoint is not a point, but a straight line—the epipolar line. The coefficient vector of its epipolar equation is expressed as: The equation of the straight line is: .

[0046] This means that, from the perspective of the target In, with The physical corresponding real point must and can only exist on this polar line. Therefore, the cross-attention mechanism of the Transformer was restructured. Instead of global feature point sampling, it now focuses on the cross-attention along a straight line. and the allowable neighborhood of small camera calibration errors (such as normal distance less than...) strip region Point sampling is performed within the area.

[0047] The specific formula derivation is as follows: Source Generate query vectors: polar neighborhood The index of the set of sampling points within is denoted as Generate key and value vectors: Perform cross-attention computation subject to epipolar constraints: in, To hide the feature channel dimension, and to scale the dot product to prevent the Softmax gradient from vanishing; is a trainable linear mapping weight matrix.

[0048] This mechanism, if the pole head is partially obscured by the left camera, will affect the feature map on the left side. The extracted features are ambiguous. When this ambiguous feature is used as... Once sent out, the feature map on the right will return strictly along the epipolar line. The information is such that, since the features of the right camera within the epipolar region are clear and complete, the attention mechanism spontaneously assigns the clear features from the right to the blurred features on the left through extremely high similarity weights. This achieves multi-view information complementarity at the feature level.

[0049] Finally, the features aggregated from various perspectives are concatenated by channel and fed into layer normalization and multilayer perceptron (MLP) for dimensionality reduction to obtain cross-perspective fused feature representations.

[0050] S400: Lie algebra pose decoding and kinematic prior loss function optimization After obtaining the cross-view fusion features, the network needs to output the target's 3D pose. To avoid gimbal lock and normalization issues, the rotation matrix representation is transformed into a Lie algebra. The translation vector is The pose decoder directly regresses the values ​​of these two variables through a series of fully connected layers.

[0051] To ensure the convergence of network training and the absolute observability of physical laws, this application innovatively constructs a joint loss function: Data item loss The Euclidean space error between the predicted 3D keypoint coordinates and the ground truth (obtained through a high-precision optical motion capture system) is evaluated using smoothed L1 loss.

[0052] Boundary Guided Loss The binary cross-entropy loss (BCE Loss) is used to force the boundary probability field output by the auxiliary branch to approximate the fuzzy boundary mask label generated by artificial or synthetic data.

[0053] Kinematic prior loss The golf swing is a typical rigid body dynamics system. According to the Newton-Euler equations of rigid body dynamics, its angular acceleration is directly related to the net torque applied to the clubhead and its moment of inertia. In extremely short exposure intervals... Within a 1-millisecond timeframe (e.g., 1 millisecond), the motion sequence should never experience abrupt jumps. The angular velocity is approximated as a difference... The angular acceleration is approximately: According to the law of rigid body rotation (in Given the known moment of inertia tensor of the cue stick, (For torque). The network must not only fit the spatial location, but also satisfy the following second-order differential time continuity constraint: In addition, a constraint term for the invariant length of the rigid body of the cue is introduced: through the positive kinematic function. By combining the near-end coordinates of the shaft with Lie algebraic rotations to deduce the clubhead coordinates, the deviation between the deduce coordinates and the directly regressed coordinates is penalized. This physical constraint can not only serve as a supervisory signal to regulate the weight evolution of the neural network during training, but also as the optimal smoothing objective function term in the actual inference testing phase, greatly improving the model's generalization robustness under unpredictable lighting conditions.

[0054] III. Physical Layout and System Hardware Platform for 3D Spatial Pose Reconstruction of Golf Club Heads Based on Explicit Boundary Guidance and Cross-View Cross-Attention Please see Figure 3 This method relies on a highly synchronized physical acquisition layout architecture. To ensure that the epipolar geometry baseline described above does not shift, the physical layout in this embodiment requires an extremely stringent industrial-grade configuration.

[0055] The hardware system includes: Synchronous Trigger Hub: Employing an FPGA core architecture, it generates a 1000Hz square wave synchronization level signal (TTL) with a pulse width accuracy of 1 microsecond. This is the lifeline ensuring the "simultaneity" of multi-view acquisition; if a millisecond-level misalignment occurs, the epipolar projection under high-speed motion will fail and collapse.

