Kinematic constraint-based re-competitive action recognition method

Abstract

The invention provides a heavy competition action recognition method based on kinematics constraints, and aims to solve the problems that the existing action recognition technology is insufficient in capturing high-speed motion joint features and lacks physiological rationality constraints, and high-precision action classification is realized by fusing a kinematics correction mechanism and a layered feature isolation architecture. Specifically, the joint kinematics correction layer is designed, illegal joint angles are automatically corrected according to biomechanical constraints of the re-competitive actions, and a skeleton sequence conforming to ergonomics is generated; meanwhile, a body and limb double-branch isolation encoder is constructed, core trunk and tail end limb features are processed respectively, and multi-scale spatial-temporal features are aggregated through a cross-level fine-grained fusion module; in addition, a dynamic calculation path selection mechanism is introduced, and the network depth is adaptively adjusted according to the complexity of the action. The method effectively solves the problems of high-speed action detail loss and non-physiological attitude interference.

Description

Technical Field

[0001] The present invention relates to heavy-duty athletic action recognition based on kinematic constraints, and solves the problem of inaccurate complex action capture in the prior art by introducing a kinematic constraint mechanism and a multi-branch feature extraction architecture. Background Art

[0002] With the increasing popularity of sports and the continuous improvement of competitive levels, motion recognition technology is becoming increasingly important in fields such as sports training, competition analysis, and sports rehabilitation. Especially for sports such as heavyweight athletics, which require extremely precise movements, accurately identifying and evaluating athletes' movement quality is key to improving training effectiveness and competition performance. However, existing motion recognition technology still faces many challenges when applied to high-speed, complex sports such as heavyweight athletics.

[0003] Traditional motion recognition methods primarily rely on computer vision and machine learning algorithms, analyzing video images or depth sensor data to identify athletes' movements. While these methods can classify and identify movements to a certain extent, they often struggle to accurately capture subtle joint changes and complex motion trajectories when dealing with high-speed sports such as heavyweight athletics. The complex movements in heavyweight athletics involve extremely rapid changes in joint angles and movement speeds, making traditional methods prone to loss of movement details and misidentification.

[0004] In recent years, with the development of deep learning technology, motion recognition methods based on deep neural networks have gradually become a research hotspot. These methods, by constructing complex neural network models, can automatically learn feature representations from motion data, thereby improving recognition accuracy. However, most existing deep learning models lack biomechanical constraints. In heavy athletic motion recognition, the movement of human joints is strictly limited by biomechanical laws, such as the range of knee flexion and hip abduction. If the model fails to fully account for these constraints, the recognition results may not conform to human physiology and may even mislead the recognition of incorrect movements.

[0005] Furthermore, existing motion recognition technology has shortcomings in assessing motion quality. Most methods only output motion categories, but fail to provide detailed kinematic analysis and quality assessment reports. This is far from sufficient for both athlete technical improvement and coaching guidance. For example, coaches need to understand which joint angles exhibit excessive deviations during movement execution, which movement phases exhibit abnormal durations, and whether there are deviations in the movement trajectory, so as to provide targeted training guidance. Summary of the Invention

[0006] In response to the problems of insufficient capture of motion joint features and lack of constraints in existing motion recognition technologies, the present invention proposes a method for heavy athletic motion recognition based on kinematic constraints. The present invention achieves high-precision motion classification by integrating a kinematic correction mechanism with a hierarchical feature isolation architecture. Specifically, the present invention designs a joint kinematic correction layer to correct illegal joint angles based on the biomechanical constraints of heavy athletic motions, generating an ergonomic skeletal sequence; at the same time, it constructs a dual-branch isolation encoder for the body and limbs to process the core trunk and terminal limb features respectively, and aggregates multi-scale spatiotemporal features through a cross-level fine-grained fusion module; in addition, it introduces a dynamic calculation path selection mechanism to adjust the network depth according to the complexity of the motion. This invention effectively solves the problem of loss of high-speed motion details and non-physiological posture interference, and provides strong support for the accurate recognition and quality assessment of heavy athletic motions.

[0007] A method for recognizing heavy athletic movements based on kinematic constraints of the present invention comprises the following steps:

[0008] S1: Receive 3D skeleton sequence input data x∈R T×N×3 , where T represents the total number of video frames, N represents the number of human joints, and each joint contains three-dimensional space coordinates (x, y, z); based on the biomechanical characteristics of heavy athletic sports, the angular range constraint set of each joint is defined as Φ = {φ knee ,φ hip ,φ elbow}, where φ knee = [0°, 150°] limits the knee flexion range, φ hip = [0°, 45°] limits the range of hip abduction, φ elbow = [-5°, 180°] limits the range of motion of the elbow joint;

[0009] For each frame of bone data X t ∈R N×3 , calculate the actual angle θ of each joint j (t), by connecting the adjacent bone vector v of the joint a =x proximal -x joint , v b =x distal -x joint , where x proximal represents the hip joint coordinates, x joint represents the knee joint coordinates, x distal Represents the ankle joint coordinates, and uses the vector angle formula to calculate the joint angle:

[0010]

[0011] Among them, v a Represents the proximal bone vector, vb represents the distal bone vector, θ j (t) represents the angle value of joint j at time t; ||v b || represents vector v b The modulus length, ||v a || represents vector v a Length of the module;

[0012] When the joint angle θ is detected j (t) exceeds the constraint range φ j When , kinematic correction is performed; the correction process first calculates the deviation Δθ between the current angle and the nearest constraint boundary j :

[0013] Δθ j =|θ j (t)-nearest(θ j (t),φ j )|

[0014] Among them, nearest(θ j (t),φ j ) function returns the constraint boundary value closest to the current angle to determine the actual correction amount:

[0015]

[0016] Among them, β represents the adjustable correction intensity coefficient, Indicates the range of motion allowed for the joint. represents the maximum allowed angle of joint j, Represents the minimum allowable angle of joint j, and corrects the joint position through three-dimensional space rotation transformation:

[0017]

[0018] Among them, R(n,θ) represents the transformation matrix of the angle θ around the rotation axis n, and the rotation axis n is obtained by normalizing the cross product of the bone vector x j (t) represents the original joint coordinates, x' j (t) represents the corrected joint coordinates, and outputs the skeleton sequence X'∈R that conforms to the kinematic constraints T×N×3 ;

[0019] S2: The corrected skeleton data is input into a two-branch feature extraction network. The body branch encodes the motion features of the core joints of the trunk, and the limb branch focuses on the feature extraction of the distal joints of the limbs. The two branches use different network structure designs to adapt to their respective feature expression requirements.

