A method for evaluating tennis shot stability based on inertial motion capture

Abstract

This invention relates to the fields of artificial intelligence and sports science, and particularly to a method for evaluating the stability of tennis shots based on inertial motion capture. The method includes constructing a spatiotemporal graph of the skeletal sequence of joints and T-frames, where the spatiotemporal graph represents the three-dimensional spatial information of each joint within a frame and the temporal information of node changes between frames; constructing a motion recognition network based on a graph neural network to identify the shot type; after identifying the user's shot type, randomly selecting two shots of the same type, and using a dynamic time warp algorithm to calculate the optimal path distance between the two shots; grouping the human joint nodes, calculating the stability and weight of each group of joint nodes, and weighting the stability of the group of joint nodes using the weights; converting the obtained weighted stability to a percentage system to obtain the final stability evaluation score; this invention can more accurately identify tennis motions and provide more accurate shot stability evaluation results.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and sports science, and in particular to a method for evaluating the stability of tennis shots based on inertial motion capture. Background Technology

[0002] As people's living standards continue to improve, sports are becoming increasingly popular. Tennis, in particular, is favored by young people due to its moderate intensity and engaging nature. However, tennis is technically demanding and has a high entry barrier, often requiring professional coaching. Therefore, by using digital technology to assess and analyze tennis movements, and providing scientific and objective guidance to users, tennis instruction can shift from relying on coach experience to data-driven approaches. This would quickly correct errors in trainees' movements, improve their technical skills, and save significant human and social resources.

[0003] In recent years, advancements in artificial intelligence technology have made progress in the field of human motion recognition, laying a solid foundation for tennis motion recognition. However, image-based methods are susceptible to environmental influences, suffer from occlusion issues, and are costly to implement. While research has explored the application of sensor technology to motion recognition in sports, a comprehensive system is lacking, encompassing tennis motion data acquisition, motion feature extraction, motion stability analysis, motion dataset creation, and recognition. This system would prevent trainees from promptly identifying and correcting technical weaknesses. Furthermore, existing research on stability assessment often relies on manual evaluation, which is difficult to quantify, or on comparing a standard motion to the athlete's own hitting technique, resulting in inaccurate assessments. Summary of the Invention

[0004] To accurately identify tennis motion recognition and assess shot stability, this invention proposes a tennis shot stability assessment method based on inertial motion capture, which specifically includes the following steps:

[0005] S1. Use an inertial motion capture kit to capture the three-dimensional coordinates of multiple joints and the inertial data measured by inertial sensors in each frame when the trainee hits the ball under different types of conditions.

[0006] S2. Based on the set number of joints, construct a spatiotemporal graph of the skeleton sequence of joint points and T frames. This spatiotemporal graph represents the three-dimensional spatial information of each joint within a frame and the temporal information of node changes between frames.

[0007] S3. Construct an action recognition network based on a graph neural network, and use this network to identify the type of ball hit;

[0008] S4. After identifying the user's hitting type, randomly select two hitting actions of the same type and use the dynamic time warp algorithm to calculate the optimal path distance between the two hitting actions.

[0009] S5. Group the human joint nodes, calculate the stability and weight of each group of joint nodes, and use the weight to weight the stability of the group of joint nodes.

[0010] S6. Convert the obtained weighted stability to a percentage system to obtain the final stability evaluation score.

[0011] Furthermore, the inertial data, which measures the three-dimensional coordinates of a joint node by the inertial sensor, includes the x-axis angular velocity, y-axis angular velocity, z-axis angular velocity, x-axis velocity, y-axis velocity, z-axis velocity, x-axis spatial coordinates, y-axis spatial coordinates, and z-axis spatial coordinates at that joint node.

