High-precision reconstruction method and system for three-dimensional model of sports person

Abstract

The invention discloses a high-precision reconstruction method and system for a three-dimensional model of a sportsman, relates to the technical field of computer vision and three-dimensional modeling, and aims to solve the problems of dependence on mark points, inconsistent model topology and detail missing in the prior art. According to the system, a depth image and inertia measurement data are synchronously acquired through a multi-modal data acquisition module, deep fusion of vision and inertia information is realized through a time sequence feature fusion module, and a symbol distance value and a non-rigid motion deformation amount are output based on an implicit dynamic field construction module; topological consistency constraint is combined to ensure that model deformation accords with biomechanical characteristics, and finally, a high-fidelity model generation module extracts contour surfaces and corrects the contour surfaces frame by frame to generate a three-dimensional human body model sequence which is coherent in time sequence and has submillimeter-level details.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and three-dimensional modeling, and particularly relates to a high-precision reconstruction method and system for a three-dimensional model of a moving person. BACKGROUND

[0002] Three-dimensional human modeling and motion capture are key technologies in the field of computer vision and graphics, and are widely used in virtual reality, sports training, special effects for films and television, and human-computer interaction. The technology aims to reconstruct a three-dimensional human model with accurate geometric and motion information from sensor data or image sequences. Among them, high-precision reconstruction of a three-dimensional model of a moving person is an important direction of current research, and the core goal is to restore a high-fidelity, detailed and motion-consistent three-dimensional human form from a dynamic sequence.

[0003] Existing technologies mainly rely on marker-based optical motion capture systems or multi-view video reconstruction methods, but these schemes have significant limitations. The marker-based system can provide high precision, but requires a large number of marker points to be pasted on the human body surface, which is costly and severely limits the use of the scene. The multi-view reconstruction method does not require markers, but is easily affected by environmental lighting, occlusion and clothing texture, and the reconstruction result often has geometric noise, missing details and motion artifacts. In addition, traditional methods are difficult to maintain model topological consistency in dynamic sequences, and the accuracy of pose estimation is insufficient for fast motion or self-occlusion, resulting in deformation distortion or motion inconsistency of the reconstructed model in the joint area.

[0004] Existing three-dimensional reconstruction algorithms generally face the balance problem of model precision and computational efficiency when dealing with high dynamic and multi-degree-of-freedom human motion. High-precision models usually require dense point clouds or complex mesh deformation, but the computational overhead is large and difficult to apply in real time; while lightweight methods can improve efficiency, but at the expense of model details and motion smoothness. Especially in professional scenes such as sports training, it is required that the model can accurately capture the subtle deformation of muscle groups, and the reconstruction process must be real-time and robust, which constitutes a technical bottleneck that needs to be solved in the field of high-precision reconstruction of three-dimensional models of moving persons. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a high-precision reconstruction method and system for a three-dimensional model of a moving person, which overcomes the defects of relying on marker points, being easily affected by environmental interference, model topological inconsistency, missing details, and difficulty in balancing computational efficiency and precision in the prior art.

[0006] The technical scheme of the present application is to construct a high-precision reconstruction system for a three-dimensional model of a moving person, which includes a multi-modal data acquisition module, a time sequence feature fusion module, an implicit dynamic field construction module, a topological consistency constraint module, and a high-fidelity model generation module. The multi-modal data acquisition module is used to synchronously acquire depth image sequences obtained by at least 8 depth camera arrays and inertial measurement unit data streams. The time sequence feature fusion module is connected to the multi-modal data acquisition module and is used to perform space-time alignment on the depth image sequences and extract multi-scale space-time features, while converting the inertial measurement unit data into skeletal joint rotation parameters, and then realizing deep fusion of visual features and inertial features through a feature cross attention mechanism to output enhanced time sequence feature representations. The implicit dynamic field construction module is connected to the time sequence feature fusion module and is used to input the enhanced time sequence feature representations as conditions to construct a continuous implicit function with space-time coordinates as input, which directly outputs the signed distance value of any point in space at the current time and the non-rigid motion deformation amount. The topological consistency constraint module interacts with the implicit dynamic field construction module to introduce physically-based elastic potential energy constraints and kinematic smoothing constraints in the optimization process of the implicit dynamic field to ensure that the model surface deformation between adjacent frames conforms to the biomechanical properties and maintains time continuity. The high-fidelity model generation module is connected to the implicit dynamic field construction module and is used to extract isosurfaces of the implicit dynamic field at each time to generate a triangular mesh model with sub-millimeter level geometric details, and correct the model vertices frame by frame according to the non-rigid motion deformation amount, and finally output a high-precision three-dimensional human body model sequence in time sequence.