[0020] S3: A multi-scale temporal convolution module is constructed in the limb branch. Different temporal convolution kernel sizes are set to capture the short-term burst characteristics and long-term trajectory characteristics of the kicking action respectively. Residual connections are introduced to ensure feature transfer.

[0021] S4: Design a cross-level feature fusion module based on the attention mechanism to dynamically integrate the global posture features of the body branch and the local motion features of the limb branch. This module can automatically learn the importance weights of features at different levels.

[0022] S5: Set up early exit mechanisms at each layer of the network to dynamically decide whether to output results early based on the discriminant confidence of the current feature representation. For simple and clear actions, the calculation can be terminated early, while complex actions continue to be analyzed in depth.

[0023] S6: While outputting the final motion recognition results, a detailed kinematic analysis report is generated, which contains auxiliary information such as the original angle deviation of each joint, the correction amplitude statistics, and the frequency distribution of constraint violations, providing data support for motion quality evaluation and technical improvement.

[0024] According to a specific implementation of an embodiment of the present invention, the specific steps of S2 are:

[0025] S21. Separate the trunk core joint set J from the input data core = {hip, spine, shoulder} coordinate data, construct a spatiotemporal graph convolutional network for feature encoding; the network input is The spatial graph convolution is:

[0026]

[0027] Among them, A k represents the kth sub-adjacency matrix based on the natural connection of the human body, represents the trainable weight of the k-th sub-adjacency in the l-th layer, Represents the global posture feature output by the body branch, σ represents the ReLU activation function, and a one-dimensional convolution kernel K is used in the time dimension. t ∈R 3×d Perform sliding calculations and output a feature tensor F with global perception capabilities body ∈R T×d , where d represents the feature dimension and T represents the total number of video frames;

[0028] S22, on the limb joints limb ={hand, elbow, knee, ankle}, design a temporal convolution module with multi-scale perception, input After group convolution processing, the upper and lower limb joints are separated and calculated:

[0029]

[0030] Among them, TCN arm Represents the upper limb dedicated temporal convolutional network, TCN leg Represents the lower limb dedicated temporal convolutional network, J upper represents the upper limb joint point subset, J lower represents the lower limb joint point subset, F limb Represents the local motion features output by the limb branch. The TCN module contains three parallel convolution paths, and the kernel sizes are set to k t =3,5,7:

[0031]

[0032] in, Multi-scale feature tensor representing limb branches, k t Represents the convolution kernel size, X' represents the corrected skeleton sequence, and the original motion information is retained through residual connection. The attention gating mechanism is used for the hand joints to enhance the weight of fine movements:

[0033] α j =sigmoid(W a x′ j )

[0034] x″ j =α j ·x′ j

[0035] Among them, W a The weight matrix representing the attention gate of the hand joints, x″ j represents the joint coordinates after the gating mechanism, α j represents the attention gating weight of the j-th joint.

[0036] According to a specific implementation of an embodiment of the present invention, the specific steps of S3 are:

[0037] S31, limb branch feature F based on step S2 limb ∈R T×2d , action features are extracted layer by layer through multi-scale temporal convolution modules, with kernel size k t =3 one-dimensional convolution to capture the instantaneous burst characteristics of the kicking action, covering the time window of three consecutive frames:

[0038] F short =ReLU(W3*F limb +b3)

[0039] Among them, F short ∈R T×2d Represents the instantaneous action feature of the short-term convolution output, b3∈R2d Represents the bias term, convolution kernel parameter W3∈R 3×2d×2d Specifically captures instantaneous signals of sudden changes in knee joint angles. Each output time point correlates to the motion state of the previous and next frames. The time dimension is kept constant through step size and padding, and the ReLU activation function enhances nonlinear expression capabilities.

[0040] S32, by kernel size k t = 5 mid-range dilated convolution expands the time window to nine frames:

[0041] F mid =DilatedConv(W5*F limb +b5)

[0042] Among them, F mid ∈R T×2d Represents the action fragment features output by the mid-range dilated convolution, DilatedConv represents the operation function for mid-range feature extraction, and the parameter W5∈R 5×2d×2d Covering the nine-frame time window, analyzing the entire process of the roundhouse kick, the bias term b5∈R 2d After adding the convolution result, the output features encode the movement pattern of hip rotation and leg swing;

[0043] S33, use wide convolution kernel k t =7 Establish action cycle:

[0044] F long =ConvlD k＝7 (W7*F limb +b7)

[0045] Among them, F long ∈R (T-6)×2d Represents the action cycle characteristics of the long-term convolution output, b7∈R 2d Represents the bias term, parameter W7∈R 7×2d×2d By filling;

[0046] S34. Apply residual connection to the three-way features to retain the original information:

[0047] F short =F short +F limb

[0048] F mid =F mid +F limb

[0049] F long =F long +F limb

[0050] Among them, F'short Represents the features of short-term features enhanced by residual connections, F' mid represents the enhanced features of the mid-range features after residual connection, F' long Represents the features of the long-term features after zero padding and residual connection. The features of each path are compressed into a 1x2d vector through global average pooling, and the adaptive weight is calculated:

[0051]

[0052] Among them, α, β, and γ represent dynamic weight coefficients, corresponding to the weights of short-term, medium-term, and long-term features, respectively. GAP(@) represents the global average pooling function. W s ∈R 6d×3 Represents the weight matrix, learns the importance of different time scales, and finally outputs multi-scale features through channel-by-channel weighted fusion:

[0053] F' limb =α⊙F' short +β⊙F' mid +γ⊙F' long .

[0054] Among them, F' limb ∈R T×2d represents the position encoding matrix, and ⊙ represents the element-wise multiplication operation.

