Athlete trajectory and posture ai fusion analysis method and system

Abstract

The present application relates to the technical field of athlete trajectory and posture analysis, and particularly relates to an AI fusion analysis method and system for athlete trajectory and posture, which encodes and models global trajectory coordinate sequences, local posture joint node coordinate sequences and motion speed vector field data to obtain features and map them to a unified Riemann manifold space, calculates the correlation measure of the trajectory and the posture, and constructs cross-modal fusion features by using dynamic consistency correction, establishes a trajectory-driven posture evolution causal chain based on the features, performs topological analysis on the causal chain to extract abnormal nodes, generates fusion analysis results by using causal interpolation compensation of causal paths unaffected by the abnormality, so as to improve the accuracy and reliability of trajectory and posture fusion analysis.

Description

Technical Field

[0001] This invention relates to the field of athlete trajectory and posture analysis technology, and in particular to an AI fusion analysis method and system for athlete trajectory and posture. Background Technology

[0002] In the field of sports video analytics, the joint analysis of athlete trajectories and human postures is a key technology for motion recognition, technical evaluation, and anomaly detection. Current conventional approaches typically treat trajectory data (such as global position coordinate sequences) and posture data (such as joint coordinate sequences) as independent information streams, employing recurrent neural networks, 3D convolutional networks, or graph convolutional networks for feature extraction respectively. Trajectory analysis focuses on the spatial path and temporal changes of global motion, while posture analysis focuses on the spatial geometric relationships and dynamic evolution between local joints. In the fusion stage, a common method is to simply concatenate or weightedly sum trajectory and posture features, followed by deep fusion through fully connected layers or attention mechanisms. Furthermore, some studies have employed multi-task learning frameworks to simultaneously optimize trajectory prediction and posture classification, but feature interaction remains at the high-level semantic level. These methods perform reasonably well in regular movements such as running and jumping, but have significant limitations in complex and dynamic competitive scenarios.

[0003] The fundamental flaw in existing fusion methods lies in the assumption that both trajectory and posture feature extraction and fusion are based on data existing in Euclidean space. However, an athlete's trajectory should essentially be viewed as a spatiotemporal path on a high-dimensional manifold, and posture keypoints are embedded in specific Lie groups or symmetric structures. Linearly combining these two in Euclidean space ignores their inherent nonlinear geometric relationships, such as the dynamic coupling between the trajectory's tangent direction and the posture's orientation. This geometric mismatch leads to fused features that fail to accurately reflect the athlete's actual physical constraints during movement. For example, when an athlete makes a sharp turn, the shift in the posture's center of gravity and the change in trajectory curvature should be synchronized, but Euclidean fusion weakens or distorts this inherent synergy.

[0004] Another significant drawback is the lack of explicit modeling of the causal dependencies between trajectory and pose in existing methods. Most fusion frameworks focus on feature matching of synchronized frames, neglecting the driving effect of the trajectory's preceding state on the current pose state, and the reverse influence of pose changes on the subsequent evolution of the trajectory. When athletes perform complex movements (such as cushioning after a jump landing), there is a clear temporal causal chain between the continuity of the trajectory and the adjustment of the pose. When abnormal events occur (such as loss of balance or a fall), existing methods cannot distinguish which trajectory-pose associations are normal causal relationships and which are abnormal perturbations, resulting in the fusion results being contaminated by abnormal information and making it difficult to generate robust analytical conclusions. This lack of causal relationships also prevents the system from effectively compensating for features using undisturbed causal paths when facing occlusion or noise. Summary of the Invention

[0005] This invention provides an AI fusion analysis method and system for athlete trajectory and posture, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an AI fusion analysis method for athlete trajectory and posture, comprising:

[0007] Spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling are performed on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space.

[0008] In the Riemannian manifold space, the geodesic distance between the motion feature and the attitude feature is calculated as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is corrected for dynamic consistency, and a cross-modal fusion feature is constructed.

[0009] Based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data to obtain the trajectory-driven attitude evolution causal chain.

[0010] A topological structure analysis is performed on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude, and the corresponding set of abnormal nodes is extracted.

[0011] Based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using the causal paths in the causal chain that are not affected by the abnormality. Based on the compensated features, a fusion analysis result of trajectory and attitude is generated.

[0012] The athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data are subjected to spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space, including:

[0013] The global trajectory coordinate sequence is convolved using multi-scale spatiotemporal convolution kernels to extract local trajectory pattern features of the global trajectory coordinate sequence under different time windows and spatial scales. The motion features are then concatenated in the time and spatial dimensions.

[0014] By constructing a skeletal topology graph with each joint in the local pose joint coordinate sequence as nodes and the skeletal connection relationships between joints as edges, graph convolution propagation is performed on the skeletal topology graph to aggregate the coordinate information of each joint with the coordinate information of adjacent joints to obtain the topology-aware features of the joints, and global pooling is performed on the topology-aware features of all joints to obtain the pose features;

[0015] The motion velocity vector field data is subjected to time difference calculation and spatial curl calculation to obtain acceleration vector field and angular velocity components. The magnitude and direction angle of the acceleration vector field are calculated as translational dynamic parameters, and the magnitude and rotation axis direction of the angular velocity components are calculated as rotational dynamic parameters. The translational dynamic parameters and the rotational dynamic parameters are vector-concatenated to construct a dynamic state descriptor and encoded as the motion trend feature.

[0016] The motion features, posture features, and motion trend features are respectively mapped to the Riemannian manifold space through nonlinear transformation and assigned manifold coordinates.

[0017] In the Riemannian manifold space, the geodesic distance between the motion feature and the attitude feature is calculated as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is corrected for dynamic consistency, and cross-modal fusion features are constructed, including:

[0018] In the Riemannian manifold space, a path integral is performed along the geodesic connecting the manifold coordinates of the motion feature and the attitude feature. The Riemann arc length of the geodesic is calculated as the geodesic distance between the motion feature and the attitude feature, and the geodesic distance is used as a spatial correlation metric between the trajectory and the attitude.

[0019] The tangent space of the Riemannian manifold space is calculated at the manifold coordinates of the motion trend feature. The tangent vector corresponding to the motion trend feature is solved in the tangent space, and the directional component of the tangent vector is extracted as the dynamic evolution direction.

[0020] Based on the angle between the dynamic evolution direction and the tangent vector of the geodesic, the dynamic consistency deviation is calculated, and the spatial correlation metric is weighted and fused with the dynamic consistency deviation to obtain the corrected spatial correlation metric.

[0021] Using the manifold coordinates of the motion features and the manifold coordinates of the attitude features as row and column indices, and the modified spatial correlation metric as matrix elements, a manifold coupling matrix for trajectory and attitude is constructed.

[0022] The manifold coupling matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​that satisfy the preset cumulative contribution rate condition are selected as the cross-modal fusion features.

[0023] Based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data, resulting in a trajectory-driven attitude evolution causal chain including:

[0024] Modal coupling weights between trajectory and attitude are extracted from the cross-modal fusion features, and the preceding moments that drive the local attitude joint coordinate sequence are identified in the global trajectory coordinate sequence based on the modal coupling weights.