[0007] Further, the depth camera array in the multi-modal data acquisition module is arranged in a ring shape, the field of view angle overlap degree of adjacent cameras is not less than 60%, the sampling frequency is 90 Hz, and the depth image resolution is not less than 1280*720. The inertial measurement unit includes a 9-axis sensor fixed to the main limb segments of the human body, and the data is synchronized with the depth image sequence through a wireless transmission protocol at a hardware level, and the synchronization error is less than 1 ms.

[0008] Further, the specific operation process of the time sequence feature fusion module is as follows. First, the depth image sequence is preprocessed by denoising and background removal. Then, a 3D convolutional neural network is used to extract multi-scale space-time features from the preprocessed depth image sequence. In parallel, the inertial measurement unit data is calibrated and the attitude is solved to obtain a skeletal joint rotation sequence represented by a quaternion. Then, a feature cross attention mechanism is designed, which takes visual space-time features as query vectors, and inertial features as key and value vectors, realizes information fusion by calculating attention weights, and outputs enhanced time sequence features. The enhanced time sequence features encode both apparent geometric changes and internal motion posture information.

[0009] Further, the implicit dynamic field construction module employs a multi-layer perceptron network to realize the implicit function. The input of the multi-layer perceptron network is a 4-dimensional vector composed of 3-dimensional spatial coordinates and a time stamp, and the network parameters are dynamically modulated by the enhanced temporal features output by the temporal feature fusion module. The output of the network includes a scalar signed distance value and a 3-dimensional non-rigid motion deformation vector. The signed distance value represents the shortest signed distance from the point to the surface of the human body, and the non-rigid motion deformation vector represents the displacement of the point from the canonical space to the observation space at the current time. By querying the signed distance value of the entire spatial grid point and applying the marching cubes algorithm, the three-dimensional model surface at the current time can be extracted.

[0010] Further, the topological consistency constraint module imposes constraints including elastic potential energy constraints and kinematic smoothness constraints. The elastic potential energy constraint is realized by calculating the deformation gradient of adjacent vertices on the model surface between consecutive frames and punishing the degree of deviation from rigid transformation, aiming to maintain the shape stability of the local area. The kinematic smoothness constraint is realized by constructing an auxiliary motion model based on skinning and minimizing the difference between the vertex positions predicted by the implicit field and the positions calculated by the skinning, aiming to ensure the smoothness and reasonableness of the overall motion trajectory. These two constraints are added to the training loss function of the implicit dynamic field construction module as regularization terms.

[0011] Further, the high-fidelity model generation module adopts a hierarchical strategy in the process of isosurface extraction. First, coarse sampling is performed in the global space to locate the approximate area of the model, and then adaptive dense sampling is performed near the model surface, with a minimum sampling interval of 0.5 mm. The signed distance field calculated by the dense sampling is applied to the marching cubes algorithm to generate an initial grid. Finally, the non-rigid motion deformation vector predicted by the implicit dynamic field construction module is used to fine-tune the position of each vertex of the initial grid to compensate for the reconstruction error caused by fast motion or occlusion, thereby outputting the final high-fidelity model.

[0012] Further, the system also includes an online adaptive learning engine. The engine continuously monitors the residual error between the reconstructed model and the real-time acquisition data. When the residual error exceeds the preset threshold of 2 mm, the engine triggers fine-tuning of the multi-layer perceptron network parameters in the implicit dynamic field construction module. The fine-tuning process only uses the latest 32 frames of data to ensure that the system can quickly adapt to slow changes in the body shape or motion pattern of the moving person.

[0013] Further, the entire system is deployed on a parallel computing architecture. The multi-modal data acquisition and temporal feature fusion are performed on the central processing unit, and the implicit dynamic field query and model generation are performed on the graphics processing unit in parallel. Through pipeline design, the overall processing delay of the system is less than 33 milliseconds, meeting the needs of real-time interactive applications.