[0055] According to a specific implementation of an embodiment of the present invention, the specific steps of S4 are:

[0056] S41. The interaction between body posture features and limb movement features is realized through the bidirectional attention mechanism, and a dynamic association between the two features is established, so that the global posture information can guide the screening of local movement features, and the local movement features can also in turn affect the understanding of the global posture. In specific implementation, the body features output by step S2 and the limb features output by step S3 are linearly transformed respectively to generate the query (Query), key (Key) and value (Value) vectors Q required for attention calculation body , K limb 、V limb :

[0057]

[0058] Among them, d h =d / 2 represents the dimension of each attention head, W q 、W k 、W v Represents the query, key, and value weight matrices of the attention mechanism. d represents the feature dimension. When calculating the attention weight, the kinematic constraint matrix M∈{0,-∞} is introduced.T×T , ensuring temporal causality and preventing information from future frames from affecting the judgment of the current frame:

[0059]

[0060] Among them, d h =d / 2 represents the dimension of each attention head, A body→limb Represents the attention weight matrix normalized by softmax, and the final attention output is the weighted summation of the value vector by the weight matrix:

[0061] F body→limb =A body→limb V limb

[0062] Among them, F body→limb represents the attention fusion feature from the body branch to the limb branch, V limb The value vector representing the attention;

[0063] S42. After obtaining the bidirectional attention output, the final feature representation is optimized through the gated fusion mechanism, and the attention outputs of the two directions are spliced ​​together:

[0064] F concat =[F body→limb ; F limb→body ]∈R T×4d

[0065] Among them, F body→limb represents the attention fusion feature from the body branch to the limb branch, F limb→body represents the attention fusion feature from the limb branch to the body branch, T represents the total number of video frames, and F concat Represents the concatenated bidirectional features. The optimal fusion weights are learned through a gated convolutional layer. The convolutional layer uses a sliding window of 5 frames to effectively capture the local dynamic patterns of the action clip:

[0066] G=σ(Conv1D k＝5 (F concat )W g )

[0067] Among them, W g ∈R 4d×2d Represents the gate weight matrix, Conv1D k＝5Represents a one-dimensional convolution operation, and σ represents a sigmoid activation function, which limits the output to the range [0, 1], forming a switching mechanism. The gate value is automatically adjusted according to the complexity of the action: for simple actions, the gate value maintains a relatively balanced fusion; for complex actions, the gate value is increased, emphasizing the features after the enhanced attention. The output is dynamically combined through the gating mechanism:

[0068] F final =G⊙F body→limb +(1-G)⊙F limb→body

[0069] Among them, F final ∈R T×2d represents the fused features, F body→limb represents the attention fusion feature from the body branch to the limb branch, F limb→body Represents the attention fusion feature from the limb branch to the body branch, ⊙ represents the element-by-element multiplication operation, G represents the gate vector activated by sigmoid, G represents the gate vector activated by sigmoid. This design enables the network to adaptively adjust the feature fusion strategy according to the characteristics of different actions and output the feature F final It not only retains the detailed information of the original action, but also integrates the global posture context.

[0070] According to a specific implementation of an embodiment of the present invention, the specific steps of S5 are:

[0071] S51. Set up lightweight classification subnetworks at key levels of the network to evaluate feature quality in real time and guide computational path selection. Each classification subnetwork performs global average pooling on input features to compress the time dimension.

[0072]

[0073] Among them, h (l) ∈R d represents the global average pooling result of the l-th layer feature, represents the feature representation of the lth layer network at time point t, T represents the total number of video frames, and the category probability distribution is calculated through a two-layer fully connected network:

[0074]

[0075] in, and represents the trainable parameters, C represents the total number of action categories, and p (l) ∈R C represents the probability distribution of category prediction at layer l, ReLU(·) represents the activation function, and the confidence score function is calculated as:

[0076]

[0077] Among them, H(p (l) ) represents the entropy value of the predicted distribution of the lth layer, s (l) ∈[0,1] represents the confidence score, which takes into account the prediction certainty and distribution concentration. The higher the value, the stronger the discriminative power of the current level feature.

[0078] S52. Design a dynamic calculation strategy based on hierarchical confidence, and use a linear attenuation design for the confidence threshold function:

[0079] τ(l)=uv*l

[0080] Where u and v represent linear parameters, and τ(l) represents the confidence threshold of the lth layer. Deep networks require higher confidence to trigger early exit and ensure the recognition accuracy of complex actions. The prediction consistency condition requires that the output of consecutive layers remain stable:

[0081] ||p (l) -p (l-1) ||1 <e1,||p (l) -p (l-2) ||1 <e2

[0082] Among them, ||p (l) -p (l-1) ||1 represents the absolute difference between the predicted distribution of the first layer and the previous layer, ||p (l) -p (l-2) ||1 represents the absolute difference between the predicted distribution of the first layer and the first two layers, e1 and e2 represent thresholds, and the final early exit decision function is:

[0083] EarlyExit (l) =[s (l) >τ(l)∧Consistency (l) ]

[0084] Among them, EarlyExit (l) Indicates the early exit decision flag of the lth layer, s (l) Represents the confidence score, automatically adjusts the calculation depth according to the complexity of the action, optimizes the calculation efficiency while ensuring the recognition accuracy, and outputs the final action category prediction The actual calculated depth l*.

[0085] According to a specific implementation of the embodiment of the present invention, the specific steps of S6 are:

[0086] S6. Based on the kinematic correction data recorded in step S1, the average angle deviation value of each joint in the entire action sequence is calculated. This calculation process can be expressed as:

[0087]

[0088] Among them, θ j (t) represents the actual angle of joint j at time point t, T represents the total number of video frames, Indicates the angle value of the joint in the standard action. Deviation values ​​are calculated for key parts such as the hip joint, knee joint, and ankle joint, and the overall correction amplitude of each frame is analyzed:

[0089]

[0090] Among them, x' j (t) represents the corrected coordinates of joint j at time point t, x j (t) represents the original coordinates of joint j at time point t, ||·||2 represents the Euclidean distance, and Γ(t) represents the sum of the joint correction amplitudes at time point t. When the correction amplitude exceeds the preset threshold, these abnormal frames are marked. A comprehensive score is calculated based on multiple dimensions, taking into account key factors such as average angle deviation and maximum correction amplitude:

[0091]

[0092] Among them, λ1 and λ2 are weight coefficients determined through a large number of experiments, and Q is the comprehensive quality score of the action. represents the average angle deviation of all joints in the action sequence, Γ max Indicates the maximum correction amplitude of all frames in the action sequence. Based on the scoring results, the action quality is divided into three levels: excellent, good, and needs improvement, and corresponding text evaluation is given;

[0093] Based on kinematic constraints and temporal logic, it identifies three common types of incorrect movements: joint angles exceeding their normal range of motion; abnormal durations of movement phases; and excessive deviations from the motion trajectory. Each detected error is precisely located on the timeline and associated with a specific joint or limb.