[0012] Furthermore, the graph neural network-based action recognition network comprises a normalization layer, a first GCN layer, a first TCN layer, a first activation function ReLU, a second GCN layer, a second TCN layer, a second activation function ReLU, a third GCN layer, a third TCN layer, a third activation function ReLU, a global pooling layer, and a fully connected layer, connected sequentially. The outputs of the first TCN layer and the first GCN layer have 64 channels, the outputs of the second TCN layer and the second GCN layer have 128 channels, and the outputs of the third TCN layer and the third GCN layer have 256 channels.

[0013] Furthermore, the process of calculating the optimal path distance between two shots using the dynamic time warp algorithm includes:

[0014] Collect m shots of the same shot type from the trainee, selecting 60 frames of data for each shot. The data for the i-th shot is represented as follows:

[0015] M = [A 1(t -30),…,A n(t -30),…,A 1t ,…,A nt ,…,A 1(t+29) ,…,A N(t+29) ];

[0016] Among them, A nt This represents the hitting characteristics of the nth joint at time t in frame 0. The frame number ranges from {-30,-29,…,0,…,29}, and m is an integer greater than 2.

[0017] Two shot data points are randomly selected and combined to obtain m / 2 combination data points. The optimal path distance for each combination is calculated using the dynamic time warp algorithm.

[0018] Furthermore, the optimal path distance for each combination is calculated using the dynamic time warp algorithm, expressed as:

[0019]

[0020] Among them, DTW(M i N j M represents a shot feature sequence of length i. i and the shot feature sequence N of length j j The optimal path distance between them; d(M[i], N[j]) represents the distance between the i-th element M[i] in the ball feature sequence M and the j-th element N[j] in the hitting feature sequence N. Each element is composed of features collected by sensors such as the triaxial angular velocity, triaxial velocity and triaxial coordinate at the joint node.

[0021] Furthermore, the weighted stability is expressed as:

[0022]

[0023] Among them, S tab For weighted stability; ω θ S represents the weight of the θ-th set of joint nodes; θ Let θ be the striking stability of the θ-th set of joint nodes; Let M be the DTW distance of the set of joint nodes in the θ-th group of the r-th hitting sequence. If m hitting action sequences are collected, the hitting sequences are combined pairwise to obtain m / 2 groups of hitting sequences. i N j This represents two shot sequences within a set of shot sequences.

[0024] Furthermore, the weight ω of the θ-th group of joint nodes θ Represented as:

[0025]

[0026] Among them, Q θ Let |Q| represent the set of the θ-th joint nodes, where k is the corresponding joint node. θ |For set Q θ Number of elements in the middle; I(A) k C) represents the mutual information between the characteristics of the shot sequence and the shot category, A k Let C represent the set of features measured by the IMU sensor in the shot sequence, and let C represent the set of shot types.

[0027] Furthermore, the mutual information I(A) between the shot sequence features and the shot category k c) is represented as:

[0028]

[0029] Where p(x, y) represents the joint probability distribution of feature x and shot type label y; p(x) represents the probability distribution of feature x; and p(y) represents the probability distribution of shot type label y.

[0030] Furthermore, the final stability evaluation score is expressed as follows:

[0031] S sorce =100-93*S tab 0.96

[0032] Among them, S sorce The final stability evaluation score; S tab For weighted stability.

[0033] Compared with existing technologies, the present invention can more accurately identify tennis motions and provide more accurate assessment results of shot stability. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the process for preparing a tennis hitting motion dataset according to the present invention;

[0035] Figure 2 This is a schematic diagram of the tennis motion recognition method of the present invention;

[0036] Figure 3 This is a schematic diagram of the method for improving ball striking stability according to the present invention;

[0037] Figure 4 This is a schematic diagram of the motion capture suit of the present invention being worn;

[0038] Figure 5 This is a schematic diagram of the main joints of the human body in this invention;

[0039] Figure 6 This is a schematic diagram showing the grouping of the main joints of the human body according to the present invention;

[0040] Figure 7 This is a schematic diagram of the architecture of a tennis ball recognition device according to the present invention. Detailed Implementation

[0041] 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.