[0014] The beneficial effects of the present application are:

[0015] 1、The present application effectively overcomes the weakness of single visual data being easily affected by light and occlusion through multi-modal data acquisition and deep fusion. The internal motion parameters provided by the inertial measurement unit and the depth visual information form a complement, and precise fusion is realized through a feature cross attention mechanism, which significantly improves the pose estimation accuracy and robustness in fast motion and self-occlusion conditions, thereby ensuring high precision of the reconstructed model from the data source.

[0016] 2、The present application uses implicit dynamic field to represent three-dimensional human body, and converts the dynamic three-dimensional reconstruction problem into learning a continuous spatio-temporal function. This representation method has the characteristics of infinite resolution, which can naturally model complex surface details and topological changes, fundamentally avoiding the difficulty of maintaining topological consistency in dynamic sequences in traditional grid or point cloud methods, ensuring the high coherence and geometric integrity of the generated model in time sequence.

[0017] 3、The topological consistency constraint module introduced in the present application embeds biomechanical priors and kinematic smoothness into the model optimization process in the form of computable constraints. The elastic potential energy constraint ensures the physical reasonableness of local deformation of the model, and the kinematic smoothness constraint suppresses the generation of motion artifacts. This physical constraint mechanism enables the system to output three-dimensional models that are more consistent with the laws of real human motion while pursuing high geometric accuracy, and is particularly suitable for fields such as sports science analysis that require strict action details.

[0018] 4、The present application realizes the unification of high precision and high efficiency through the coordinated design of implicit field representation and parallel computing architecture. The combination of efficient query of implicit field and parallel computing capability of graphics processing unit enables the system to realize a reconstruction speed of 30 frames per second at sub-millimeter level precision, successfully breaking through the technical bottleneck of mutual restriction between precision and computational overhead in traditional methods, and providing a feasible technical path for real-time high-precision motion analysis. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is the overall technical scheme architecture diagram of the high-precision reconstruction system of the three-dimensional model of the moving personnel proposed by the present application;

[0020] Figure 2 is the core principle framework diagram of the time sequence feature fusion and implicit dynamic field construction in the present application; DETAILED DESCRIPTION

[0021] The present application will be further described below in conjunction with the drawings and specific embodiments, and the illustrative embodiments and descriptions of the present application are used to explain the present application, but not as a limitation of the present application.

[0022] Example 1

[0023] Reference is made to the accompanying drawings Figure 1 The embodiment details a specific implementation of a high-precision reconstruction system for a moving person's three-dimensional model. The system aims to achieve high-precision, real-time reconstruction of a moving person's dynamic three-dimensional model through multi-modal data fusion and advanced artificial intelligence technology. The core components of the system include a multi-modal data acquisition module, a time-series feature fusion module, an implicit dynamic field construction module, a topological consistency constraint module, and a high-fidelity model generation module. These modules work together to form a complete processing chain from data acquisition to model generation.

[0024] The multi-modal data acquisition module is the data input end of the entire system, responsible for synchronously capturing the depth visual information and inertial motion data of the moving person. This module includes a ring-shaped array composed of at least 8 depth cameras. These depth cameras are uniformly distributed in a ring shape in space, ensuring that the field of view angle overlap of adjacent cameras is not less than 60%, thereby achieving 360-degree coverage of the moving person's entire body. The sampling frequency of the depth camera array is fixed at 90 Hz, and the resolution of each depth image is not less than 1280x720 pixels to ensure that the acquired geometric information has sufficient temporal and spatial resolution. The depth camera usually integrates an infrared laser projector and an infrared camera inside, calculating the distance from each point in the scene to the camera through time-of-flight or structured light methods to generate depth images. In addition to the depth camera array, this module also includes a set of inertial measurement units. These inertial measurement units are 9-axis sensors that integrate three-axis accelerometers, three-axis gyroscopes, and three-axis magnetometers. They are firmly bound to the main limb segments of the moving person, such as the thighs, calves, upper arms, forearms, and torso. The inertial measurement units send the raw acceleration, angular velocity, and magnetic field strength data streams collected in real time to the central processing unit through low-power Bluetooth or other wireless transmission protocols. To achieve accurate correspondence between multi-modal data, the system uses a hardware-level synchronization mechanism. Specifically, a global synchronization signal generator sends synchronization pulses to all depth cameras and inertial measurement units simultaneously, ensuring that each frame of depth image and each set of inertial measurement unit data are strictly aligned in timestamp, with a synchronization error controlled within 1 millisecond. The raw depth image sequence and inertial measurement unit data stream collected will be temporarily stored in the cache memory for subsequent module processing.