[0094] The final report is stored in a structured data format, and personalized improvement suggestions are generated based on the results to provide guidance for athletes to improve their skills. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 Flowchart of this method;

[0096] Figure 2 This is the architectural diagram of this method. DETAILED DESCRIPTION

[0097] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to examples and drawings.

[0098] As attached Figure 1 and attached Figure 2 As shown, a method for recognizing heavy athletic movements based on kinematic constraints includes the following steps:

[0099] Step 1: Receive 3D skeleton sequence input data x∈R T×N×3 , where T represents the total number of video frames, N represents the number of human joints, and each joint contains three-dimensional space coordinates (x, y, z); based on the biomechanical characteristics of heavy athletic sports, the angular range constraint set of each joint is defined as Φ = {φ knee ,φ hip ,φ elbow}, where φ knee = [0°, 150°] limits the knee flexion range, φ hip = [0°, 45°] limits the range of hip abduction, φ elbow = [-5°, 180°] limits the range of motion of the elbow joint;

[0100] For each frame of bone data X t ∈R N×3 , calculate the actual angle θ of each joint j (t), by connecting the adjacent bone vector v of the joint a =x proximal -x joint , v b =x distal -x joint , where x proximal represents the hip joint coordinates, x joint represents the knee joint coordinates, x distal Represents the ankle joint coordinates, and uses the vector angle formula to calculate the joint angle:

[0101]

[0102] Among them, v a Represents the proximal bone vector, v b represents the distal bone vector, θ j (t) represents the angle value of joint j at time t; ||v b || represents vector v b The modulus length, ||v a || represents vector v a Length of the module;

[0103] When the joint angle θ is detected j (t) exceeds the constraint range φ jWhen , kinematic correction is performed; the correction process first calculates the deviation Δθ between the current angle and the nearest constraint boundary j :

[0104] Δθ j =|θ j (t)-nearest(θ j (t),φ j )|

[0105] Among them, nearest(θ j (t),φ j ) function returns the constraint boundary value closest to the current angle to determine the actual correction amount:

[0106]

[0107] Among them, β represents the adjustable correction intensity coefficient, Indicates the range of motion allowed for the joint. represents the maximum allowed angle of joint j, Represents the minimum allowable angle of joint j, and corrects the joint position through three-dimensional space rotation transformation:

[0108]

[0109] Among them, R(n,θ) represents the transformation matrix of the angle θ around the rotation axis n, and the rotation axis n is obtained by normalizing the cross product of the bone vector x j (t) represents the original joint coordinates, x' j (t) represents the corrected joint coordinates, and outputs the skeleton sequence X'∈R that conforms to the kinematic constraints T×N×3 ;

[0110] Step 2: The corrected skeleton data is input into a two-branch feature extraction network. The body branch encodes the motion characteristics of the core joints of the trunk, while the limb branch focuses on the feature extraction of the distal joints of the limbs. The two branches use different network structure designs to adapt to their respective feature expression requirements.

[0111] Separate the torso core joint set J from the input data core = {hip, spine, shoulder} coordinate data, construct a spatiotemporal graph convolutional network for feature encoding; the network input is The spatial graph convolution is:

[0112]

[0113] Among them, A k represents the kth sub-adjacency matrix based on the natural connection of the human body, represents the trainable weight of the k-th sub-adjacency in the l-th layer, Represents the global posture feature output by the body branch, σ represents the ReLU activation function, and a one-dimensional convolution kernel K is used in the time dimension. t ∈R 3×d Perform sliding calculations and output a feature tensor F with global perception capabilities body ∈R T×d , where d represents the feature dimension and T represents the total number of video frames;

[0114] Joints of the limbs limb ={hand, elbow, knee, ankle}, design a temporal convolution module with multi-scale perception, input After group convolution processing, the upper and lower limb joints are separated and calculated:

[0115]

[0116] Among them, TCN arm Represents the upper limb dedicated temporal convolutional network, TCN leg Represents the lower limb dedicated temporal convolutional network, J upper represents the upper limb joint point subset, J lower represents the lower limb joint point subset, F limb Represents the local motion features output by the limb branch. The TCN module contains three parallel convolution paths, and the kernel sizes are set to k t =3,5,7:

[0117]

[0118] in, Multi-scale feature tensor representing limb branches, k t Represents the convolution kernel size, X' represents the corrected skeleton sequence, and the original motion information is retained through residual connection. The attention gating mechanism is used for the hand joints to enhance the weight of fine movements:

[0119] α j =sigmoid(W a x j ')

[0120] x″ j =α′ j ·x j

[0121] Among them, W a The weight matrix representing the attention gate of the hand joints, x″ j represents the joint coordinates after the gating mechanism, α j represents the attention gating weight of the j-th joint;

[0122] Step 3: Construct a multi-scale temporal convolution module in the limb branch. By setting different temporal convolution kernel sizes, it captures the short-term burst characteristics and long-term trajectory characteristics of the kicking action respectively. At the same time, it introduces residual connections to ensure feature transfer.

