Football motion recognition method and system based on deep learning

Abstract

The invention discloses a football motion recognition method and system based on deep learning. The method comprises the steps of data measurement and acquisition, optimized data processing, sensing motion recognition, motion strategy optimization and adaptive feedback adjustment. The invention relates to the technical field of football motion analysis, in particular to a football motion recognition method and system based on deep learning, and the scheme adopts a method of combining basic preprocessing with dynamic time window feature extraction and fusion to carry out data optimization processing; constructing a recognition model by adopting an output activation removal bidirectional long-short-term memory neural network combined with a self-attention mechanism; an integrated differential evolution algorithm combined with reinforcement learning is adopted, an initial prediction model is optimized, and an action strategy optimization model is further obtained through training.

Description

Technical Field

[0001] The present invention relates to the technical field of football motion analysis, and specifically to a football motion recognition method and system based on deep learning. Background Art

[0002] The deep learning-based football motion recognition method uses sensor data, such as motion data collected by accelerometers, gyroscopes and other devices, and analyzes these sensor data through a deep learning model. This method can accurately identify and classify sensor indicators of various actions in football, such as passing, shooting, running, etc. Its function is to improve the analysis efficiency of athlete training and competition, and provide a scientific basis for tactical formulation, sports performance evaluation and injury prevention.

[0003] However, among the existing methods for football motion recognition, traditional methods often focus on image motion recognition, while football motion recognition based on sensor data faces the technical problem of poor quality of original data, which affects the accuracy of subsequent prediction tasks; in the existing football motion recognition, the data of football motion changes frequently and in actual motion recognition, the data source must consider the real-time motion process, and thus the continuity of the results must be guaranteed, while the traditional time series data prediction model, the bidirectional long short-term memory neural network, generally uses a nonlinear activation function for activation at output. This method improves the accuracy but affects the continuity of the final result, resulting in a technical problem that the availability of the prediction result is reduced; in the existing football motion recognition, there is a technical problem that a simple sensor motion recognition method only predicts the results of the sensor motion, and cannot provide intuitive support for motion recognition and even the entire process of football motion, thereby reducing the referenceability of the method itself. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a method and system for football motion recognition based on deep learning. In view of the fact that in the existing football motion recognition methods, traditional methods often focus on image motion recognition, while football motion recognition based on sensor data faces the technical problem of poor original data quality, which affects the accuracy of subsequent prediction tasks, this scheme creatively adopts basic preprocessing combined with dynamic time window feature extraction and fusion methods to perform data optimization processing, and improves the class balance and data quality of the data through noise enhancement, translation and scaling, and further improves the quality of feature data through dynamic time window feature detection extraction and subsequent feature fusion, providing a good data foundation for subsequent prediction; in view of the fact that in the existing football motion recognition, the data of football motion changes frequently and in actual motion recognition, the data source must consider the real-time motion process, and thus the continuity of the results must be guaranteed, while the traditional time series data prediction model bidirectional long short-term memory neural network, in the input When the prediction is performed, a nonlinear activation function is generally used for activation. This method improves the accuracy but affects the continuity of the final result, resulting in a technical problem that the availability of the prediction result is reduced. This solution creatively removes the output activation of the bidirectional long short-term memory neural network combined with the self-attention mechanism, optimizes the feature capture capability by combining the self-attention mechanism, and improves the continuity of the prediction result and the overall availability of the method by removing the activation function of the output layer; in view of the technical problem that in the existing football motion recognition, there is a simple sensor motion recognition method that only predicts the result of the sensor motion, and cannot intuitively provide support for the motion recognition and even the entire process of football, thereby reducing the reference of the method itself, this solution creatively adopts an integrated differential evolution algorithm combined with reinforcement learning, which not only optimizes the initial prediction model, but also further trains the action strategy optimization model, so that it can adaptively adapt the corresponding action plan reference and support for the results of football motion recognition, making the output of the entire method more intuitive and reliable.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a method for recognizing football motions based on deep learning, the method comprising the following steps:

[0006] Step S1: data measurement and collection;

[0007] Step S2: Optimizing data processing;

[0008] Step S3: sensing action recognition;

[0009] Step S4: action strategy optimization;

[0010] Step S5: Adaptive feedback adjustment.