[0025] The global trajectory coordinates of the preceding time step are used as the preceding state node, and the joint coordinates of the current time step of the local attitude joint coordinate sequence are used as the current state node. A directed connection is established between the preceding state node and the current state node as a causal edge.

[0026] Based on the motion velocity vector field data, the momentum transfer between the previous time step and the current time step is extracted, and a momentum flow density field is constructed in the state space spanned by the previous state node and the current state node.

[0027] Calculate the flux integral value of the momentum flux density field along the causal edge direction, and determine the causal influence strength of the previous state node on the current state node as the edge weight of the causal edge based on the ratio between the flux integral value and the mass distribution of the previous state node;

[0028] Using the edge weights of the causal edges as the connection strengths of the edges, the preceding state nodes, the current state nodes, and the causal edges are organized into a directed graph structure, which serves as the trajectory-driven causal chain for attitude evolution.

[0029] Perform topological analysis on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude, including:

[0030] From the directed graph structure of the attitude evolution causal chain, the out-degree value of each trajectory state node is counted, and an out-degree distribution sequence of trajectory state nodes is constructed. The information entropy of the out-degree distribution sequence is calculated as the trajectory topological entropy. The in-degree value of each attitude state node is counted, and an in-degree distribution sequence of attitude state nodes is constructed. The information entropy of the in-degree distribution sequence is calculated as the attitude topological entropy.

[0031] The difference between the trajectory topological entropy and the attitude topological entropy is calculated as the topological entropy difference.

[0032] The momentum flow vectors on each causal edge in the momentum flow density field are reconstructed into a vector field in the topological space of the attitude evolution causal chain. The curl operator is calculated on the reconstructed vector field to obtain the momentum flow curl field. The circulation integral of the momentum flow curl field is calculated as the momentum conservation deviation.

[0033] The momentum conservation deviation is converted into a topological stability boundary quantity through a Lyapunov exponential mapping transformation. Based on the functional relationship between the topological stability boundary quantity and the geometric mean of the trajectory topological entropy and the attitude topological entropy, the allowable upper bound of the topological entropy difference is determined.

[0034] When the difference in topological entropy exceeds the allowable upper bound, it is determined that there is a topological inconsistency anomaly between the trajectory and attitude.

[0035] Based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using the causal paths in the causal chain that are not affected by the abnormalities. The fusion analysis results of the generated trajectory and pose based on the compensated features include:

[0036] In the attitude evolution causal chain, identify all incoming and outgoing edges of each abnormal node in the abnormal node set, mark the incoming and outgoing edges of the abnormal nodes as abnormal causal edges, mark the remaining causal edges as normal causal edges, and take the causal path composed only of the normal causal edges as the causal path unaffected by the abnormality;

[0037] The trajectory feature vectors and attitude feature vectors propagating on each normal causal edge are extracted from the causal path that is not affected by the anomaly, and used as normal feature samples;

[0038] Calculate the topological distance between the anomalous node and each node on the causal path unaffected by the anomalous node in the topological space of the attitude evolution causal chain, and construct an interpolation weight distribution based on the topological distance;

[0039] The normal feature samples corresponding to each node on the causal path that is not affected by the anomaly are weighted and combined according to the interpolation weight distribution to generate the compensation feature vector of the abnormal node. The initial feature vector of the abnormal node is replaced with the compensation feature vector to obtain the compensated attitude evolution causal chain.

[0040] The compensated trajectory feature sequence and the compensated posture feature sequence are extracted from the compensated posture evolution causal chain and spatiotemporally aligned and fused to generate the fusion analysis result of athlete trajectory and posture.

[0041] A second aspect of this invention provides an AI fusion analysis system for athlete trajectory and posture, comprising:

[0042] The Riemann coding unit is used to perform spatiotemporal convolutional coding, graph topology propagation, and dynamic modeling on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, and map them to a unified Riemannian manifold space.

[0043] A cross-modal fusion unit is used to calculate the geodesic distance between the motion feature and the attitude feature in the Riemannian manifold space as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is dynamically consistent and cross-modal fusion features are constructed.

[0044] A causal chain building unit is used to establish a causal edge between the previous state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence based on the cross-modal fusion features, and to quantify the strength of the causal edge according to the dynamic constraints of the motion velocity vector field data to obtain a trajectory-driven attitude evolution causal chain.

[0045] An anomaly detection unit is used to perform topological structure analysis on the attitude evolution causal chain, determine whether there is an anomaly of topological inconsistency between trajectory and attitude, and extract the corresponding set of anomaly nodes.

[0046] The causal compensation unit is used to perform causal interpolation compensation on abnormal features based on the abnormal node set and the causal paths in the causal chain that are not affected by the abnormality, and to generate a fusion analysis result of trajectory and attitude based on the compensated features.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] 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.

[0052] The beneficial effects of this application are as follows:

[0053] By mapping athlete trajectories and posture features uniformly to a Riemannian manifold space and using geodesic distance to accurately measure spatial correlation, combined with dynamic consistency correction based on the tangent vector direction of motion trends, cross-modal feature fusion becomes more consistent with the laws of motion physics. This significantly improves the spatiotemporal alignment accuracy of complex movements such as jumps and rotations, avoiding the geometric distortions caused by traditional linear fusion. By constructing a trajectory-driven causal chain of posture evolution and quantifying the dynamic constraints of the velocity vector field on posture changes with causal edge strength, interpretable path reasoning from global motion trends to local joint responses is achieved, effectively capturing the causal dependencies between movements.

[0054] Topological analysis of the causal chain of attitude evolution automatically identifies topological inconsistencies between trajectory and attitude, and accurately locates sets of anomalous nodes, resolving local mismatch issues in multimodal data under scenarios of drastic speed changes or occlusion. Causal interpolation compensation is applied to anomalous features using causal paths unaffected by anomalies, repairing missing or erroneous data while maintaining motion continuity. This avoids anomaly propagation contaminating subsequent analysis and significantly enhances the system's fault tolerance to incomplete input data.

[0055] This method can autonomously learn the normal manifold structure of motion data during training without relying on manually labeled anomalous samples, and it has a natural sensitivity to atypical movements or injury states. The fusion analysis results simultaneously preserve the macroscopic coherence of the trajectory and the microscopic details of the posture. The generated causal chain of posture evolution provides a physically verifiable reasoning basis for subsequent technical and tactical analysis and rehabilitation assessment, demonstrating significant application advantages in the fields of intelligent analysis of competitive sports, sports rehabilitation monitoring, and digital human animation generation. Attached Figure Description

[0056] Figure 1 A flowchart illustrating the AI ​​fusion analysis method for athlete trajectory and posture;

[0057] Figure 2 A schematic diagram of the process for constructing cross-modal fusion features for Riemannian manifold space. Detailed Implementation

[0058] 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.

[0059] 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.

[0060] Figure 1 This is a flowchart illustrating the AI ​​fusion analysis method for athlete trajectory and posture according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling are performed on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space.

[0062] In the Riemannian manifold space, the geodesic distance between the motion feature and the attitude feature is calculated as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is corrected for dynamic consistency, and a cross-modal fusion feature is constructed.

[0063] Based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data to obtain the trajectory-driven attitude evolution causal chain.

[0064] A topological structure analysis is performed on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude, and the corresponding set of abnormal nodes is extracted.