[0042] This invention proposes a method for evaluating the stability of tennis shots based on inertial motion capture, specifically including the following steps:

[0043] S1. Use an inertial motion capture kit to capture the three-dimensional coordinates of multiple joints and the inertial data measured by inertial sensors in each frame when the trainee hits the ball under different types of conditions.

[0044] S2. Based on the set number of joints, construct a spatiotemporal graph of the skeleton sequence of joint points and T frames. This spatiotemporal graph represents the three-dimensional spatial information of each joint within a frame and the temporal information of node changes between frames.

[0045] S3. Construct an action recognition network based on a graph neural network, and use this network to identify the type of ball hit;

[0046] S4. After identifying the user's hitting type, randomly select two hitting actions of the same type and use the dynamic time warp algorithm to calculate the optimal path distance between the two hitting actions.

[0047] S5. Group the human joint nodes, calculate the stability and weight of each group of joint nodes, and use the weight to weight the stability of the group of joint nodes.

[0048] S6. Convert the obtained weighted stability to a percentage system to obtain the final stability evaluation score.

[0049] In this embodiment, an implementation method for collecting subject motion data is given, such as... Figure 1 Specifically, it includes the following steps:

[0050] S101 uses an inertial motion capture suit to capture the three-dimensional coordinates of multiple joints and the data measured by inertial sensors in each frame when the trainee hits the ball in different types of situations.

[0051] Specifically, taking the forehand stroke as an example, the inertial kinetic capture device is worn with reference to the attached... Figure 4 The diagram is worn by the trainee. Inertial motion capture is used to collect the three-axis coordinates of each major joint of the trainee relative to the Earth coordinate system, including the three-axis angular velocities, three-axis velocities, and three-dimensional spatial coordinates of each joint. The data for the I-th joint is represented as AI = [gyro-x, gyro-y, gyro-z, velo-x, velo-y, velo-z, posi-x, posi-y, posi-z], where gyro-x represents the angular velocity along the x-axis, gyro-y represents the angular velocity along the y-axis, and gyro-z represents the angular velocity along the z-axis; velo-x represents the velocity along the x-axis, velo-y represents the velocity along the y-axis, and velo-z represents the velocity along the z-axis; posi-x represents the spatial coordinates along the x-axis, posi-y represents the spatial coordinates along the y-axis, and posi-y represents the spatial coordinates along the z-axis; I = 1, 2, 3, ..., N, where N is the number of selected major joints. In this embodiment, N is 23. Specific major joints of the human body are shown in the attached diagram. Figure 5 As shown.

[0052] S102 uses a peak detection algorithm to detect the combined velocity collected at the wrist joint of the racket-holding hand to obtain the moment of impact, and extracts the coordinates of each joint point captured in 60 frames before and after the moment of impact.

[0053] Based on sports biomechanics theory, the velocity of the racket-holding hand reaches its maximum at the instant of impact. Therefore, the resultant velocity Stro is calculated based on the collected triaxial velocity components of the racket-holding hand. it The formula is as follows:

[0054]

[0055] The resulting combined velocity is multi-frame data that changes continuously over time. Therefore, the peak detection algorithm is used to detect the frame number of the moment of impact.

[0056] Using the frame number of the moment of impact as a baseline, data from the first 30 frames and the last 30 frames are extracted using a fixed window. This yields all the data for one impact.

[0057] The joint 3D coordinates of each frame in the measured data are extracted, and the extracted 3D coordinates are represented as [m1, m2, ..., m 60 ], m i Let m represent the coordinates of the i-th frame (i = {1, 2, ..., 60}). Each frame includes the coordinates of all joint nodes. In this embodiment, there are 23 joint nodes. The coordinates of the i-th frame are m. i Represented as [m i1 ,m i2 ,…,m i23 ], m ij This represents the coordinates of the j-th joint node in the i-th frame (j = {1, 2, ..., 23}).

[0058] S103, which uses the trainee's hip joint as the origin of the coordinate system and normalizes all other joints to obtain the root joint relative coordinates of the main joints.