[0025] The time sequence feature fusion module is connected to the multi-modal data acquisition module, and its core task is to preprocess, extract features and deeply fuse the original data from different modalities. The module first processes the depth image sequence. Because the depth camera is susceptible to environmental infrared interference or surface material, the original depth image may have noise or invalid points. Therefore, the preprocessing stage includes time domain median filtering and spatial domain bilateral filtering to suppress noise, while using background difference method to remove static scenes and preliminarily separate the moving person area. The preprocessed depth image sequence is sent to a pre-trained 3D convolutional neural network for multi-scale spatio-temporal feature extraction. The 3D convolutional neural network adopts an encoder structure, which gradually reduces the sampling in the time and space dimensions through consecutive 3D convolutional layers, 3D pooling layers and activation functions, while increasing the number of feature channels, thereby extracting spatio-temporal feature maps containing different levels of abstraction. These feature maps encode the continuous change information of the apparent geometry of the moving person. On the parallel path, the time sequence feature fusion module processes the inertial measurement unit data stream. The original inertial measurement unit data is first calibrated for sensor errors, including gyro zero offset compensation, accelerometer scale factor correction, and magnetic compass hard and soft iron interference compensation. The calibrated data is calculated through attitude solving algorithms such as complementary filtering or gradient descent method to obtain the rotation parameters of the main skeletal joints of the human body in the global coordinate system with the pelvis as the origin, which are represented in the form of quaternions to form the skeletal joint rotation sequence. In order to effectively fuse visual features and inertial features, the system designs and implements a feature cross-attention mechanism. In this mechanism, the visual spatio-temporal features extracted from the 3D convolutional neural network are reshaped into a series of feature vectors as query vectors in the attention mechanism. Correspondingly, the skeletal joint rotation sequence is encoded through a fully connected network to serve as the key vector and value vector. The calculation process of the attention weight is as follows:

[0026]

[0027] where Q represents the query vector, i.e. the visual spatio-temporal feature, K represents the key vector, i.e. the encoded inertial feature, V represents the value vector, which is also the encoded inertial feature, and d_k is the dimension of the key vector. This formula describes the process of calculating the similarity between the query and the key, normalizing the weight, and then weighting and summing the value vector. Its physical meaning is to dynamically focus on the internal motion posture information that is most relevant to the current visual appearance. For example, when the visual information is blurred or self-occluded due to rapid limb swinging, the attention mechanism will give higher weight to the inertial feature, thereby using reliable internal motion data to supplement or even correct the visual estimation. The fused feature vector, i.e. the enhanced time sequence feature representation, encodes both external geometric deformation and internal skeletal motion information, providing a more robust and richer spatio-temporal context for the subsequent modules.

[0028] The implicit dynamic field construction module is connected to the temporal feature fusion module, and its function is to convert the enhanced temporal feature representation described above into a continuous, dynamic 3-dimensional geometric representation. The core of this module is a parameterized multi-layer perception network that learns a mapping function from a 4-dimensional space-time coordinate to a signed distance value and a non-rigid motion deformation amount. Specifically, the input of the multi-layer perception network is a 4-dimensional vector composed of 3-dimensional spatial coordinates x, y, z and 1-dimensional time stamp t. The parameters of the network are not fixed, but are dynamically modulated by the enhanced temporal features output by the temporal feature fusion module. This modulation is usually achieved through techniques such as feature conditioning layers or adaptive instance normalization, allowing the network to flexibly change its internal function form according to different time and motion states. The multi-layer perception network usually contains more than 8 fully connected layers, each followed by a non-linear activation function such as ReLU or a sine function. The final output layer of the network is divided into two branches: one branch outputs a scalar, i.e. the signed distance value; the other branch outputs a 3-dimensional vector, i.e. the non-rigid motion deformation amount. The signed distance value defines the signed distance between the query point and the surface of the moving person model, a positive value indicates that the point is outside the model, a negative value indicates that the point is inside the model, and a zero value is exactly on the surface. The non-rigid motion deformation vector represents the displacement amount required to move the point from a standard pose in the canonical space coordinate to the observed space coordinate at the current time. Therefore, the implicit dynamic field actually defines a time-evolving, continuous 3-dimensional shape and motion field. In order to obtain an explicit 3-dimensional model from this implicit field, it is necessary to densely sample the spatial region of interest and query the signed distance value of each sampling point, thereby forming a discrete 3-dimensional signed distance field.