[0065] Based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using the causal paths in the causal chain that are not affected by the abnormality. Based on the compensated features, a fusion analysis result of trajectory and attitude is generated.

[0066] In one optional implementation, spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling are performed on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space, including:

[0067] The global trajectory coordinate sequence is convolved using multi-scale spatiotemporal convolution kernels to extract local trajectory pattern features of the global trajectory coordinate sequence under different time windows and spatial scales. The motion features are then concatenated in the time and spatial dimensions.

[0068] By constructing a skeletal topology graph with each joint in the local pose joint coordinate sequence as nodes and the skeletal connection relationships between joints as edges, graph convolution propagation is performed on the skeletal topology graph to aggregate the coordinate information of each joint with the coordinate information of adjacent joints to obtain the topology-aware features of the joints, and global pooling is performed on the topology-aware features of all joints to obtain the pose features;

[0069] The motion velocity vector field data is subjected to time difference calculation and spatial curl calculation to obtain acceleration vector field and angular velocity components. The magnitude and direction angle of the acceleration vector field are calculated as translational dynamic parameters, and the magnitude and rotation axis direction of the angular velocity components are calculated as rotational dynamic parameters. The translational dynamic parameters and the rotational dynamic parameters are vector-concatenated to construct a dynamic state descriptor and encoded as the motion trend feature.

[0070] The motion features, posture features, and motion trend features are respectively mapped to the Riemannian manifold space through nonlinear transformation and assigned manifold coordinates.

[0071] For the global trajectory coordinate sequence, multi-scale spatiotemporal convolutional kernels are used for convolution operations to capture the athlete's motion patterns at different temporal granularities and spatial ranges. A set of spatiotemporal convolutional layers with different kernel sizes is designed, each corresponding to a specific temporal window length and spatial receptive field range. For the temporal dimension, convolutional kernels with short windows (e.g., covering 3 consecutive frames), medium windows (e.g., covering 7 consecutive frames), and long windows (e.g., covering 15 consecutive frames) are set respectively, so that the extracted features can reflect both the instantaneous trajectory change details and capture the macroscopic motion trend over a longer period of time. For the spatial dimension, spatial convolutional kernels with different receptive fields are designed to extract local coordinate change patterns and coupling features across coordinate components, based on the spatial correlation between the components of the two-dimensional or three-dimensional trajectory coordinates. The output feature maps of each scale convolutional kernel are kept aligned in the temporal dimension, and then features are concatenated in the temporal and spatial dimensions, merging the local trajectory pattern features extracted at different scales along the channel direction to form a multi-scale fused motion feature representation. This multi-scale stitching strategy ensures that motion features have sufficient descriptive power for both rapid explosive movements and slow displacement trends in the trajectory sequence.

[0072] For the local pose joint coordinate sequence, a skeletal topology graph is constructed, with each joint as a node and the skeletal connections between joints as edges. The human skeletal structure has a fixed hierarchical topological relationship; for example, there is a thigh bone connection between the hip and knee joints, and a calf bone connection between the knee and ankle joints. The adjacency relationships of each joint are predefined by the human anatomical skeletal structure and encoded as an adjacency matrix. Graph convolution propagation is performed on the constructed skeletal topology graph. During propagation, each joint aggregates its own coordinate information as well as the coordinate information of all directly connected adjacent joints. The message passing mechanism of graph convolution ensures that the feature representation of each joint not only includes its own 3D coordinates but also incorporates the spatial context information of the surrounding skeletal neighborhood, thus obtaining topology-aware joint features, i.e., topology-aware features. After multi-layer graph convolution propagation, the topology-aware features of each joint have fully integrated the skeletal structure information within the multi-hop neighborhood. Global pooling is performed on the topology-aware features of all joints, averaging or maximizing the features across the node dimension to obtain a compact representation of the overall pose, i.e., pose features. Global pooling makes pose features robust to changes in the number of joints and preserves the overall semantic information of the whole body pose.

[0073] For the velocity vector field data, time difference calculation and spatial curl calculation are performed to extract dynamic information. Time difference calculation obtains the acceleration vector field by subtracting the velocity vector fields at adjacent time points, reflecting the rate of change of velocity over time and demonstrating the athlete's acceleration and deceleration dynamics. For each sampling point in the acceleration vector field, its magnitude is calculated. With direction angle ,in The magnitude of the acceleration is reflected in the translational dynamic intensity. The direction reflecting acceleration is the orientation of the athlete's translational tendency; together, they constitute the translational dynamics parameters. Spatial curl calculation involves applying a curl operator to the velocity vector field in the spatial domain to extract the angular velocity component, reflecting the athlete's rotational motion characteristics within a local spatial region. The magnitude of the angular velocity component is then calculated. relative to the direction of rotation axis ,in Intensity characterizing rotational rate The spatial orientation of the rotation axis, together with the rotational dynamics parameter, constitutes the rotational dynamics parameter. The translational dynamics parameter (including...) and ) and rotational dynamics parameters (including and The components of the vector are concatenated along the feature dimension to construct a dynamic state descriptor. This descriptor comprehensively characterizes the athlete's translational acceleration and rotational motion characteristics at the current moment, providing a complete encoding of the dynamic information of the motion velocity vector field. After undergoing nonlinear transformation through a multi-layer fully connected network, the dynamic state descriptor is encoded as a motion trend feature, which is used to characterize the direction of the athlete's motion dynamic trend in the subsequent Riemannian manifold space.

[0074] Motion features, attitude features, and motion trend features are mapped to a unified Riemannian manifold space through independent nonlinear transformation networks, and corresponding manifold coordinates are assigned to each type of feature. Compared to Euclidean space, the Riemannian manifold space can more naturally represent motion data with nonlinear structures. Its inherent curvature structure allows the geometric relationships between different modal features to be measured using geodesic distance rather than Euclidean distance, thus more accurately reflecting the essential spatial relationship between motion and attitude features. The nonlinear transformation is implemented by designing a fully connected mapping network for each type of feature, connected by multiple layers of nonlinear activation functions (such as ReLU or Tanh). The output dimension of the network is consistent with the embedding dimension of the Riemannian manifold. The parameters of the mapping network are optimized through end-to-end training, ensuring that the mapped manifold coordinates satisfy the local Euclidean constraint of the manifold. That is, near each point on the manifold, its tangent space can be approximated as Euclidean space, thus guaranteeing the mathematical rationality of subsequent geodesic distance calculations and tangent vector operations. After mapping motion features to the manifold, trajectory manifold coordinates are obtained; after mapping attitude features, attitude manifold coordinates are obtained; and after mapping motion trend features, dynamic manifold coordinates are obtained. All three types of manifold coordinates are located in the same Riemannian manifold space, providing a unified geometric computational basis for subsequent cross-modal fusion and dynamic consistency correction.

[0075] In practical applications, the three branches mentioned above—multi-scale spatiotemporal convolutional coding, graph topology propagation, and dynamic modeling—can be executed in parallel to reduce overall computational latency. For real-time scenarios in sports analysis, the number of scales in the spatiotemporal convolution kernels and the number of propagation layers in the graph convolution can be appropriately reduced to improve processing efficiency while ensuring the quality of feature representation. During the construction of the dynamic state descriptor, the step size of the temporal difference and the calculation window size of the spatial curl can be adaptively adjusted according to the movement frequency characteristics of the athlete's sport to ensure that the extraction accuracy of the acceleration vector field and angular velocity components matches the physical characteristics of the sport.