[0059] Specifically, the relative coordinate values ​​are obtained by subtracting the hip joint's coordinates from the three-dimensional coordinates of each joint point, and the three-dimensional coordinates of the hip joint are set to zero. After combining these values, the normalized coordinates of the main joints of the whole body are obtained. The 60 frames of data during the shot are then processed using this method to obtain a tensor with dimensions [60, 23, 3], where 60 is the number of frames selected in this embodiment, 23 is the number of joint nodes selected in this embodiment, and 3 represents the x, y, and z axes of the joint nodes in the spatial coordinate system.

[0060] S104: The processed joint timing data are merged into a tensor and filled into a dictionary. The data name and action type label are also filled into the dictionary. Finally, the dictionary is generated into a JSON file.

[0061] The dataset includes six action types: forehand stroke, serve, two-handed backhand stroke, one-handed backhand stroke, overhead smash, volley, and slice. The strokes are labeled sequentially as follows: forehand stroke-1, serve-2, two-handed backhand stroke-3, one-handed backhand stroke-4, overhead smash-5, volley-6, and slice-7. Files are named according to the action type and its corresponding number, such as forehand_1 and serve_01. The data is written into a dictionary in the order of name, keypoint tensor data, and action label, ultimately generating a JSON file.

[0062] After collecting the shot data, it is necessary to identify the shot type, such as... Figure 2 In this embodiment, the process of identifying the ball hit includes the following steps:

[0063] S201, In this invention, the input graph structure of the action recognition model is constructed based on the human skeletal joint diagram structure. A spatiotemporal graph G = (V, E) of the skeletal sequence with N nodes (23 joint nodes in this embodiment) and T frames (60 frames in this embodiment) is constructed. This spatiotemporal graph can represent the three-dimensional spatial information of each joint within a frame and the temporal information of node changes between frames.

[0064] Specifically, the spatiotemporal graph is represented as G = (V, E), where V represents node features and E represents edge features. The node features of the spatiotemporal graph constructed in this invention are V = {Node1, Node2, ..., Node...}. n}, 23 skeletal joints are used to construct a spatiotemporal graph structure. A single frame contains 23 nodes, and there are a total of T frames of information. Therefore, a node is represented as V = {v ti |t=1,...,T;i=1,...,N},v ti Let T represent the feature of the i-th joint in the t-th frame, where T is the data duration and N is the total number of human joints used.

[0065] The edge feature edge E = {E} of the spacetime graph constructed in this invention S E F In the same frame, joints are connected to form edges according to the order of human joint connection. Considering that the joints at the ends of the human body have a significant impact on the motion recognition results, edges are also constructed and connected to the ends of each limb. The spatial domain information representing the combined edge is E. S This refers to the connection of joints in the human body within a single frame; connecting joints at the same position between consecutive frames creates an edge connection, where the temporal information of the edge is represented by E. FThat is, the temporal information of consecutive frames. E S The specific representation is as follows:

[0066] E S ={v ti v tj |(i,j)∈H}

[0067] Among them, v ti v tj H represents the edge (i, j) between joint point i and joint point j in frame t, and H is a set of naturally connected human joints and end joints connection information.

[0068] E F The specific representation is as follows:

[0069] E F ={v ti v (t+1)i}

[0070] Among them, v ti v (t+1)i This represents the edge of joint i from frame t to frame t+1.

[0071] The final result is a graph containing both temporal and spatial information. The node information is (x, y, z), representing the node's position in the spatial coordinate system; the edge information is (E...). S E F ).

[0072] In summary, the spatiotemporal graph construction method based on skeleton sequences proposed in this invention can realize the collection of skeleton data in both time and space dimensions.

[0073] S202, the action recognition method based on graph neural networks constructed in this invention can realize the modeling of the temporal and spatial information of skeletal data, thereby improving the accuracy of tennis hitting action recognition.