[0123] Based on the limb branch feature F in step S2 limb ∈R T×2d , action features are extracted layer by layer through multi-scale temporal convolution modules, with kernel size k t =3 one-dimensional convolution to capture the instantaneous burst characteristics of the kicking action, covering the time window of three consecutive frames:

[0124] F short =ReLU(W3*F limb +b3)

[0125] Among them, F short ∈R T×2d Represents the instantaneous action feature of the short-term convolution output, b3∈R 2d Represents the bias term, convolution kernel parameter W3∈R 3×2d×2d Specifically captures instantaneous signals of sudden changes in knee joint angles. Each output time point correlates to the motion state of the previous and next frames. The time dimension is kept constant through step size and padding, and the ReLU activation function enhances nonlinear expression capabilities.

[0126] By kernel size k t = 5 mid-range dilated convolution expands the time window to nine frames:

[0127] F mid =DilatedConv(W5*F limb +b5)

[0128] Among them, F mid ∈R T×2d Represents the action fragment features output by the mid-range dilated convolution, DilatedConv represents the operation function for mid-range feature extraction, and the parameter W5∈R 5×2d×2d Covering the nine-frame time window, analyzing the entire process of the roundhouse kick, the bias term b5∈R 2d After adding the convolution result, the output features encode the movement pattern of hip rotation and leg swing;

[0129] Use a wide convolution kernel k t =7 Establish action cycle:

[0130] F long =ConvlD k＝7 (W7*F limb +b7)

[0131] Among them, F long ∈R (T-6)×2dRepresents the action cycle characteristics of the long-term convolution output, b7∈R 2d Represents the bias term, parameter W7∈R 7×2d×2d By filling;

[0132] Applying residual connections to the three-way features preserves the original information:

[0133] F short =F short +F limb

[0134] F mid =F mid +F limb

[0135] F long =F long +F limb

[0136] Among them, F' short Represents the features of short-term features enhanced by residual connections, F' mid represents the enhanced features of the mid-range features after residual connection, F' long Represents the features of the long-term features after zero padding and residual connection. The features of each path are compressed into a 1x2d vector through global average pooling, and the adaptive weight is calculated:

[0137]

[0138] Among them, α, β, and γ represent dynamic weight coefficients, corresponding to the weights of short-term, medium-term, and long-term features, respectively. GAP(·) represents the global average pooling function, and W s ∈R 6d×3 Represents the weight matrix, learns the importance of different time scales, and finally outputs multi-scale features through channel-by-channel weighted fusion:

[0139] F' limb =α⊙F' short +β⊙F' mid +γ⊙F' long .

[0140] Among them, F' limb ∈R T×2d represents the position encoding matrix, ⊙ represents the element-by-element multiplication operation;

[0141] Step 4: Design a cross-level feature fusion module based on the attention mechanism to dynamically integrate the global posture features of the body branch and the local motion features of the limb branch. This module can automatically learn the importance weights of features at different levels.

[0142] The interaction between body posture features and limb movement features is realized through the bidirectional attention mechanism, and a dynamic association between the two features is established, so that the global posture information can guide the screening of local movement features, and the local movement features can in turn affect the understanding of the global posture. In specific implementation, the body features output by step S2 and the limb features output by step S3 are linearly transformed respectively to generate the query (Query), key (Key) and value (Value) vectors Q required for attention calculation body , K limb 、V limb :

[0143]

[0144] Among them, d h =d / 2 represents the dimension of each attention head, W q 、W k 、W v Represents the query, key, and value weight matrices of the attention mechanism. d represents the feature dimension. When calculating the attention weight, the kinematic constraint matrix M∈{0,-∞} is introduced. T×T , ensuring temporal causality and preventing information from future frames from affecting the judgment of the current frame:

[0145]

[0146] Among them, d h =d / 2 represents the dimension of each attention head, A body→limb Represents the attention weight matrix normalized by softmax, and the final attention output is the weighted summation of the value vector by the weight matrix:

[0147] F body→limb =A body→limb V limb

[0148] Among them, F body→limb represents the attention fusion feature from the body branch to the limb branch, V limb The value vector representing the attention;

[0149] After obtaining the bidirectional attention output, the final feature representation is optimized through the gated fusion mechanism, and the attention outputs of the two directions are spliced ​​together:

[0150] F concat =[F body→limb ; F limb→body ]∈R T×4d

[0151] Among them, F body→limbrepresents the attention fusion feature from the body branch to the limb branch, F limb→body represents the attention fusion feature from the limb branch to the body branch, T represents the total number of video frames, and F concat Represents the concatenated bidirectional features. The optimal fusion weights are learned through a gated convolutional layer. The convolutional layer uses a sliding window of 5 frames to effectively capture the local dynamic patterns of the action clip:

[0152] G=σ(Conv1D k＝5 (F concat )W g )

[0153] Among them, W g ∈R 4d×2d Represents the gate weight matrix, Conv1D k＝5 Represents a one-dimensional convolution operation, and σ represents a sigmoid activation function, which limits the output to the range [0, 1], forming a switching mechanism. The gate value is automatically adjusted according to the complexity of the action: for simple actions, the gate value maintains a relatively balanced fusion; for complex actions, the gate value is increased, emphasizing the features after the enhanced attention. The output is dynamically combined through the gating mechanism:

[0154] F final =G⊙F body→limb +(1-G)⊙F limb→body

[0155] Among them, F final ∈R T×2d represents the fused features, F body→limb represents the attention fusion feature from the body branch to the limb branch, F limb→body Represents the attention fusion feature from the limb branch to the body branch, ⊙ represents the element-by-element multiplication operation, G represents the gate vector activated by sigmoid, G represents the gate vector activated by sigmoid. This design enables the network to adaptively adjust the feature fusion strategy according to the characteristics of different actions and output the feature F final It not only preserves the details of the original action, but also integrates the global posture context;

[0156] Step 5: Set up an early exit mechanism at each layer of the network to dynamically decide whether to output the result early based on the judgment confidence of the current feature representation. For simple and clear actions, the calculation can be terminated early, while complex actions continue to be analyzed in depth:

[0157] A lightweight classification subnetwork is set up at the key layers of the network to evaluate feature quality in real time and guide the selection of computational paths. Each classification subnetwork performs global average pooling on the input features to compress the time dimension:

[0158]

[0159] Among them, h (l) ∈R d represents the global average pooling result of the l-th layer feature, represents the feature representation of the lth layer network at time point t, T represents the total number of video frames, and the category probability distribution is calculated through a two-layer fully connected network:

[0160]

[0161] in, and represents the trainable parameters, C represents the total number of action categories, and p (l) ∈R C represents the probability distribution of category prediction at layer l, ReLU(·) represents the activation function, and the confidence score function is calculated as:

[0162]

[0163] Among them, H(p (l) ) represents the entropy value of the predicted distribution of the lth layer, s (l) ∈[0,1] represents the confidence score, which takes into account the prediction certainty and distribution concentration. The higher the value, the stronger the discriminative power of the current level feature.