[0076] In one optional implementation, the geodesic distance between the motion feature and the attitude feature in the Riemannian manifold space is calculated as a spatial correlation metric between trajectory and attitude. The spatial correlation metric is then dynamically consistent based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, and cross-modal fusion features are constructed, including:

[0077] In the Riemannian manifold space, a path integral is performed along the geodesic connecting the manifold coordinates of the motion feature and the attitude feature. The Riemann arc length of the geodesic is calculated as the geodesic distance between the motion feature and the attitude feature, and the geodesic distance is used as a spatial correlation metric between the trajectory and the attitude.

[0078] The tangent space of the Riemannian manifold space is calculated at the manifold coordinates of the motion trend feature. The tangent vector corresponding to the motion trend feature is solved in the tangent space, and the directional component of the tangent vector is extracted as the dynamic evolution direction.

[0079] Based on the angle between the dynamic evolution direction and the tangent vector of the geodesic, the dynamic consistency deviation is calculated, and the spatial correlation metric is weighted and fused with the dynamic consistency deviation to obtain the corrected spatial correlation metric.

[0080] Using the manifold coordinates of the motion features and the manifold coordinates of the attitude features as row and column indices, and the modified spatial correlation metric as matrix elements, a manifold coupling matrix for trajectory and attitude is constructed.

[0081] The manifold coupling matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​that satisfy the preset cumulative contribution rate condition are selected as the cross-modal fusion features.

[0082] like Figure 2 As shown, the method includes:

[0083] In Riemannian manifold space, both motion and attitude features exist as point coordinates on the manifold. The distance between them cannot be simply measured by a straight-line distance in Euclidean space, but needs to be measured along the intrinsic geometry of the manifold itself. Specifically, along the manifold coordinates connecting the motion features... With attitude feature manifold coordinates The path integral is performed on the geodesic, which is the shortest curve between two points on a Riemannian manifold, and its arc length is given by the Riemannian metric tensor. Decision. Assume the ground wire parameterization is as follows: , ,satisfy , Then the Riemann arc length of the geodesic Calculated by the following formula:

[0084]

[0085] in For geodesics in parameters The tangent vector at that point, Let be the Riemannian metric tensor of the manifold at this point. In practical calculations, numerical iterative methods using exponential and logarithmic maps are typically employed to solve for the geodesic, and the result is obtained through discretized path integrals. The approximate value. The obtained This refers to the geodesic distance between motion features and attitude features. Using it as a spatial correlation metric between trajectory and attitude can accurately reflect the intrinsic differences between the two types of features under the manifold geometry, avoiding the problem of Euclidean distance failing in high curvature manifold regions.

[0086] After obtaining the spatial correlation metric, it is necessary to further incorporate the dynamic evolution information carried by the motion trend features to correct the aforementioned metric. This correction is made in the manifold coordinates of the motion trend features. At this point, calculate the tangent space of the Riemannian manifold. The tangent space is a linear approximation of the manifold at that point, and its dimension is equal to the intrinsic dimension of the manifold. In the tangent space, the corresponding tangent vector is solved by locally linearizing the characteristics of the motion trend. The specific method is to incorporate the characteristics of the movement trend into... Within the neighborhood of , a logarithmic mapping is used to project the vector onto the tangent space, and the directional component of the projection vector is taken as the direction of dynamic evolution. The direction of dynamic evolution reflects the local orientation of the motion trend on the manifold, representing the expected direction of change of the motion state in a short period of time.

[0087] If the evolution direction of the motion trend features is highly consistent with the direction of the geodesic tangent vector connecting the motion and attitude features, it indicates that the trajectory change trend and the attitude change trend corroborate each other in manifold geometry, and the reliability of the spatial correlation measure is high. Conversely, if the two directions deviate significantly, it indicates a dynamic inconsistency, requiring corresponding adjustments to the original spatial correlation measure. Therefore, the dynamic evolution direction is calculated. At the starting point of the geodesic Tangent vector at point The angle between : ,in For geodesics in The unit tangent vector at point A, through a logarithmic mapping Obtained through normalization. Angle The larger the value, the greater the deviation between the direction of dynamic evolution and the direction of trajectory-attitude correlation, and the worse the dynamic consistency. Substitute in consistency deviation The calculation, take When the directions are completely consistent When the directions are completely opposite This transforms angular deviation into a continuous scalar deviation measure. The corrected spatial correlation measure. The weighted fusion yielded: ,in The dynamic consistency correction weight is used to balance the relative contribution between geometric distance and dynamic deviation, and can be adaptively adjusted according to the dynamic characteristics of the specific sport. By combining information from both manifold geometric distance and dynamic direction deviation, it can more comprehensively characterize the coupling relationship between trajectory and attitude at the spatiotemporal dynamic level compared to geodesic distance alone.

[0088] A set of manifold coordinates with motion characteristics For row indexing, the set of manifold coordinates of the pose features. For column indexes, the corresponding corrected spatial correlation measure will be used. Fill in the matrix elements to construct the manifold coupling matrix between trajectory and attitude. ,in The number of samples for motion features. This represents the number of samples for the pose features. Manifold coupling matrix. The structured matrix format fully records the dynamic correction correlation strength between all motion and attitude features, providing a unified mathematical framework for subsequent cross-modal feature extraction. The smaller the element value in the matrix, the more consistent the corresponding motion and attitude features are in the manifold geometry and dynamic direction, and the tighter their cross-modal coupling.

[0089] manifold coupling matrix Perform eigenvalue decomposition. Because... Typically, it is a non-square matrix; in practice, it is used for... Perform eigenvalue decomposition to obtain the eigenvalue sequence. and the corresponding feature vectors Sort by eigenvalue from largest to smallest. Based on the preset cumulative contribution rate condition, select the top [values] that meet the condition. The eigenvectors corresponding to each eigenvalue. The cumulative contribution rate is defined as the top... The ratio of the sum of 1 eigenvalues ​​to the sum of all eigenvalues ​​is denoted as . :

[0090]

[0091] when When the preset threshold (e.g., 95%) is reached, the corresponding selection is made. eigenvectors These feature vectors serve as cross-modal fusion features. They capture the most important structural information in the manifold coupling matrix, representing the most significant coupling patterns between trajectory and attitude in the manifold space. The retained dimension is automatically determined based on the cumulative contribution rate condition. This approach avoids the loss of key features due to excessive information compression and suppresses noise interference carried by low-contribution feature vectors, resulting in cross-modal fusion features with good compactness and discriminativeness. The final cross-modal fusion features will serve as input for subsequent causal chain construction and anomaly detection, supporting in-depth analysis of the co-evolutionary patterns of athlete trajectory and posture.

[0092] In one optional implementation, based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data to obtain the trajectory-driven attitude evolution causal chain, including:

[0093] Modal coupling weights between trajectory and attitude are extracted from the cross-modal fusion features, and the preceding moments that drive the local attitude joint coordinate sequence are identified in the global trajectory coordinate sequence based on the modal coupling weights.