[0074] Specifically, in this embodiment, the action recognition network based on graph neural network comprises a normalization layer, a first GCN layer, a first TCN layer, a first activation function ReLU, a second GCN layer, a second TCN layer, a second activation function ReLU, a third GCN layer, a third TCN layer, a third activation function ReLU, a global pooling layer, and a fully connected layer connected in sequence. The outputs of the first TCN layer and the first GCN layer have 64 channels, the outputs of the second TCN layer and the second GCN layer have 128 channels, and the outputs of the third TCN layer and the third GCN layer have 256 channels.

[0075] This embodiment uses the backpropagation algorithm to adjust the weights of the graph neural network-based action recognition network, employs the SGD optimizer as the optimization function, uses cross-entropy as the loss function, and evaluates the recognition results based on accuracy (Acc). The accuracy (Acc) is calculated as follows:

[0076]

[0077] Wherein, TP indicates that the predicted class is true and the actual class is true; FP indicates that the predicted class is false and the actual class is true; TN indicates that the predicted class is true and the actual class is false; and FN indicates that the predicted class is false and the actual class is false.

[0078] S203 uses the training dataset for training, randomly dropping features with a probability of 0.5; the learning rate is set to 0.01, decreasing by 0.1 every 10 epochs; the batch size for training samples is set to 64. Specifically, at the start of training, the weights are initialized, with He initialization selected for parameter initialization. Training data, the model, and the optimizer are loaded. Training ends when the cross-entropy loss function output by the last softmax classification layer decreases and converges, retaining the trained model weights.

[0079] The process for stability evaluation of each identified action in this embodiment is as follows: Figure 3 Specifically, it includes the following steps:

[0080] S301, the data from m consecutive shots performed by the trainee using the same hitting motion are processed according to the method described in S101, with the data segmented into a fixed window and joint coordinates normalized. The trainee's single shot data is represented as M = [A...]. 1(t -30),…,A n(t -30),…,A 1t ,…,A nt ,…,A 1(t+29) ,…,A N(t+29) The data is collected in M ​​= [A1, A2, ..., A], where N represents the total number of joints in the body, t represents the moment of impact, and 60 frames of data are collected before and after that moment. Specifically, during implementation, the index is changed to M = [A1, A2, ..., A]. N*60 ], where N represents the total number of joints, A N*60 The joint point data collected by the inertial sensor in S101 represents the data of the Nth joint node in the 60th frame. After changing the index, the i-th data in sequence M is represented as... in Let x, y, z represent the angular velocities of the i-th data point along the x-axis, y-axis, and z-axis, respectively.

[0081] S302, Based on the principle of random division of the obtained m stroke data, randomly select two strokes as a group, and any two groups of strokes M i M j Where (i≠j), a total of m / 2 pairs of shots M are obtained. i M j The data is processed using the algorithm in S303 for each pair of data. In this invention, m is assumed to be an even number greater than or equal to 2. At least two hitting sequences are required to form a set of hitting sequences for path recognition.

[0082] S303 uses the dynamic time warp algorithm to calculate the optimal path distance between two shots. The algorithm is as follows:

[0083]

[0084] Among them, DTW(M i N j d(M[i],N[j]) represents the optimal path distance between the i-th element M[i] in the feature sequence M and the j-th element N[j] in the feature sequence N, and the calculation formula is as follows:

[0085]

[0086] in, Let M[i] represent the nth feature, which includes the triaxial angular velocity, triaxial velocity and triaxial coordinates at the joint nodes; Let N[j] represent the nth feature.

[0087] S304 (This step is optional) According to the principles of tennis kinematics, different body parts cooperate with each other to complete the hitting action during the hitting process, and each joint contributes to the hitting result to a different degree.