[0164] A dynamic calculation strategy is designed based on hierarchical confidence, and the confidence threshold function adopts a linear attenuation design:

[0165] τ(l)=uv*l

[0166] Where u and v represent linear parameters, and τ(l) represents the confidence threshold of the lth layer. Deep networks require higher confidence to trigger early exit and ensure the recognition accuracy of complex actions. The prediction consistency condition requires that the output of consecutive layers remain stable:

[0167] ||p (l) -p (l-1) ||1 <e1,||p (l) -p (l-2) ||1 <e2

[0168] Among them, ||p (l) -p (l-1) ||1 represents the absolute difference between the predicted distribution of the first layer and the previous layer, ||p (l) -p (l-2) ||1 represents the absolute difference between the predicted distribution of the first layer and the first two layers, e1 and e2 represent thresholds, and the final early exit decision function is:

[0169] EarlyExit (l) =[s(l) >τ(l)∧Consistency (l) ]

[0170] Among them, EarlyExit (l) Indicates the early exit decision flag of the lth layer, s (l) Represents the confidence score, automatically adjusts the calculation depth according to the complexity of the action, optimizes the calculation efficiency while ensuring the recognition accuracy, and outputs the final action category prediction The actual calculated depth l*;

[0171] Step 6: When outputting the final motion recognition results, a detailed kinematic analysis report is generated, which includes auxiliary information such as the original angle deviation of each joint, the statistics of the correction amplitude, and the frequency distribution of constraint violations, providing data support for motion quality evaluation and technical improvement;

[0172] Based on the kinematic correction data recorded in step S1, the average angle deviation value of each joint in the entire action sequence is calculated. This calculation process can be expressed as:

[0173]

[0174] Among them, θ j (t) represents the actual angle of joint j at time point t, T represents the total number of video frames, Indicates the angle value of the joint in the standard action. Deviation values ​​are calculated for key parts such as the hip joint, knee joint, and ankle joint, and the overall correction amplitude of each frame is analyzed:

[0175]

[0176] Among them, x' j (t) represents the corrected coordinates of joint j at time point t, x j (t) represents the original coordinates of joint j at time point t, ||·||2 represents the Euclidean distance, and Γ(t) represents the sum of the joint correction amplitudes at time point t. When the correction amplitude exceeds the preset threshold, these abnormal frames are marked. A comprehensive score is calculated based on multiple dimensions, taking into account key factors such as average angle deviation and maximum correction amplitude:

[0177]

[0178] Among them, λ1 and λ2 are weight coefficients determined through a large number of experiments, and Q is the comprehensive quality score of the action. represents the average angle deviation of all joints in the action sequence, Γ maxIndicates the maximum correction amplitude of all frames in the action sequence. Based on the scoring results, the action quality is divided into three levels: excellent, good, and needs improvement, and corresponding text evaluation is given;

[0179] Based on kinematic constraints and temporal logic, it identifies three common types of incorrect movements: joint angles exceeding their normal range of motion; abnormal durations of movement phases; and excessive deviations from the motion trajectory. Each detected error is precisely located on the timeline and associated with a specific joint or limb.

[0180] The final report is stored in a structured data format, and personalized improvement suggestions are generated based on the results to provide guidance for athletes to improve their skills.

Claims

1. A method for recognizing heavy athletic movements based on kinematic constraints, characterized in that The following steps are involved: S1: Receive 3D skeleton sequence input data x∈R T×N×3 , where T represents the total number of video frames, N represents the number of human joints, and each joint contains three-dimensional space coordinates (x, y, z); based on the biomechanical characteristics of heavy athletic sports, the angular range constraint set of each joint is defined as Φ = {φ knee ,φ hip ,φ elbow }, where φ knee = [0°, 150°] limits the knee flexion range, φ hip = [0°, 45°] limits the range of hip abduction, φ elbow = [-5°, 180°] limits the range of motion of the elbow joint; For each frame of bone data X t ∈R N×3 , calculate the actual angle θ of each joint j (t), by connecting the adjacent bone vector v of the joint a =x proximal -x joint , v b =x distal -x joint , where x proximal represents the hip joint coordinates, x joint represents the knee joint coordinates, x distal Represents the ankle joint coordinates, and uses the vector angle formula to calculate the joint angle: Among them, v a Represents the proximal bone vector, v b represents the distal bone vector, θ j (t) represents the angle value of joint j at time t; ||v b || represents vector v b The modulus length, ||v a || represents vector v a Length of the module; When the joint angle θ is detected j (t) exceeds the constraint range φ j When , kinematic correction is performed; the correction process first calculates the deviation Δθ between the current angle and the nearest constraint boundary j : Dth j =|θ j (t)-nearest(θ j (t),φ j )| Among them, nearest(θ j (t),φ j ) function returns the constraint boundary value closest to the current angle to determine the actual correction amount: Among them, β represents the adjustable correction intensity coefficient, Indicates the range of motion allowed for the joint. represents the maximum allowed angle of joint j, Represents the minimum allowable angle of joint j, and corrects the joint position through three-dimensional space rotation transformation: Among them, R(n,θ) represents the transformation matrix of the angle θ around the rotation axis n, and the rotation axis n is obtained by normalizing the cross product of the bone vector x j (t) represents the original joint coordinates, x′ j (t) represents the corrected joint coordinates, and outputs the bone sequence X that conforms to the kinematic constraints ' ∈R T×N×3 ; S2: The corrected skeleton data is input into a two-branch feature extraction network. The body branch encodes the motion features of the core joints of the trunk, and the limb branch focuses on the feature extraction of the distal joints of the limbs. The two branches use different network structure designs to adapt to their respective feature expression requirements. S3: A multi-scale temporal convolution module is constructed in the limb branch. Different temporal convolution kernel sizes are set to capture the short-term burst characteristics and long-term trajectory characteristics of the kicking action respectively. Residual connections are introduced to ensure feature transfer. S4: Design a cross-level feature fusion module based on the attention mechanism to dynamically integrate the global posture features of the body branch and the local motion features of the limb branch. This module can automatically learn the importance weights of features at different levels. S5: Set up early exit mechanisms at each layer of the network to dynamically decide whether to output results early based on the discriminant confidence of the current feature representation. For simple and clear actions, the calculation can be terminated early, while complex actions continue to be analyzed in depth. S6: While outputting the final motion recognition results, a detailed kinematic analysis report is generated, which contains auxiliary information such as the original angle deviation of each joint, the correction amplitude statistics, and the frequency distribution of constraint violations, providing data support for motion quality evaluation and technical improvement.