[0094] The global trajectory coordinates of the preceding time step are used as the preceding state node, and the joint coordinates of the current time step of the local attitude joint coordinate sequence are used as the current state node. A directed connection is established between the preceding state node and the current state node as a causal edge.

[0095] Based on the motion velocity vector field data, the momentum transfer between the previous time step and the current time step is extracted, and a momentum flow density field is constructed in the state space spanned by the previous state node and the current state node.

[0096] Calculate the flux integral value of the momentum flux density field along the causal edge direction, and determine the causal influence strength of the previous state node on the current state node as the edge weight of the causal edge based on the ratio between the flux integral value and the mass distribution of the previous state node;

[0097] Using the edge weights of the causal edges as the connection strengths of the edges, the preceding state nodes, the current state nodes, and the causal edges are organized into a directed graph structure, which serves as the trajectory-driven causal chain for attitude evolution.

[0098] When extracting modal coupling weights between trajectory and attitude from cross-modal fusion features, it is necessary to quantify the correlation between trajectory and attitude components at each time step in the fusion feature matrix. Specifically, the cross-modal fusion features are expanded along the time axis, and for each historical time step of the global trajectory coordinate sequence, the inner product response between the corresponding trajectory embedding vector and the current time step's attitude embedding vector is calculated to form a modal coupling weight sequence. Let the th... The trajectory embedding vector at each historical moment is The pose embedding vector at the current moment is Then the first Modal coupling weights at each historical moment satisfy ,in This represents the total number of historical moments. According to... The historical moments are sorted by their size, and the weights selected exceed a preset threshold. The historical time set is used as the preceding time set. These times are the moments in the global trajectory coordinate sequence that have a significant driving effect on the local attitude joint coordinate sequence. The threshold can be adjusted according to the dynamic characteristics of the sport. For example, for high-speed change-of-direction sports, the influence window of the preceding time is short, and the threshold should be relatively high to retain the most critical driving moments; while for continuous and smooth sports, the influence range of the preceding time is wide, and the threshold can be appropriately reduced.

[0099] After determining the set of preceding time steps, the global trajectory coordinates corresponding to each preceding time step are extracted as preceding state nodes, and the coordinates of each joint in the local attitude joint coordinate sequence at the current time step are extracted as current state nodes. The preceding state nodes carry information including their 3D spatial coordinates, timestamp, and corresponding modal coupling weights at that time. The current state nodes carry the 3D coordinates and joint topology numbers of each joint. A directed connection is established between each preceding state node and each current state node, with the direction of the connection pointing from the preceding state node to the current state node, representing temporal causality, i.e., the driving relationship between the historical trajectory state and the current attitude state. The establishment of directed connections follows temporal causality constraints, i.e., only preceding state nodes with timestamps earlier than the current time step are allowed to establish causal edges with the current state node, strictly excluding non-causal connections that affect the current state from future time steps, thus ensuring the temporal rationality of the causal chain.

[0100] When constructing the momentum flux density field in the state space spanned by the previous and current state nodes, velocity field sampling data within the time interval between the previous and current moments is extracted from the motion velocity vector field data. Let the mass distribution function of the moving body within this time interval be... The velocity vector field is Then the momentum flux density field Defined as ,in These are the spatial coordinates in the state space. The momentum flux density field is a time variable. It describes the direction and intensity of momentum transfer per unit volume of a moving body as it evolves from a previous state to the current state. In actual calculations, the velocity vector field data is obtained by finite difference estimation of tracking data of various parts of the athlete's body, while the mass distribution is assigned values ​​based on the standard mass ratio model of each body segment to ensure the accuracy of the physical meaning of the momentum flux density field.

[0101] When calculating the flux integral of the momentum flux density field along the causal edge direction, the causal edge is considered as a directed line segment in the state space connecting the previous state node and the current state node, and its unit direction vector is denoted as... , indicating from the first The preceding state node points to the first... The direction of the current state node. The integral value of momentum flux along the causal edge direction. It is obtained by performing a line integral on the momentum flux density field along the state space path traversed by the causal edge, i.e. ,in For the integral path corresponding to the causal edge, Let be the path element length. This integral value reflects the magnitude of the physical quantity by which the preceding trajectory state influences the current attitude state through momentum transfer. The larger the flux integral value, the stronger the dynamic driving force of the preceding state on the current state.

[0102] When determining the intensity of causal influence based on the ratio of the flux integral value to the mass distribution of the preceding state node, let the first... The total mass distribution corresponding to each preceding state node is Then the first The preceding state node pairs with the _th ... The causal influence strength of each current state node satisfy ,in To prevent small positive numbers from being divided by zero. The edge weight, representing the weight of a corresponding causal edge, physically signifies the strength of the causal influence exerted by a preceding state node per unit mass on the current state node through its momentum flux density field. A higher edge weight indicates a stronger causal relationship between the trajectory and attitude evolution represented by that edge, and this edge will be assigned a higher priority in subsequent topology analysis. The edge weights between all preceding state nodes and the current state node are normalized so that the sum of all causal edge weights received by each current state node is 1, thus ensuring the comparability of causal strength at the node level.

[0103] Using the edge weights of causal edges as the connection strength, all previous state nodes, current state nodes, and causal edges are organized into a directed graph structure. The node set in the directed graph consists of the union of the previous state node set and the current state node set, and the edge set consists of all causal edges that have established directed connections. Each edge carries a corresponding edge weight. Time difference information and modal coupling weights The topological structure of the directed graph reflects the driving relationship network of global trajectory evolution on local attitude joint changes. The trajectory state at different previous time steps affects the attitude of each joint at the current time step through causal edges of varying strengths, thus forming a causal chain of attitude evolution with temporal causal directionality and dynamic physical constraints. This directed graph structure is the trajectory-driven causal chain of attitude evolution, which can be used in subsequent topology analysis steps to determine whether there are topological inconsistencies between the trajectory and attitude, and to provide a structural basis for the location and compensation of anomalous nodes.

[0104] In one optional implementation, performing topological analysis on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude includes:

[0105] The out-degree value of each trajectory state node is counted from the directed graph structure of the attitude evolution causal chain, and an out-degree distribution sequence of trajectory state nodes is constructed. The information entropy of the out-degree distribution sequence is calculated as the trajectory topological entropy.

[0106] The in-degree value of each attitude state node is counted, and an in-degree distribution sequence of attitude state nodes is constructed. The information entropy of the in-degree distribution sequence is calculated as the attitude topological entropy.

[0107] The difference between the trajectory topological entropy and the attitude topological entropy is calculated as the topological entropy difference.

[0108] The momentum flow vectors on each causal edge in the momentum flow density field are reconstructed into a vector field in the topological space of the attitude evolution causal chain. The curl operator is calculated on the reconstructed vector field to obtain the momentum flow curl field. The circulation integral of the momentum flow curl field is calculated as the momentum conservation deviation.

[0109] The momentum conservation deviation is converted into a topological stability boundary quantity through a Lyapunov exponential mapping transformation. Based on the functional relationship between the topological stability boundary quantity and the geometric mean of the trajectory topological entropy and the attitude topological entropy, the allowable upper bound of the topological entropy difference is determined.

[0110] When the difference in topological entropy exceeds the allowable upper bound, it is determined that there is a topological inconsistency anomaly between the trajectory and attitude.