[0088] Specifically, to more scientifically assess a player's hitting stability, the body's 17 joints are divided into 3 groups, and the value of each data point after the division is represented as follows:

[0089] data 1 =[A 1(t-30) ,...,A 8(t-30) A lt ,...,A 8t ,...,A 1(t+29) A 8(t+29) ]

[0090] data 2 =[A9(t-30) ,...,A 12(t-30) ,...,A 9t ,...,A 12t ,...,A 9(t+29) A 12(t+29) ]

[0091] data 3 =[A 15(t-30) ,...,A 23(t-30) ,...,A 15t ,...,A 23t ,...,A 15(t+29) ,...,A 23(t+29) ]

[0092] In this embodiment, the joints of the arm holding the racket are grouped into one group, the joints of the lower body of the subject are grouped into another group, and all other joints except those of the head and neck are grouped into a third group.

[0093] The consistency of movement is calculated for each joint combination in the body, and the final combined result is the following formula for calculating the stability of the shot:

[0094]

[0095]

[0096] Where n = m / 2, m is the number of the same type of action, which is grouped into pairs to obtain n groups, each group including two sequences of the same type of hitting action; S θ This represents the similarity of the θ-th joint combination, where θ takes values ​​from 1 to 3. A schematic diagram of body joint combinations is shown below. Figure 6 As shown; ω θ The weight parameter represents the weight of the θ-th joint combination. Specifically, in this embodiment, the mutual information between features and labels is used to describe it.

[0097] Let the feature set of dataset D be Data = {A1, A2, ..., A...} n}, where n is the number of features in the dataset, A i Let X represent the i-th feature, and C represent the labels of the dataset. Then the mutual information formula for two random variables X and Y is defined as:

[0098]

[0099] Therefore, the joint combination weight parameter ω in this application θ Defined as:

[0100]

[0101] Among them, Qθ Let |Q| represent the set of joints contained in the θ-th joint combination. θ | Indicates the size of the set; A k Describe the set Q θ The characteristics of the k-th joint. I(A) k ;C) represents the mutual information between the characteristics of the shot sequence and the shot category, where C represents the shot type.

[0102] S305, for the S obtained in the previous step tab The final motion stability assessment score is obtained by converting the score to a percentage. The calculation formula is as follows:

[0103] S sorce =100–93×S tab 0.96

[0104] The proposed method will be analyzed below with a specific example.

[0105] The ball-hitting action recognition method proposed in this invention classifies the types of ball-hitting actions of trainees. In specific implementation, it is necessary to first filter out 60 complete action frames based on the moment of hitting the ball, process the skeletal joint data into a spatiotemporal graph structure, and then input it into the proposed graph neural network model for ball-hitting action classification. Table 1 shows the accuracy (Acc) for the classified actions.

[0106] Table 1

[0107]

[0108] The method for evaluating the stability of a hitting motion proposed in this invention assesses the stability of a trainee's hitting motion. In practice, the same incoming tennis ball speed and direction are set, and the trainee hits the ball to collect data. This data is then processed and input into a stability calculation method for evaluation. This invention selected six trainees of different skill levels to participate in the experiment: two beginners, two intermediate players, and two professionals. Taking the forehand hitting motion as an example, Table 2 shows the hitting stability of players at different skill levels.

[0109] Table 2

[0110] level Beginner 1 Beginner 2 Beginner 1 Beginner 2 Professional 1 Professional 2 stability 36 29 58 63 88 91

[0111] Corresponding to the specific implementation methods described above, refer to Figure 7 This invention discloses a device system for comprehensively and accurately identifying tennis hitting motion types and assessing stability, comprising:

[0112] Acquisition module 401: Based on the inertial motion capture suit, it acquires different hitting actions to obtain the three-dimensional coordinates of the multi-joint points of the trainee in each frame when hitting the ball and the data collected by the inertial sensor;

[0113] Detection module 402: Based on the peak detection algorithm, it detects the combined velocity collected at the wrist joint of the trainee's racket-holding hand to obtain the moment of impact, and extracts the coordinates of each joint point captured in 60 frames before and after the moment of impact.