[0029] The topology consistency constraint module interacts closely with the implicit dynamic field construction module. It plays the role of a regularizer in the training and optimization process of the implicit dynamic field, ensuring that the generated dynamic model not only has high geometric accuracy, but also conforms to physical laws and motion continuity. This module imposes two main constraints: elastic potential energy constraint and kinematic smoothing constraint. The elastic potential energy constraint targets the local deformation characteristics of the model surface. It first defines a set of adjacent vertex pairs on the model surface. For each pair of vertices, it calculates the deformation gradient between the two consecutive frames. The goal of the elastic potential energy constraint is to penalize deformation gradients that deviate from an isometric transformation (i.e., rigid motion). Its calculation is based on the Green strain tensor, aiming to encourage local surface regions to maintain their original shape as much as possible during deformation, avoiding unnatural stretching or shearing. The kinematic smoothing constraint imposes constraints from the perspective of the global motion trajectory. It constructs a simplified skeletal skinning model defined by a set of skeletal joints and their hierarchical relationships, with joint motion driven by rotation parameters calculated from inertial measurement units. Through a linear blend skinning algorithm, the expected position of each surface vertex under the skeletal model can be calculated. The kinematic smoothing constraint then calculates the sum of the squared Euclidean distances between the vertex positions predicted by the implicit dynamic field and the expected positions calculated by the skeletal skinning, and takes it as a loss term. This constraint ensures that the overall motion predicted by the implicit field is consistent with the skeletal motion trend based on physical sensors, effectively filtering out high-frequency motion jitter or noise. During the system training phase, these two constraints are added to the total loss function of the implicit dynamic field construction module in the form of a weighted sum, together with the regression loss of the signed distance value, to guide the parameter learning of the multi-layer perception network.

[0030] The high-fidelity model generation module is connected to the implicit dynamic field construction module, and is responsible for converting the implicit dynamic field representation into an explicit 3D triangular mesh model sequence that can be displayed and used. The core of this module is the isosurface extraction algorithm, which usually adopts the marching cubes algorithm. In order to improve efficiency, the extraction process adopts a hierarchical strategy. First, coarse uniform sampling is performed in the global bounding box space, with a sampling interval of about 5 mm. By querying the implicit field, the region where the signed distance value crosses zero, i.e. the approximate location of the model surface, is quickly located. Then, adaptive dense sampling is started in the vicinity of these surfaces, and the sampling interval is dynamically adjusted according to the gradient of the signed distance value, with a minimum interval of 0.5 mm, in order to capture sub-millimeter-level geometric details such as fingertips and wrinkles. The dense 3D signed distance field obtained by dense sampling is applied to the marching cubes algorithm to generate an initial high-resolution triangular mesh model. However, due to fast motion or instantaneous occlusion, the model reconstructed based on the current frame information alone may still have minor errors. Therefore, this module further utilizes the non-rigid motion deformation vectors predicted by the implicit dynamic field construction module. For each vertex on the initial mesh, its spatial coordinates are mapped back to the canonical space, and then combined with the predicted non-rigid motion deformation vector of the vertex to correct its position. This step is equivalent to smoothing and optimizing the single-frame reconstruction result using temporal and spatial continuity information, effectively compensating for the reconstruction errors caused by motion blur or partial occlusion. Finally, the output is a sequence of 3D human body triangular mesh models with high-fidelity geometric details, which are corrected frame by frame and temporally coherent.