[0011] Furthermore, in step S1, the data measurement and collection is used to collect the basic data set required for football motion recognition, specifically by measuring data through sensors and inertial measurement devices, and collecting raw data for football motion recognition;

[0012] The sensors specifically include a plantar pressure sensor and an electromyography sensor;

[0013] The inertial measurement device is specifically used to detect acceleration and angular velocity data during football movement;

[0014] The football motion recognition raw data specifically includes plantar pressure data, electromyography data and posture motion data.

[0015] Furthermore, in step S2, the optimized data processing is used to perform data optimization processing on the original data, specifically, to perform data enhancement and time window enhancement on the original data of football motion recognition to obtain optimized football motion data.

[0016] Further, in step S3, the sensor action recognition is used to construct a deep learning model for football motion recognition applied to the sensor data, specifically, based on the optimized football action data, an output activation removal bidirectional long short-term memory neural network combined with a self-attention mechanism is used to construct a prediction model to obtain an initial model for football action recognition;

[0017] The output activation removal long short-term memory neural network combined with the self-attention mechanism specifically includes a bidirectional long short-term memory layer, an attention layer, and an optimized output layer;

[0018] The bidirectional long short-term memory layer is used as a basic model for data prediction;

[0019] The attention layer is used to optimize the learning effect of the gait features of football movements;

[0020] The optimized output layer is used to remove the nonlinear activation function of the output layer and perform continuity optimization of the prediction results;

[0021] The step of using the output activation removal of the long short-term memory neural network combined with the self-attention mechanism to construct a prediction model to obtain an initial model for football action recognition includes:

[0022] Step S31: constructing a bidirectional long short-term memory layer, specifically by respectively constructing a forward long short-term memory subnet and a backward long short-term memory subnet, performing bidirectional hidden state calculation and merging, and obtaining a merged hidden state output;

[0023] Step S32: construct an attention layer, specifically including the following steps:

[0024] Step S321: Calculate the attention score;

[0025] Step S322: normalizing attention weights;

[0026] Step S323: Calculate the context vector;

[0027] Step S324: constructing an attention layer, specifically, optimizing the performance of the model in capturing features from time series data by calculating the attention score, normalizing the attention weight, and calculating the context vector, so as to construct the attention layer;

[0028] Step S33: constructing an optimized output layer to enhance the temporal continuity of football motion recognition, specifically by removing the nonlinear activation function in the output gate of the traditional bidirectional long short-term memory subnet to optimize the output layer;

[0029] Step S34: Sensor action recognition model training, specifically, by constructing a bidirectional long short-term memory layer, constructing an attention layer, and constructing an optimized output layer, the sensor action recognition model is trained to obtain a football action recognition initial model Model OP ;

[0030] Step S35: football motion recognition, specifically by obtaining the football motion recognition initial model Model OP , perform preliminary football action recognition and obtain initial football action recognition data;

[0031] The soccer action recognition initial data specifically includes plantar pressure prediction data, electromyography prediction data, posture movement prediction data, weight distribution prediction data and gait parameter data.

[0032] Further, in step S4, the action strategy optimization is used to optimize the action strategy for the prediction model, specifically based on the football action recognition initial model Model OP The football action recognition optimization model Model is obtained by using the integrated differential evolution algorithm combined with reinforcement learning to optimize the model parameters and action strategy. PR And action strategy optimization model Model CP , specifically including the following steps:

[0033] Step S41: constructing an integrated differential evolution algorithm, including the following steps:

[0034] Step S411: Initialize population parameters, specifically by randomly generating an initial population to represent candidate solutions of model parameters and action strategies. The calculation formula for randomly generating an initial population is:

[0035] ;