[0114] Processing module 403: Based on the trainee's hip joint as the origin of the coordinate system, normalize the other joint points and generate a JSON file according to the above dataset preparation method;

[0115] Classification module 404: Identifies the action type corresponding to the joint coordinate data of each frame during the hitting period using a graph neural network-based method;

[0116] Evaluation module 405: Based on the DTW algorithm, the data generated during two hitting periods are compared and calculated, and the average is used to obtain the stability of the trainee's hitting.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the stability of tennis shots based on inertial motion capture, characterized in that, Specifically, the following steps are included: S1. Use an inertial motion capture kit to capture the three-dimensional coordinates of multiple joints and the inertial data measured by inertial sensors in each frame when the trainee hits the ball under different types of conditions. S2. Based on the set number of joints, construct a spatiotemporal graph of the skeleton sequence of joint points and T frames. This spatiotemporal graph represents the three-dimensional spatial information of each joint within a frame and the temporal information of node changes between frames. S3. Construct an action recognition network based on a graph neural network, and use this network to identify the type of ball hit; S4. After identifying the user's shot type, randomly select two shots of the same type and use the dynamic time warp algorithm to calculate the optimal path distance between the two shots, including: Collect m shots of the same shot type from the trainee, selecting 60 frames of data for each shot. The data for the i-th shot is represented as follows: M=[A 1(t-30) ,…,A n(t-30) ,…,A 1t ,…,A nt ,…,A 1(t+29) ,…,A N(t+29) ]; Among them, A nt This represents the hitting characteristics of the nth joint at time t in frame 0. The frame number ranges from {-30,-29,…,0,…,29}, and m is an integer greater than 2. Two shot data points are randomly selected and combined to obtain n combinations. The optimal path distance for each combination is calculated using the dynamic time warp algorithm, and is represented as: in, Represents a shot feature sequence of length i and a shot feature sequence of length j The optimal path distance between them; Represents the i-th element in the spherical feature sequence M The j-th element in the shot feature sequence N The distance between them; S5. Group the human joint nodes, calculate the stability and weight of each group of joint nodes, and use the weights to weight the stability of the joint nodes in the group; the weighted stability is expressed as: in, For weighted stability; Let θ be the striking stability of the θ-th set of joint nodes; Let be the DTW distance of the set of joint nodes in the θ-th joint node set in the r-th hitting sequence. If m hitting action sequences are collected, and the hitting sequences are combined pairwise, then... Group shot sequence, , This represents two shot sequences within a set of shot sequences; Weights of the θ-th set of joint nodes Represented as: in, Indicates the first A set of joint nodes For the corresponding joint points, For set Number of elements in the middle; This represents the set of features measured by the IMU sensor in the shot sequence. Represents the set of shot types; Mutual information between shot sequence characteristics and shot type Represented as: in, This represents the joint probability distribution of feature x and shot type label y; This represents the probability distribution of feature x; This represents the probability distribution of the shot type label y; S6. Convert the obtained weighted stability to a percentage system to obtain the final stability evaluation score, expressed as: S sorce = 100 – 93× 0.96 Among them, S sorce The final stability evaluation score; For weighted stability.

2. The method for evaluating tennis shot stability based on inertial motion capture according to claim 1, characterized in that, The inertial data, which measures the three-dimensional coordinates of a joint node by an inertial sensor, includes the x-axis angular velocity, y-axis angular velocity, z-axis angular velocity, x-axis velocity, y-axis velocity, z-axis velocity, x-axis spatial coordinates, y-axis spatial coordinates, and z-axis spatial coordinates at that joint node.

3. The method for evaluating tennis shot stability based on inertial motion capture according to claim 1, characterized in that, The graph neural network-based action recognition network consists of a normalization layer, a first GCN layer, a first TCN layer, a first activation function ReLU, a second GCN layer, a second TCN layer, a second activation function ReLU, a third GCN layer, a third TCN layer, a third activation function ReLU, a global pooling layer, and a fully connected layer, connected in sequence. The outputs of the first TCN layer and the first GCN layer have 64 channels, the outputs of the second TCN layer and the second GCN layer have 128 channels, and the outputs of the third TCN layer and the third GCN layer have 256 channels.

Images