2. A method for recognizing heavy athletic movements based on kinematic constraints according to claim 1, characterized in that The specific method of step S2 is: S21. Separate the trunk core joint set J from the input data core = {hip, spine, shoulder} coordinate data, construct a spatiotemporal graph convolutional network for feature encoding; the network input is The spatial graph convolution is: Among them, A k represents the kth sub-adjacency matrix based on the natural connection of the human body, represents the trainable weight of the k-th sub-adjacency in the l-th layer, Represents the global posture feature output by the body branch, σ represents the ReLU activation function, and a one-dimensional convolution kernel K is used in the time dimension. t ∈R 3×d Perform sliding calculations and output a feature tensor F with global perception capabilities body ∈R T×d , where d represents the feature dimension and T represents the total number of video frames; S22, on the limb joints limb ={hand, elbow, knee, ankle}, design a temporal convolution module with multi-scale perception, input After group convolution processing, the upper and lower limb joints are separated and calculated: Among them, TCN arm Represents the upper limb dedicated temporal convolutional network, TCN leg Represents the lower limb dedicated temporal convolutional network, J upper represents the upper limb joint point subset, J lower represents the lower limb joint point subset, F limb Represents the local motion features output by the limb branch. The TCN module contains three parallel convolution paths, and the kernel sizes are set to k t =3,5,7: in, Multi-scale feature tensor representing limb branches, k t represents the convolution kernel size, X ' Represents the corrected skeleton sequence, retaining the original motion information through residual connections, and using an attention gating mechanism for hand joints to enhance the weight of fine movements: α j =sigmoid(W a x′ j ) x″ j =a j ·x j Among them, W a Represents the weight matrix of hand joint attention gating, x j " represents the joint coordinates after the gating mechanism, α j represents the attention gating weight of the j-th joint.

3. A method for recognizing heavy athletic movements based on kinematic constraints according to claim 1, characterized in that The specific method in step S3 is: S31, limb branch feature F based on step S2 limb ∈R T×2d , action features are extracted layer by layer through multi-scale temporal convolution modules, with kernel size k t =3 one-dimensional convolution to capture the instantaneous burst characteristics of the kicking action, covering the time window of three consecutive frames: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> short <h2 style=";text-align:left;direction:ltr"> =ReLU(W3*F<h2 style=";text-align:left;direction:ltr"> limb <h2 style=";text-align:left;direction:ltr"> +b3) Among them, F short ∈R T×2d Represents the instantaneous action feature of the short-term convolution output, b3∈R 2d Represents the bias term, convolution kernel parameter W3∈R 3×2d×2d Specifically captures instantaneous signals of sudden changes in knee joint angles. Each output time point correlates to the motion state of the previous and next frames. The time dimension is kept constant through step size and padding, and the ReLU activation function enhances nonlinear expression capabilities. S32, by kernel size k t = 5 mid-range dilated convolution expands the time window to nine frames: F mid =DilatedConv(W5*F limb +b5) Among them, F mid ∈R T×2d Represents the action fragment features output by the mid-range dilated convolution, DilatedConv represents the operation function for mid-range feature extraction, and the parameter W5∈R 5×2d×2d Covering the nine-frame time window, analyzing the entire process of the roundhouse kick, the bias term b5∈R 2d After adding the convolution result, the output features encode the movement pattern of hip rotation and leg swing; S33, use wide convolution kernel k t =7 Establish action cycle: F long =ConvlD k＝7 (W7*F limb +b7) Among them, F long ∈R (T-6)×2d Represents the action cycle characteristics of the long-term convolution output, b7∈R 2d Represents the bias term, parameter W7∈R 7×2d×2d By filling; S34. Apply residual connection to the three-way features to retain the original information: F′ short =F short +F limb F′ mid =F mid +F limb F′ long =F′ long +F limb Among them, F ' short Represents the features of short-term features enhanced by residual connections, F' mid represents the enhanced features of the mid-range features after residual connection, F ' long Represents the features of the long-term features after zero padding and residual connection. The features of each path are compressed into a 1x2d vector through global average pooling, and the adaptive weight is calculated: Among them, α, β, and γ represent dynamic weight coefficients, corresponding to the weights of short-term, medium-term, and long-term features, respectively. GAP(·) represents the global average pooling function, and W s ∈R 6d×3 Represents the weight matrix, learns the importance of different time scales, and finally outputs multi-scale features through channel-by-channel weighted fusion: F ' limb =α⊙F ' short +β⊙F' mid +γ⊙F' long Among them, F' limb ∈R T×2d represents the position encoding matrix, and ⊙ represents the element-wise multiplication operation.