[0036] In the formula, P is the initial population, X i is the i-th initial population individual, (0) is the initialization identifier, which is used to indicate that the number of iterations is 0, i is the population individual index, N is the total number of population individuals, is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy;

[0037] Step S412: Calculate individual fitness, specifically calculate the fitness function value of each population individual, the calculation formula is:

[0038] ;

[0039] In the formula, F(X i ) is the individual fitness of the i-th initial population individual, acc(·) is the accuracy evaluation function, eff(·) is the effectiveness evaluation function, and the specific calculation values ​​of the accuracy evaluation function and the effectiveness evaluation function are specifically calculated by applying the model parameter candidate solution and the action strategy candidate solution to the football action recognition initial model Model OP , after prediction, we get is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy;

[0040] Step S413: Differential evolution variant individual generation, for each population individual, randomly selects three different population individuals as material population individuals, and generates variant individuals, the calculation formula is:

[0041] ;

[0042] Where V i is the i-th mutant individual, i is the population individual index, X a is the number of individuals X in the i-th initial population i The first material population individual, FV is the variation factor, specifically a random number in the range of [0,2], X b It is an individual of the second material population, X c It is an individual of the third material population;

[0043] Step S414: integrated differential evolution crossover, specifically generating crossover candidate individuals based on the mutated individuals, the calculation formula is:

[0044] ;

[0045] Where U i is the i-th crossover candidate individual, V iis the ith mutant individual, X i is the i-th individual of the initial population, CR is the crossover probability, and j is a random number in the range [0,1];

[0046] Step S415: integrated differential evolution selection, specifically, re-calculating the variant individuals and crossover candidate individuals according to the operation in step S412, and selecting individuals with higher fitness values ​​as new initial population individuals for iterative calculation;

[0047] Step S416: iterative solution, specifically setting a maximum number of iterations and performing iterative solution until a convergence condition is reached or the maximum number of iterations is reached, and the optimal population individual is used to represent the optimal individual model parameters and the optimal individual football action strategy;

[0048] Step S42: Reinforcement learning strategy optimization, specifically, taking the football motion problem as the basic environment of reinforcement learning, and adopting the Q learning optimization strategy to perform reinforcement learning training, so as to obtain the reinforcement learning environment and the reinforcement learning optimization strategy;

[0049] Step S43: Constructing reinforcement learning convergence judgment conditions, specifically setting a convergence threshold, and performing difference calculation on the fitness values ​​in iterative training. When the absolute value of the difference between the fitness function value of the previous iteration and the fitness function value of the current iteration is less than the set convergence threshold, it indicates that the convergence condition is met, otherwise it indicates that the convergence condition is not met, and iterative training continues;

[0050] Step S44: Calculate the optimal strategy output, which is used to combine the integrated differential evolution algorithm and the reinforcement learning strategy optimization to calculate the optimal solution for reinforcement learning;

[0051] Step S45: Optimizing the action recognition model, specifically, optimizing the initial model Model of the football action recognition by obtaining the optimal solution predicted by the reinforcement learning model OP Model optimization, and obtain the football action recognition optimization model Model PR ;

[0052] Step S46: Action strategy optimization model training, specifically, through the reinforcement learning strategy optimization, the construction of reinforcement learning convergence judgment conditions and the calculation of the optimal strategy output, the football action strategy optimization model training is performed to obtain the action strategy optimization model Model CP ;

[0053] Step S47: Action strategy optimization, specifically using the soccer action recognition optimization model Model PR, based on the original data of football motion recognition, the football motion recognition sensor data is predicted, and the action strategy is used to optimize the model Model CP , based on the football motion recognition sensor numerical data, a football motion optimization strategy is predicted.

[0054] Further, in step S5, the adaptive feedback adjustment is used to combine the action strategy optimization and the sensor action recognition model to perform football action recognition, and in actual use, perform action feedback and adjustment of adaptive football training, specifically using the football action recognition optimization model Model PR And action strategy optimization model Model CP ,According to the real-time collected original data of football motion recognition, sensor motion recognition and motion strategy optimization prediction are performed to obtain an adaptive football motion reference solution;

[0055] The adaptive soccer motion reference scheme specifically includes a visualized plantar pressure trend reference image, a visualized muscle activity gait phase image, a visualized posture change reference image and weight distribution control reference data.