[0111] For the directed graph of the causal chain of attitude evolution, degree distribution information is extracted from two dimensions: trajectory state nodes and attitude state nodes. For each trajectory state node, the number of directed edges pointing to attitude state nodes in the directed graph is counted, i.e., the out-degree value. The out-degree values ​​of all trajectory state nodes are summarized to form an out-degree distribution sequence. Let there be a total of The trajectory status node, the first The out-degree value of each node is Then the probability of this node in the out-degree distribution is Based on this probability distribution, the information entropy of the degree distribution sequence is calculated, which serves as the trajectory topological entropy. Its calculation method is as follows The trajectory topological entropy reflects the uniformity of trajectory state nodes in causal propagation: if the out-degree distribution of trajectory nodes is highly concentrated, it indicates that a few trajectory states dominate the causal drive of attitude, and the entropy value is low; if the out-degree distribution is relatively uniform, the entropy value is high.

[0112] Correspondingly, for each attitude state node, the number of directed edges it receives from the trajectory state node is counted, i.e., the in-degree value. The in-degree values ​​of all attitude state nodes are then aggregated to form an in-degree distribution sequence. Let there be a total of The attitude state node, the first The in-degree value of each node is Then the probability of this node in the in-degree distribution is Similarly, the information entropy of the in-degree distribution sequence is calculated as the attitude topological entropy. Its calculation method is as follows Attitude topological entropy reflects the distributional balance of attitude state nodes when they are driven by causal trajectories.

[0113] The difference between the trajectory topological entropy and the attitude topological entropy is obtained as the topological entropy difference value. This difference directly measures the degree of information asymmetry between the trajectory-side causal output structure and the attitude-side causal receiver structure. When the two structures are highly consistent, Approaching zero; when there is a significant structural misalignment, It will increase significantly. Based solely on This alone is insufficient to determine an anomaly; it is also necessary to combine motion dynamics information to construct an adaptive upper bound for tolerance.

[0114] Therefore, the momentum flow density field The momentum flow vectors on each causal edge are extracted, and a vector field is reconstructed in the topological space of the attitude evolution causal chain to form a continuous vector field representation covering the entire causal graph topology. A curl operator is applied to the reconstructed vector field to obtain the momentum flow curl field. The non-zero region of the curl field corresponds to a local circulation tendency of momentum flow, meaning that the momentum transfer path undergoes a non-conservative deflection. Furthermore, a closed loop selected along the causal chain topological space... The circulation integral is calculated over the curl field of the momentum flow to obtain the momentum conservation deviation. ,Right now ,in is the line element vector along the loop. The larger the absolute value, the worse the conservation of momentum transfer in the causal chain, and the more unstable the topological structure.

[0115] Momentum conservation deviation Transformed into topological stability boundary quantities through Lyapunov exponential mapping transformation. The Lyapunov exponential mapping describes the sensitivity of a dynamical system to initial disturbances, and... Mapped to this framework, the stability margin of causal chain topology under dynamic perturbations can be quantified. Specifically, ,in The scaling coefficients of the Lyapunov mapping, This is a normalized reference value used to control the sensitivity of the mapping. The larger the value, the closer the current causal chain topology is to the unstable boundary in a dynamic sense.

[0116] Obtaining the boundary values ​​of topological stability Then, it is compared with the trajectory topological entropy. and attitude topological entropy geometric mean Establish functional relationships and determine the allowable upper bound of the topological entropy difference. . Specifically, ,in To provide a baseline scaling factor for the upper limit of allowable limits, This is the decay scale parameter for the stability boundary quantity. The physical meaning of this functional relationship is: the more unstable the dynamics (…), the more stable the motion becomes (…). When the value is larger, the upper bound of the allowable topological entropy difference is smaller, meaning the judgment on topological inconsistency is more stringent; conversely, when the dynamic stability is good, a certain degree of natural difference between the trajectory and attitude topology is allowed.

[0117] When the calculated topological entropy difference Exceeding the upper limit of tolerance At that time, that is The system identifies a topological inconsistency between trajectory and attitude in the current attitude evolution causal chain. Further analysis locates the specific node causing the anomaly and extracts the corresponding set of anomalous nodes, providing precise repair targets for subsequent causal interpolation compensation steps. Anomalous nodes are located by ranking them according to their contribution to the global topological entropy difference; nodes whose contribution exceeds a preset threshold are included in the anomalous node set, thus achieving fine-grained tracing of the source of topological inconsistency.

[0118] This judgment mechanism organically combines information theory metrics (topological entropy) with dynamic constraints (momentum conservation bias and Lyapunov stability). It can capture the asymmetry between trajectory and attitude at the structural level, and adaptively adjust the judgment threshold according to the dynamic characteristics of the motion itself. This avoids the misjudgment problem of fixed threshold methods in different motion scenarios and improves the robustness and accuracy of topological inconsistency detection.

[0119] In one optional implementation, based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using causal paths in the causal chain that are not affected by the abnormality. The fusion analysis results of the trajectory and pose generated based on the compensated features include:

[0120] In the attitude evolution causal chain, identify all incoming and outgoing edges of each abnormal node in the abnormal node set, mark the incoming and outgoing edges of the abnormal nodes as abnormal causal edges, mark the remaining causal edges as normal causal edges, and take the causal path composed only of the normal causal edges as the causal path unaffected by the abnormality;

[0121] The trajectory feature vectors and attitude feature vectors propagating on each normal causal edge are extracted from the causal path that is not affected by the anomaly, and used as normal feature samples;

[0122] Calculate the topological distance between the anomalous node and each node on the causal path unaffected by the anomalous node in the topological space of the attitude evolution causal chain, and construct an interpolation weight distribution based on the topological distance;

[0123] The normal feature samples corresponding to each node on the causal path that is not affected by the anomaly are weighted and combined according to the interpolation weight distribution to generate the compensation feature vector of the abnormal node. The initial feature vector of the abnormal node is replaced with the compensation feature vector to obtain the compensated attitude evolution causal chain.

[0124] The compensated trajectory feature sequence and the compensated posture feature sequence are extracted from the compensated posture evolution causal chain and spatiotemporally aligned and fused to generate the fusion analysis result of athlete trajectory and posture.

[0125] After obtaining the set of anomalous nodes, the features carried by these anomalous nodes need to be repaired and compensated to eliminate the interference of topological inconsistencies on the final fusion analysis results. The compensation process begins with the systematic labeling of the connection relationships of anomalous nodes in the attitude evolution causal chain. Specifically, for each anomalous node in the set, all incoming edges (i.e., directed edges ending at the anomalous node) and all outgoing edges (i.e., directed edges starting at the anomalous node) in the directed graph of the causal chain are retrieved, and these incoming and outgoing edges are uniformly marked as anomalous causal edges. All causal edges in the causal chain other than anomalous causal edges are marked as normal causal edges. Based on this, directed paths consisting only of normal causal edges connected end to end are defined as causal paths unaffected by anomalous events. These paths do not pass through any anomalous nodes, and the feature information propagated on them has dynamic consistency guarantees, serving as a reliable source of information for subsequent interpolation compensation.