4. A method for recognizing heavy athletic movements based on kinematic constraints according to claim 1, characterized in that The specific steps in step S4 are: S41. The interaction between body posture features and limb movement features is realized through the bidirectional attention mechanism, and a dynamic association between the two features is established, so that the global posture information can guide the screening of local movement features, and the local movement features can also in turn affect the understanding of the global posture. In specific implementation, the body features output by step S2 and the limb features output by step S3 are linearly transformed respectively to generate the query (Query), key (Key) and value (Value) vectors Q required for attention calculation body , K limb 、V limb : Among them, d h =d / 2 represents the dimension of each attention head, W q 、W k 、W v Represents the query, key, and value weight matrices of the attention mechanism. d represents the feature dimension. When calculating the attention weight, the kinematic constraint matrix M∈{0,-∞} is introduced. T×T , ensuring temporal causality and preventing information from future frames from affecting the judgment of the current frame: Among them, d h =d / 2 represents the dimension of each attention head, A body→limb Represents the attention weight matrix normalized by softmax, and the final attention output is the weighted summation of the value vector by the weight matrix: F body→limb =A body→limb V limb Among them, F body→limb represents the attention fusion feature from the body branch to the limb branch, V limb The value vector representing the attention; S42. After obtaining the bidirectional attention output, the final feature representation is optimized through the gated fusion mechanism, and the attention outputs of the two directions are spliced ​​together: F concat =[F body→limb ;F limb→body ]∈R T×4d Among them, F body→limb represents the attention fusion feature from the body branch to the limb branch, F limb→body represents the attention fusion feature from the limb branch to the body branch, T represents the total number of video frames, and F concat Represents the concatenated bidirectional features. The optimal fusion weights are learned through a gated convolutional layer. The convolutional layer uses a sliding window of 5 frames to effectively capture the local dynamic patterns of the action clip: G=σ(Conv1D k＝5 (F concat )W g ) Among them, W g ∈R 4d×2d Represents the gate weight matrix, Conv1D k＝5 Represents a one-dimensional convolution operation, and σ represents a sigmoid activation function, which limits the output to the range [0, 1], forming a switching mechanism. The gate value is automatically adjusted according to the complexity of the action: for simple actions, the gate value maintains a relatively balanced fusion; for complex actions, the gate value is increased, emphasizing the features after the enhanced attention. The output is dynamically combined through the gating mechanism: F final =G⊙F body→limb +(1-G)⊙F limb→body Among them, F final ∈R T×2d represents the fused features, F body→limb represents the attention fusion feature from the body branch to the limb branch, F limb→body Represents the attention fusion feature from the limb branch to the body branch, ⊙ represents the element-by-element multiplication operation, G represents the gate vector activated by sigmoid, G represents the gate vector activated by sigmoid. This design enables the network to adaptively adjust the feature fusion strategy according to the characteristics of different actions and output the feature F final It not only retains the detailed information of the original action, but also integrates the global posture context.

5. A method for recognizing heavy athletic movements based on kinematic constraints according to claim 1, characterized in that The specific steps in step S5 are: S51. Set up lightweight classification subnetworks at key levels of the network to evaluate feature quality in real time and guide computational path selection. Each classification subnetwork performs global average pooling on input features to compress the time dimension. Among them, h (l) ∈R d represents the global average pooling result of the l-th layer feature, represents the feature representation of the lth layer network at time point t, T represents the total number of video frames, and the category probability distribution is calculated through a two-layer fully connected network: in, and represents the trainable parameters, C represents the total number of action categories, and p (l) ∈R C represents the probability distribution of category prediction at layer l, ReLU(·) represents the activation function, and the confidence score function is calculated as: Among them, H(p (l) ) represents the entropy value of the predicted distribution of the lth layer, s (l) ∈[0,1] represents the confidence score, which takes into account the prediction certainty and distribution concentration. The higher the value, the stronger the discriminative power of the current level feature. S52. Design a dynamic calculation strategy based on hierarchical confidence, and use a linear attenuation design for the confidence threshold function: τ(l)=uv*l Where u and v represent linear parameters, and τ(l) represents the confidence threshold of the lth layer. Deep networks require higher confidence to trigger early exit and ensure the recognition accuracy of complex actions. The prediction consistency condition requires that the output of consecutive layers remain stable: ||p (l) -p (l-1) ||1<e1,||p (l) -p (l-2) ||1<e2 Among them, ||p (l) -p (l-1) ||1 represents the absolute difference between the predicted distribution of the first layer and the previous layer, ||p (l) -p (l-2) ||1 represents the absolute difference between the predicted distribution of the first layer and the first two layers, e1 and e2 represent thresholds, and the final early exit decision function is: EarlyExit (l) =[s (l) >τ(l)∧Consistency (l) ] Among them, EarlyExit (l) Indicates the early exit decision flag of the lth layer, s (l) Represents the confidence score, automatically adjusts the calculation depth according to the complexity of the action, optimizes the calculation efficiency while ensuring the recognition accuracy, and outputs the final action category prediction The actual calculated depth l*.

6. A method for recognizing heavy athletic movements based on kinematic constraints according to claim 1, characterized in that The specific steps of step S6 are: S6. Based on the kinematic correction data recorded in step S1, the average angle deviation value of each joint in the entire action sequence is calculated. This calculation process can be expressed as: Among them, θ j (t) represents the actual angle of joint j at time point t, T represents the total number of video frames, Indicates the angle value of the joint in the standard action. Deviation values ​​are calculated for key parts such as the hip joint, knee joint, and ankle joint, and the overall correction amplitude of each frame is analyzed: Among them, x' j (t) represents the corrected coordinates of joint j at time point t, x j (t) represents the original coordinates of joint j at time point t, ||·||2 represents the Euclidean distance, and Γ(t) represents the sum of the joint correction amplitudes at time point t. When the correction amplitude exceeds the preset threshold, these abnormal frames are marked. A comprehensive score is calculated based on multiple dimensions, taking into account key factors such as average angle deviation and maximum correction amplitude: Among them, λ1 and λ2 are weight coefficients determined through a large number of experiments, and Q is the comprehensive quality score of the action. represents the average angle deviation of all joints in the action sequence, Γ max Indicates the maximum correction amplitude of all frames in the action sequence. Based on the scoring results, the action quality is divided into three levels: excellent, good, and needs improvement, and corresponding text evaluation is given; Based on kinematic constraints and temporal logic, it identifies three common types of incorrect movements: joint angles exceeding their normal range of motion; abnormal durations of movement phases; and excessive deviations from the motion trajectory. Each detected error is precisely located on the timeline and associated with a specific joint or limb. The final report is stored in a structured data format, and personalized improvement suggestions are generated based on the results to provide guidance for athletes to improve their skills.

Images