[0056] The present invention provides a football motion recognition system based on deep learning, which includes a data measurement and acquisition module, an optimization data processing module, a sensor motion recognition module, a motion strategy optimization module and an adaptive feedback adjustment module;

[0057] The data measurement and acquisition module is used for data measurement and acquisition, and obtains the original data of football motion recognition through data measurement and acquisition, and sends the original data of football motion recognition to the optimization data processing module, the action strategy optimization module and the adaptive feedback adjustment module;

[0058] The optimized data processing module is used for optimizing data processing, obtaining optimized football motion data through optimizing data processing, and sending the optimized football motion data to the sensor motion recognition module;

[0059] The sensing action recognition module is used for sensing action recognition, obtaining an initial model for football action recognition through sensing action recognition, and sending the initial model for football action recognition to the action strategy optimization module;

[0060] The action strategy optimization module is used for action strategy optimization, and obtains a football action recognition optimization model and an action strategy optimization model through action strategy optimization, and sends the football action recognition optimization model and the action strategy optimization model to the adaptive feedback adjustment module;

[0061] The adaptive feedback adjustment module is used for adaptive feedback adjustment, and an adaptive football motion reference scheme is obtained through adaptive feedback adjustment.

[0062] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0063] (1) In the existing football motion recognition methods, traditional methods often focus on image motion recognition, while football motion recognition based on sensor data faces the technical problem of poor quality of original data, which affects the accuracy of subsequent prediction tasks. This scheme creatively adopts basic preprocessing combined with dynamic time window feature extraction and fusion methods to optimize data processing. Through noise enhancement, translation and scaling, the class balance and data quality of the data are improved. Through dynamic time window feature detection extraction and subsequent feature fusion, the quality of feature data is further improved, providing a good data foundation for subsequent predictions.

[0064] (2) In the existing football motion recognition, the data of football motion changes frequently. In the actual motion recognition, the data source must consider the real-time motion process, and the continuity of the results must be guaranteed. The traditional time series data prediction model, the bidirectional long short-term memory neural network, generally uses a nonlinear activation function to activate the output. This method improves the accuracy but affects the continuity of the final result, resulting in a technical problem that the availability of the prediction result is reduced. This solution creatively uses the output activation removal of the bidirectional long short-term memory neural network combined with the self-attention mechanism. By combining the self-attention mechanism to optimize the feature capture capability, and by removing the activation function of the output layer, the continuity of the prediction result is improved, and the overall availability of the method is improved;

[0065] (3) In view of the technical problem that in the existing football motion recognition, the simple sensor motion recognition method only predicts the results of the sensor action and cannot provide intuitive support for the action recognition and even the entire process of football movement, thereby reducing the reference value of the method itself, this scheme creatively adopts an integrated differential evolution algorithm combined with reinforcement learning, which not only optimizes the initial prediction model, but also further trains the action strategy optimization model, so that it can adaptively adapt the corresponding action plan reference and support for the results of football motion recognition, making the output of the entire method more intuitive and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A flowchart of a method for recognizing football motions based on deep learning provided by the present invention;

[0067] Figure 2 A schematic diagram of a football motion recognition system based on deep learning provided by the present invention;

[0068] Figure 3 A schematic diagram of the process of optimizing data processing in step S2;

[0069] Figure 4 This is a schematic diagram of the process of sensing action recognition in step S3;

[0070] Figure 5 Schematic diagram of the process of optimizing the action strategy in step S4.

[0071] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0073] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0074] Example 1, see Figure 1 The present invention provides a method for recognizing football motions based on deep learning, the method comprising the following steps:

[0075] Step S1: data measurement and collection;

[0076] Step S2: Optimizing data processing;

[0077] Step S3: sensing action recognition;

[0078] Step S4: action strategy optimization;

[0079] Step S5: Adaptive feedback adjustment.