[0056] Multi-view high-speed camera system: This system comprises at least two (ideally three cameras: front-view, top-view, and side-view) high-speed industrial cameras equipped with global shutter CMOS sensors. The global shutter ensures simultaneous exposure of all rows of pixels on the image sensor, eliminating the rolling shutter effect. Each camera uses a wide-angle or standard distortion-free fixed-focus lens.

[0057] High-frequency strobe light system: It is directly connected to the interrupt output pin of the synchronous trigger control unit via a physical cable. When the camera begins an extremely short exposure (e.g., 200 microseconds), the high-energy LED array instantly releases an overload current to emit an extremely strong pulse beam, which illuminates the details of the camera tip while using the extremely narrow physical pulse width to help freeze motion blur.

[0058] Spatial calibration component: Employs a specially designed three-dimensional non-coplanar calibration target. Based on an evolved version of Zhang Zhengyou's calibration method, it calculates the intrinsic parameter matrix for each camera. External rotation matrix and external translation vector and radial and tangential distortion coefficients All parameters are fixed in the system's configuration file, providing precise basic matrix operation parameters for the CVCA module.

[0059] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention, characterized in that, include: Acquire a multi-view synchronous continuous image sequence of the golf hitting area and combine it with a physics-guided neural network to predict adaptive regions of interest; Feature extraction is performed on the image within the adaptive region of interest, and a pole head boundary probability field is constructed through an explicit boundary guidance module to obtain a multi-view spatial feature map with explicit boundary prior information. The multi-view spatial feature map is input into the cross-view cross-attention module. Using the epipolar geometry constraint of the multi-view camera system, cross-attention calculation based on epipolar lines is performed between the feature maps of the multi-view to obtain cross-view fused features. The cross-view fusion features are input into the pose decoder, and a loss function is constructed in conjunction with the golf double pendulum kinematic prior model for network optimization, outputting the three-dimensional spatial pose data of the golf clubhead at the current moment.

2. The method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention as described in claim 1, is characterized in that, The step of acquiring a multi-view synchronous continuous image sequence of the golf hitting area and combining it with a physically guided neural network to predict an adaptive region of interest includes: Based on multiple pre-calibrated high-speed cameras, acquire raw image sequences from multiple perspectives at the same time. The previous moment The three-dimensional motion state of the clubhead is input to the physical guidance neural network, which outputs the current moment based on the rigid body kinematics equations of the golf club. The predicted three-dimensional spatial coordinates and uncertainty covariance matrix; The predicted three-dimensional spatial coordinates are projected onto the two-dimensional pixel plane of each viewpoint through the intrinsic and extrinsic parameter matrix of the multi-view camera to obtain the two-dimensional predicted center point. Based on the projection characteristics of the uncertainty covariance matrix, spatial anisotropic scaling coefficients are calculated at each viewpoint. A dynamically changing two-dimensional adaptive region of interest sequence is generated with the two-dimensional prediction center point as the center, and the region of interest image is cropped from the original image sequence.

3. The method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention as described in claim 1, is characterized in that, The step of constructing the pole head boundary probability field through the explicit boundary guidance module to obtain a multi-view spatial feature map with explicit boundary prior information includes: The region of interest image is input into the feature extraction backbone network to extract multi-scale intermediate layer features; After extracting high-level semantic features from the feature extraction backbone network, an explicit boundary prediction branch is connected. This explicit boundary prediction branch outputs a two-dimensional pole head boundary probability field. ; Using an adaptive threshold function to evaluate the probability field of the pole head boundary Binarization and Gaussian smoothing are performed to generate a boundary attention mask. ; The boundary attention mask The network features are element-wise multiplied with the multi-scale intermediate layer features and the residuals are added, so that the network features focus on the blurred boundary regions generated by high-speed motion in the spatial dimension, thus obtaining a multi-view spatial feature map. ,in Indicates the view index. The number of feature channels, and The height and width of the feature map.