[0126] Normal feature samples are extracted from causal paths unaffected by anomalies. Each normal causal edge along the path is traversed, and the trajectory and attitude feature vectors propagated by that edge are read. The trajectory feature vectors are derived from the embedded representations of the corresponding preceding state nodes after spatiotemporal convolutional encoding, while the attitude feature vectors are derived from the keypoint feature representations of the corresponding current state nodes after graph topology propagation. The trajectory and attitude feature vectors extracted from all nodes along all normal causal paths are aggregated to form a normal feature sample set. This set encompasses the complete feature distribution of the athlete's trajectory and attitude co-evolution in regions with normal topology, providing ample reference for compensation of anomalous nodes.

[0127] In the directed graph topology space of the attitude evolution causal chain, for each anomalous node ( ,in (For the set of anomalous nodes), calculate the relationship between them and each normal node on the causal path that is not affected by the anomalous nodes. ( ,in Topological distance between (the set of all normal nodes) Topological distance is measured using the shortest path length in a directed graph, i.e., from... arrive Or from arrive The minimum number of directed edge hops (the smaller of the two directions) is used. If there is no directed reachable path between two nodes, the upper bound of the graph's diameter is used as the distance replacement value. This is based on topological distance. The interpolation weight distribution is constructed according to the inverse distance principle. A normal node versus an abnormal node interpolation weights Defined as ,in To prevent the introduction of small positive numbers with a denominator of zero, the weights of all normal nodes are then normalized to obtain normalized interpolated weights. This ensures that the interpolation weight distribution for each anomalous node satisfies the probability normalization constraint. Normal nodes with smaller topological distances contribute more to the compensation of anomalous nodes, which aligns with the principle of locality of information propagation in causal chains.

[0128] After obtaining the interpolation weight distribution, normal feature samples are weighted and combined to generate a compensation feature vector. For anomalous nodes... Its compensation trajectory feature vector By examining all normal nodes Trajectory feature vector By normalized weights We obtain the result by weighted summation, i.e. Similarly, compensate for the pose feature vector ,in Normal node The pose feature vector. The generated compensation feature vector. and Replace abnormal nodes The original initial feature vectors are used to repair the features of the abnormal node. This applies to the set of abnormal nodes. After performing the above replacement operation on all anomalous nodes in sequence, a compensated attitude evolution causal chain is obtained. In the compensated causal chain, the original anomalous node positions have been filled by topologically adjacent normal features, and the dynamic consistency of the overall causal chain is restored.

[0129] After obtaining the compensated attitude evolution causal chain, the compensated trajectory feature sequence and the compensated attitude feature sequence are extracted and spatiotemporally aligned and fused to generate the final fusion analysis result. The trajectory feature sequence arranges the trajectory feature vectors corresponding to each node in timestamp order, and the attitude feature sequence arranges the attitude feature vectors corresponding to each node in the same timestamp order. The spatiotemporal alignment operation first eliminates the temporal offset between the trajectory feature sequence and the attitude feature sequence through a dynamic time warping method, ensuring that the correspondence between the two sequences on the time axis is accurate. After alignment, the trajectory feature vector and attitude feature vector at each time step are concatenated, and then mapped to the fusion feature space through a multilayer perceptron, outputting the fusion feature vector at that time step. The fusion feature vectors at all time steps are organized in chronological order to form a complete temporal fusion feature sequence.

[0130] Based on temporal fusion feature sequences, a fusion analysis of athlete trajectory and posture is generated. The fusion analysis results include outputs in multiple dimensions: at the motion pattern recognition level, the temporal fusion feature sequences are classified, outputting the athlete's current technical movement category and confidence level; at the abnormal behavior detection level, the difference in feature vectors before and after compensation is compared, quantifying the deviation of trajectory and posture at each abnormal node, and locating the time points where technical movement errors or sports injury risks occur; at the motion quality assessment level, the fusion feature sequences are compared with a standard movement template feature library for similarity, outputting scores and improvement suggestions for each technical aspect. The entire causal interpolation compensation process utilizes the feature information of healthy paths in the causal chain to perform topology-aware repair of damaged nodes, effectively avoiding contamination of the fusion analysis results by abnormal nodes, and ensuring the consistency and reliability of the final output in terms of dynamic constraints and topological structure.

[0131] A second aspect of this invention provides an AI fusion analysis system for athlete trajectory and posture, comprising:

[0132] The Riemann coding unit is used to perform spatiotemporal convolutional coding, graph topology propagation, and dynamic modeling on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, and map them to a unified Riemannian manifold space.

[0133] A cross-modal fusion unit is used to calculate the geodesic distance between the motion feature and the attitude feature in the Riemannian manifold space as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is dynamically consistent and cross-modal fusion features are constructed.

[0134] A causal chain building unit is used to establish a causal edge between the previous state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence based on the cross-modal fusion features, and to quantify the strength of the causal edge according to the dynamic constraints of the motion velocity vector field data to obtain a trajectory-driven attitude evolution causal chain.

[0135] An anomaly detection unit is used to perform topological structure analysis on the attitude evolution causal chain, determine whether there is an anomaly of topological inconsistency between trajectory and attitude, and extract the corresponding set of anomaly nodes.

[0136] The causal compensation unit is used to perform causal interpolation compensation on abnormal features based on the abnormal node set and the causal paths in the causal chain that are not affected by the abnormality, and to generate a fusion analysis result of trajectory and attitude based on the compensated features.

[0137] A third aspect of the present invention provides an electronic device, comprising:

[0138] processor;

[0139] Memory used to store processor-executable instructions;

[0140] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0141] 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.

[0142] 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.

[0143] 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. An AI fusion analysis method for athlete trajectory and posture, characterized in that, include: Spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling are performed on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space. In the Riemannian manifold space, the geodesic distance between the motion feature and the attitude feature is calculated as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is corrected for dynamic consistency, and a cross-modal fusion feature is constructed. Based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data to obtain the trajectory-driven attitude evolution causal chain. A topological structure analysis is performed on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude, and the corresponding set of abnormal nodes is extracted. Based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using the causal paths in the causal chain that are not affected by the abnormality. Based on the compensated features, a fusion analysis result of trajectory and attitude is generated.

2. The method according to claim 1, characterized in that, The athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data are subjected to spatiotemporal convolutional encoding, graph topology propagation, and dynamic modeling to obtain motion features, posture features, and motion trend features, which are then mapped to a unified Riemannian manifold space, including: The global trajectory coordinate sequence is convolved using multi-scale spatiotemporal convolution kernels to extract local trajectory pattern features of the global trajectory coordinate sequence under different time windows and spatial scales. The motion features are then concatenated in the time and spatial dimensions. By constructing a skeletal topology graph with each joint in the local pose joint coordinate sequence as nodes and the skeletal connection relationships between joints as edges, graph convolution propagation is performed on the skeletal topology graph to aggregate the coordinate information of each joint with the coordinate information of adjacent joints to obtain the topology-aware features of the joints, and global pooling is performed on the topology-aware features of all joints to obtain the pose features; The motion velocity vector field data is subjected to time difference calculation and spatial curl calculation to obtain acceleration vector field and angular velocity components. The magnitude and direction angle of the acceleration vector field are calculated as translational dynamic parameters, and the magnitude and rotation axis direction of the angular velocity components are calculated as rotational dynamic parameters. The translational dynamic parameters and the rotational dynamic parameters are vector-concatenated to construct a dynamic state descriptor and encoded as the motion trend feature. The motion features, posture features, and motion trend features are respectively mapped to the Riemannian manifold space through nonlinear transformation and assigned manifold coordinates.