[0080] Example 2, see Figure 1 and Figure 2In step S1, the data measurement and collection is used to collect the basic data set required for football motion recognition, specifically by measuring data through sensors and inertial measurement devices, and collecting raw data for football motion recognition;

[0081] The sensors specifically include a plantar pressure sensor and an electromyography sensor;

[0082] The inertial measurement device is specifically used to detect acceleration and angular velocity data during football movement;

[0083] The football motion recognition raw data specifically includes plantar pressure data, electromyography data and posture motion data.

[0084] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the optimized data processing is used to perform data optimization processing on the original data, specifically to perform data enhancement and time window enhancement on the original data of football motion recognition to obtain optimized football motion data, which specifically includes the following steps:

[0085] Step S21: random noise enhancement, which is used to add random noise to the original data to prevent the model from overfitting. Specifically, random noise data is added to the original data of football motion recognition to obtain noise enhancement data. The calculation formula is:

[0086] ;

[0087] In the formula, is the noise enhanced data, x is the original data of football motion recognition, is random noise data;

[0088] Step S22: data translation enhancement, which is used to perform translation processing on the basis of the noise enhanced data to enhance the data's ability to represent gait data at different spatial positions. Specifically, the original data of the football motion recognition is translated to obtain translation enhanced data. The calculation formula is:

[0089] ;

[0090] In the formula, is the translation enhancement data, t is the time step index, x is the original data of football motion recognition, is the translation time step parameter, used to calculate the translation time step;

[0091] Step S23: data scaling enhancement, which is used to perform scaling processing on the basis of the translation enhancement data to enhance the data's ability to represent gait signals of different amplitudes, specifically, scaling processing is performed on the original data of football motion recognition to obtain scaling enhancement data, and the calculation formula is:

[0092] ;

[0093] In the formula, is the scaled enhanced data, x is the original data of football motion recognition, is the scaling factor used to represent the translation augmentation data The zoom factor;

[0094] Step S24: data enhancement, specifically, performing data enhancement on the original data of football motion recognition by randomly combining and executing the random noise enhancement, the data translation enhancement and the data scaling enhancement, so as to obtain enhanced football motion data;

[0095] The enhanced football action data specifically includes enhanced pressure data, enhanced electromyography data and enhanced posture motion data;

[0096] Step S25: dynamic time window feature extraction, specifically, using dynamic time window technology to perform feature detection, and performing real-time feature detection during the feature extraction process, dynamically adjusting the time window length, and extracting gait cycle time feature data;

[0097] Step S26: feature fusion optimization, specifically, feature fusion of the enhanced pressure data, enhanced electromyography data and enhanced posture motion data in the enhanced football action data to obtain football action fusion feature data;

[0098] Step S27: optimizing data processing, specifically, performing optimized data processing of the original data by performing data enhancement processing and the feature fusion optimization to obtain optimized football action data;

[0099] The optimized football action data specifically includes optimized pressure feature data, optimized electromyography feature data, optimized posture motion feature data and weight distribution prediction fusion data.

[0100] By performing the above operations, in view of the fact that in the existing football motion recognition methods, traditional methods often focus on image motion recognition, while football motion recognition based on sensor data faces the technical problem of poor original data quality, which affects the accuracy of subsequent prediction tasks, this solution creatively adopts basic preprocessing combined with dynamic time window feature extraction and fusion methods to optimize data processing, improves the class balance and data quality of the data through noise enhancement, translation and scaling, and further improves the quality of feature data through dynamic time window feature detection extraction and subsequent feature fusion, providing a good data foundation for subsequent predictions.