4. The method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention as described in claim 3, is characterized in that, The step of inputting the multi-view spatial feature map into the cross-view cross-attention module, and using the epipolar geometry constraints of the multi-view camera system to perform epipolar-based cross-attention calculation between the feature maps of the multiple views, includes: From the source perspective Feature pixels in Utilizing the source perspective From the perspective of the target The fundamental matrix between Calculate the corresponding polar lines ,in For feature pixels homogeneous coordinates; From the perspective of the target Multi-view spatial feature map In, along the polar line Sample the feature region where it is located to generate a set of candidate key-value pairs; Based on query vector Key vector Sum value vector The calculation formula for performing cross-attention fusion is as follows: middle, Spatial feature map from the source perspective At pixel The features at that location are generated through a linear transformation. and From the target perspective in the polar neighborhood The features of the sampling points within are generated through a linear transformation. For the feature vector dimension, The sampling band is dynamically adjusted to tolerate epipolar error. By concatenating the cross-attention output features from each perspective with the original perspective features and performing dimensionality reduction using a multilayer perceptron, cross-perspective fusion features are obtained.

5. The method for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention as described in claim 1 or 4, characterized in that, The three-dimensional spatial pose data output by the pose decoder includes the three-dimensional coordinates of the clubhead's center of mass. and three-dimensional rotation parameters based on Lie algebra representation The steps for constructing a loss function based on the prior kinematic model of a golf double pendulum and optimizing the network include: Construct a loss that includes data items Boundary-guided loss and kinematic prior loss The joint loss function is expressed as: in, For dynamic balancing weights; The kinematic prior loss The formula is expressed as: in, Let Lie algebras be the rotations at the current time and the previous two time points, respectively. Let the rod head inertia tensor be... The equivalent joint torque is derived based on the double pendulum model. Angular velocity, This is the inter-frame time interval. For positive kinematic functions, The parameter is the prior parameter for the rod length.

6. A layout for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-viewpoint cross-attention, characterized in that, include: Synchronous trigger control unit, used to generate a global synchronous trigger level signal with microsecond-level precision; A multi-view high-speed camera system includes at least two high-speed industrial cameras with global shutters, distributed in orthogonal or oblique positions in the golf hitting area. The high-speed industrial cameras receive the global synchronization trigger level signal to perform synchronous exposure acquisition. The high-frequency strobe lighting system is physically linked with the global synchronization trigger level signal to output a high-intensity narrow pulse light source within the camera exposure window; The spatial calibration component employs a three-dimensional calibration target, which is positioned at the center of the impact area to extract the external parameter matrix, internal parameter matrix, and distortion coefficients of each high-speed industrial camera in the world coordinate system.

7. The layout for reconstructing the three-dimensional spatial pose of a golf club head based on explicit boundary guidance and cross-view cross-attention as described in claim 6, is characterized in that, The multi-view high-speed camera system includes a top-down camera, a front-level camera, and a side-rear oblique camera. The optical axes of each camera converge at the origin of the ball's center, and the angle between the principal optical axes of any two cameras is greater than 30 degrees and less than 150 degrees.

8. A three-dimensional spatial pose reconstruction system for golf club heads based on explicit boundary guidance and cross-viewpoint cross-attention, characterized in that, include: The data acquisition module is used to acquire a multi-view synchronous continuous image sequence of the golf hitting area and combine it with a physics-guided neural network to predict adaptive regions of interest. The explicit boundary feature extraction module is used to extract features from the image within the adaptive region of interest. The explicit boundary guidance module constructs a pole head boundary probability field to obtain a multi-view spatial feature map with explicit boundary prior information. The cross-view fusion module is used to input the multi-view spatial feature map into the cross-view cross-attention module, and use the epipolar geometry constraint of the multi-view camera system to perform epipolar-based cross-attention calculation between the feature maps of the multi-view to obtain cross-view fused features. The pose decoding and optimization module is used to input the cross-view fusion features into the pose decoder, and combine the golf double pendulum kinematic prior model to construct a loss function for network optimization, and output the three-dimensional spatial pose data of the golf club head at the current moment.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the golf club head three-dimensional spatial pose reconstruction method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the golf club head three-dimensional spatial pose reconstruction method as described in any one of claims 1 to 5.

Images