3. The method according to claim 1, characterized in that, In the Riemannian manifold space, the geodesic distance between the motion feature and the attitude feature is calculated as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is corrected for dynamic consistency, and cross-modal fusion features are constructed, including: In the Riemannian manifold space, a path integral is performed along the geodesic connecting the manifold coordinates of the motion feature and the attitude feature. The Riemann arc length of the geodesic is calculated as the geodesic distance between the motion feature and the attitude feature, and the geodesic distance is used as a spatial correlation metric between the trajectory and the attitude. The tangent space of the Riemannian manifold space is calculated at the manifold coordinates of the motion trend feature. The tangent vector corresponding to the motion trend feature is solved in the tangent space, and the directional component of the tangent vector is extracted as the dynamic evolution direction. Based on the angle between the dynamic evolution direction and the tangent vector of the geodesic, the dynamic consistency deviation is calculated, and the spatial correlation metric is weighted and fused with the dynamic consistency deviation to obtain the corrected spatial correlation metric. Using the manifold coordinates of the motion features and the manifold coordinates of the attitude features as row and column indices, and the modified spatial correlation metric as matrix elements, a manifold coupling matrix for trajectory and attitude is constructed. The manifold coupling matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​that satisfy the preset cumulative contribution rate condition are selected as the cross-modal fusion features.

4. The method according to claim 1, characterized in that, Based on the cross-modal fusion features, a causal edge is established between the preceding state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence. The strength of the causal edge is quantified according to the dynamic constraints of the motion velocity vector field data, resulting in a trajectory-driven attitude evolution causal chain including: Modal coupling weights between trajectory and attitude are extracted from the cross-modal fusion features, and the preceding moments that drive the local attitude joint coordinate sequence are identified in the global trajectory coordinate sequence based on the modal coupling weights. The global trajectory coordinates of the preceding time step are used as the preceding state node, and the joint coordinates of the current time step of the local attitude joint coordinate sequence are used as the current state node. A directed connection is established between the preceding state node and the current state node as a causal edge. Based on the motion velocity vector field data, the momentum transfer between the previous time step and the current time step is extracted, and a momentum flow density field is constructed in the state space spanned by the previous state node and the current state node. Calculate the flux integral value of the momentum flux density field along the causal edge direction, and determine the causal influence strength of the previous state node on the current state node as the edge weight of the causal edge based on the ratio between the flux integral value and the mass distribution of the previous state node; Using the edge weights of the causal edges as the connection strengths of the edges, the preceding state nodes, the current state nodes, and the causal edges are organized into a directed graph structure, which serves as the trajectory-driven causal chain for attitude evolution.

5. The method according to claim 4, characterized in that, Perform topological analysis on the attitude evolution causal chain to determine whether there are any topological inconsistencies between the trajectory and attitude, including: From the directed graph structure of the attitude evolution causal chain, the out-degree value of each trajectory state node is counted, and an out-degree distribution sequence of trajectory state nodes is constructed. The information entropy of the out-degree distribution sequence is calculated as the trajectory topological entropy. The in-degree value of each attitude state node is counted, and an in-degree distribution sequence of attitude state nodes is constructed. The information entropy of the in-degree distribution sequence is calculated as the attitude topological entropy. The difference between the trajectory topological entropy and the attitude topological entropy is calculated as the topological entropy difference. The momentum flow vectors on each causal edge in the momentum flow density field are reconstructed into a vector field in the topological space of the attitude evolution causal chain. The curl operator is calculated on the reconstructed vector field to obtain the momentum flow curl field. The circulation integral of the momentum flow curl field is calculated as the momentum conservation deviation. The momentum conservation deviation is converted into a topological stability boundary quantity through a Lyapunov exponential mapping transformation. Based on the functional relationship between the topological stability boundary quantity and the geometric mean of the trajectory topological entropy and the attitude topological entropy, the allowable upper bound of the topological entropy difference is determined. When the difference in topological entropy exceeds the allowable upper bound, it is determined that there is a topological inconsistency anomaly between the trajectory and attitude.

6. The method according to claim 1, characterized in that, Based on the set of abnormal nodes, causal interpolation compensation is performed on the abnormal features using the causal paths in the causal chain that are not affected by the abnormalities. The fusion analysis results of the generated trajectory and pose based on the compensated features include: In the attitude evolution causal chain, identify all incoming and outgoing edges of each abnormal node in the abnormal node set, mark the incoming and outgoing edges of the abnormal nodes as abnormal causal edges, mark the remaining causal edges as normal causal edges, and take the causal path composed only of the normal causal edges as the causal path unaffected by the abnormality; The trajectory feature vectors and attitude feature vectors propagating on each normal causal edge are extracted from the causal path that is not affected by the anomaly, and used as normal feature samples; Calculate the topological distance between the anomalous node and each node on the causal path unaffected by the anomalous node in the topological space of the attitude evolution causal chain, and construct an interpolation weight distribution based on the topological distance; The normal feature samples corresponding to each node on the causal path that is not affected by the anomaly are weighted and combined according to the interpolation weight distribution to generate the compensation feature vector of the abnormal node. The initial feature vector of the abnormal node is replaced with the compensation feature vector to obtain the compensated attitude evolution causal chain. The compensated trajectory feature sequence and the compensated posture feature sequence are extracted from the compensated posture evolution causal chain and spatiotemporally aligned and fused to generate the fusion analysis result of athlete trajectory and posture.

7. An AI fusion analysis system for athlete trajectory and posture, used to implement the method as described in any one of claims 1-6, characterized in that, include: The Riemann coding unit is used to perform spatiotemporal convolutional coding, graph topology propagation, and dynamic modeling on the athlete's global trajectory coordinate sequence, local posture joint coordinate sequence, and motion velocity vector field data to obtain motion features, posture features, and motion trend features, and map them to a unified Riemannian manifold space. A cross-modal fusion unit is used to calculate the geodesic distance between the motion feature and the attitude feature in the Riemannian manifold space as a spatial correlation metric between trajectory and attitude. Based on the tangent vector direction of the motion trend feature in the Riemannian manifold space, the spatial correlation metric is dynamically consistent and cross-modal fusion features are constructed. A causal chain building unit is used to establish a causal edge between the previous state of the global trajectory coordinate sequence and the current state of the local attitude joint coordinate sequence based on the cross-modal fusion features, and to quantify the strength of the causal edge according to the dynamic constraints of the motion velocity vector field data to obtain a trajectory-driven attitude evolution causal chain. An anomaly detection unit is used to perform topological structure analysis on the attitude evolution causal chain, determine whether there is an anomaly of topological inconsistency between trajectory and attitude, and extract the corresponding set of anomaly nodes. The causal compensation unit is used to perform causal interpolation compensation on abnormal features based on the abnormal node set and the causal paths in the causal chain that are not affected by the abnormality, and to generate a fusion analysis result of trajectory and attitude based on the compensated features.

8. 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 6.

9. 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 6.

Images