[0031] The system also integrates an online adaptive learning engine, which runs as a background process continuously. It monitors the difference between the reconstructed model output by the high-fidelity model generation module and the original data collected by the multi-modal data acquisition module at the current time, i.e. the reconstruction residual. Specifically, it renders the reconstructed model into a depth map and compares it with the measured depth map. When the average residual calculated exceeds the preset threshold of 2 mm, the engine determines that the current model may deviate from the actual state of the moving person (such as changes in clothing fit due to sweating, or subtle changes in movement patterns). At this time, the engine automatically triggers a fine-tuning process. This process extracts the last 32 frames of multi-modal data, including depth images and inertial measurement unit data, as well as their corresponding enhanced temporal features, from the cache. Using these data, the multi-layer perceptron network parameters in the implicit dynamic field construction module are updated with small-scale and fast gradient descent. The learning rate of the fine-tuning process is set low, and only recent data is used, aiming to quickly adapt to slow changes without forgetting previously learned general knowledge, thereby maintaining the long-term reconstruction accuracy and adaptability of the system.

[0032] The whole system is built on a high-efficiency parallel computing architecture to guarantee its real-time performance. The driving of the multi-modal data acquisition module, the data reception, and the forward inference of the 3D convolutional neural network in the time sequence feature fusion module, the pose solving and attention fusion calculation, these tasks are usually deployed on multi-core central processing units to take advantage of their powerful general-purpose computing and logical control capabilities. The dense multi-layer perception network query in the implicit dynamic field construction module, the large-scale spatial sampling and marching cubes algorithm in the high-fidelity model generation module, these computationally intensive tasks are offloaded to graphics processing units. With thousands of computing cores, graphics processing units can perform parallel signed distance value queries and grid generation on hundreds of thousands of spatial points, greatly improving processing speed. The system uses a pipeline design to overlap the acquisition, feature fusion, implicit field query, model generation, and other stages of a frame of data. When the Nth frame is being generated, the N+1th frame of data is being fused, and the N+2th frame of data is being acquired. This design strictly controls the end-to-end processing delay from data acquisition to the final model output to be within 33 milliseconds, enabling real-time high-precision three-dimensional reconstruction at 30 frames per second, meeting the stringent requirements of sports training, virtual reality interaction, and other application scenarios.

[0033] In summary, the embodiment discloses a specific implementation scheme of the high-precision three-dimensional model reconstruction system of the moving person by describing the internal working mechanism, data flow, parameter setting, and interaction logic of each module in detail. Through a series of technical innovations such as multi-modal depth fusion, implicit dynamic field representation, physical constraint optimization, and parallel computing acceleration, the system successfully generates a high-precision three-dimensional human model that is time-sequential, detailed, and consistent with the laws of biomechanics in complex motion scenarios.

[0034] The technical scheme of the present application is not limited to the above specific embodiments, and any technical modification made according to the technical scheme of the present application falls within the scope of protection of the present application.

Claims

1. A high-precision reconstruction system for a three-dimensional model of an athlete, characterized in that, include: A multimodal data acquisition module is used to simultaneously acquire depth image sequences obtained from at least eight depth camera arrays and inertial measurement unit data streams; The temporal feature fusion module is connected to the multimodal data acquisition module. It is used to perform spatiotemporal alignment on depth image sequences and extract multi-scale spatiotemporal features. At the same time, it converts inertial measurement unit data into skeletal joint rotation parameters. Then, through the feature cross-attention mechanism, it achieves deep fusion of visual features and inertial features and outputs an enhanced temporal feature representation. The implicit dynamic field construction module, connected to the temporal feature fusion module, is used to construct a continuous implicit function with spatiotemporal coordinates as input by taking the enhanced temporal feature representation as a conditional input. This function outputs the symbolic distance value of any point in space at the current time and the amount of non-rigid motion deformation. The topology consistency constraint module interacts with the implicit dynamic field construction module to introduce physics-based elastic potential energy constraints and kinematic smoothness constraints during the optimization process of the implicit dynamic field, ensuring that the surface deformation of the model between adjacent frames conforms to biomechanical characteristics and maintains temporal continuity. The high-fidelity model generation module is connected to the implicit dynamic field construction module. It is used to extract the isosurface of the implicit dynamic field at each time step, generate a triangular mesh model with sub-millimeter geometric details, and perform frame-by-frame correction of the model vertices based on the non-rigid motion deformation, outputting a high-precision three-dimensional human body model sequence with coherent time sequence.

2. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The depth camera array in the multimodal data acquisition module is arranged in a ring, with the overlap of the field of view of adjacent cameras not less than 60%, the sampling frequency is 90 Hz, and the depth image resolution is not less than 1280 x 720. The inertial measurement unit includes a 9-axis sensor, which is fixed to the main limbs of the human body. Its data is synchronized with the depth image sequence at the hardware level through a wireless transmission protocol, with a synchronization error of less than 1 millisecond.

3. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The specific operation method of the temporal feature fusion module is as follows: First, the depth image sequence is preprocessed by denoising and background removal; then, a 3D convolutional neural network is used to extract multi-scale spatiotemporal features from the preprocessed depth image sequence; in parallel, the inertial measurement unit data is calibrated and attitude is calculated to obtain the skeletal joint rotation sequence represented by quaternions; then, a feature cross-attention mechanism is designed, which uses visual spatiotemporal features as query vectors and inertial features as key and value vectors, and realizes information fusion by calculating attention weights, and outputs the fused enhanced temporal features.

4. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The implicit dynamic field construction module uses a multilayer perceptron network to implement implicit functions. The input of the multilayer perceptron network is a 4-dimensional vector composed of 3-dimensional spatial coordinates and timestamps. The network parameters are dynamically modulated by the enhanced temporal features output by the temporal feature fusion module. The output of the network includes a scalar symbolic distance value and a 3-dimensional non-rigid motion deformation vector. The symbolic distance value represents the shortest symbolic distance from the point to the human body surface, and the non-rigid motion deformation vector represents the displacement of the point from the norm space to the observation space at the current moment.

5. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The constraints imposed by the topology consistency constraint module include elastic potential energy constraints and kinematic smoothness constraints. The elastic potential energy constraint is implemented by calculating the deformation gradient of adjacent vertices on the model surface between consecutive frames and penalizing the degree to which they deviate from the rigid transformation. The kinematic smoothness constraint is implemented by constructing an auxiliary motion model based on skeleton skin and minimizing the difference between the vertex position predicted by the implicit field and the position calculated by the skeleton skin. These two constraints are added to the training loss function of the implicit dynamic field construction module in the form of regularization terms.

6. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The high-fidelity model generation module employs a layered strategy for isosurface extraction: first, coarse-grained sampling is performed in the global space to locate the approximate model region; then, adaptive densification sampling is performed near the model surface, with a minimum sampling interval of 0.5 mm; the moving cube algorithm is applied to the symbolic distance field calculated from the densification sampling to generate the initial mesh. Finally, the non-rigid motion deformation vectors predicted by the implicit dynamic field construction module are used to finely adjust the position of each vertex of the initial mesh.

7. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, It also includes an online adaptive learning engine; the online adaptive learning engine continuously monitors the residual between the reconstructed model and the real-time acquired data; When the residual exceeds the preset threshold of 2 mm, the engine triggers a fine-tuning of the parameters of the multilayer perceptron network in the implicit dynamic field construction module. The fine-tuning process uses only the most recent 32 frames of data.

8. The high-precision reconstruction system for a three-dimensional model of an athlete according to claim 1, characterized in that, The entire system is deployed on a parallel computing architecture; multimodal data acquisition and time-series feature fusion are executed on the central processing unit, while implicit dynamic field query and model generation are performed in parallel on the graphics processing unit. Through pipeline design, the overall system processing latency is less than 33 milliseconds.

9. A high-precision reconstruction system for a three-dimensional model of an athlete according to claim 3, characterized in that, The calculation method of the feature cross-attention mechanism is as follows: visual spatiotemporal features are used as query vectors, and encoded inertial features are used as key vectors and value vectors; attention weights are obtained by calculating and normalizing the similarity between the query vector and the key vector, and then the value vectors are weighted and summed to achieve deep fusion of visual features and inertial features.

10. A high-precision reconstruction system for a three-dimensional model of an athlete according to claim 4, characterized in that, The implicit dynamic field construction module contains a multilayer perceptron network with more than 8 fully connected layers, each followed by a nonlinear activation function; the network parameters are dynamically modulated by enhanced temporal features through feature condition layers or adaptive instance normalization techniques.

Images