[0101] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the sensor action recognition is used to construct a deep learning model for football motion recognition applied to sensor data. Specifically, based on the optimized football action data, a bidirectional long short-term memory neural network with output activation removal combined with a self-attention mechanism is used to construct a prediction model to obtain an initial model for football action recognition;

[0102] The output activation removal long short-term memory neural network combined with the self-attention mechanism specifically includes a bidirectional long short-term memory layer, an attention layer, and an optimized output layer;

[0103] The bidirectional long short-term memory layer is used as a basic model for data prediction;

[0104] The attention layer is used to optimize the learning effect of the gait features of football movements;

[0105] The optimized output layer is used to remove the nonlinear activation function of the output layer and perform continuity optimization of the prediction results;

[0106] The step of using the output activation removal of the long short-term memory neural network combined with the self-attention mechanism to construct a prediction model to obtain an initial model for football action recognition includes:

[0107] Step S31: constructing a bidirectional long short-term memory layer, specifically by respectively constructing a forward long short-term memory subnet and a backward long short-term memory subnet, performing bidirectional hidden state calculation and merging, and obtaining a merged hidden state output, including the following steps:

[0108] Step S311: construct a forward long short-term memory subnet, the calculation formula is:

[0109] ;

[0110] In the formula, is the output hidden state of the forward LSTM subnetwork at time t, t is the time step index, LSTM(·) is the basic LSTM subnetwork representation function, including input gate, forget gate and output gate, x t is a model input data sample, used to represent the optimized football action data, is the output hidden state of the forward LSTM subnetwork at time t-1;

[0111] Step S312: Construct a backward long short-term memory subnet, and the calculation formula is:

[0112] ;

[0113] In the formula, is the output hidden state of the backward LSTM subnetwork at time t, t is the time step index, and LSTM(·) is the basic LSTM subnetwork representation function x t is a model input data sample, used to represent the optimized football action data, is the output hidden state of the backward long short-term memory subnetwork at time t+1;

[0114] Step S313: Bidirectional hidden state merging, the calculation formula is:

[0115] ;

[0116] In the formula, H t is the merged hidden state output, is the output hidden state of the forward LSTM subnetwork at time t, is the output hidden state of the backward LSTM subnetwork at time t, where t is the time step index;

[0117] Step S32: construct an attention layer, specifically including the following steps:

[0118] Step S321: Calculate the attention score, the calculation formula is:

[0119] ;

[0120] In the formula, e t is the attention score, H t is the merged hidden state output, is the set of output hidden states of the bidirectional long-short term subnetwork, is the dot product operator;

[0121] Step S322: normalize the attention weight, and the calculation formula is:

[0122] ;

[0123] In the formula, is the normalized attention weight, exp(·) is the natural base function, e t is the attention score value, where t is the time step index, T is the total number of time steps, which is used to represent the total duration in the time series data, k is the normalized time step index, which is used to calculate the time-by-time summation of the time step weights, and e k is the normalized time step score value;

[0124] Step S323: Calculate the context vector, the calculation formula is:

[0125] ;

[0126] Where C is the context vector, which is used to represent the weighted sum of the combined hidden states output by the bidirectional long short-term memory layer, T is the total number of time steps, t is the time step index, is the normalized attention weight, H t is the merged hidden state output;

[0127] Step S324: constructing an attention layer, specifically, optimizing the performance of the model in capturing features from time series data by calculating the attention score, normalizing the attention weight, and calculating the context vector, so as to construct the attention layer;

[0128] Step S33: construct an optimized output layer to enhance the temporal continuity of football motion recognition. Specifically, the output layer is optimized by removing the nonlinear activation function in the output gate of the traditional bidirectional long short-term memory subnet. The calculation formula is:

[0129] ;

[0130] Where y is the predicted output without activation, W C is the output layer weight, C is the context vector, which is used to represent the weighted sum of the combined hidden states output by the bidirectional long short-term memory layer, and b is the output layer bias term;

[0131] Step S34: Sensor action recognition model training, specifically, by constructing a bidirectional long short-term memory layer, constructing an attention layer, and constructing an optimized output layer, the sensor action recognition model is trained to obtain a football action recognition initial model Model OP ;

[0132] Step S35: football motion recognition, specifically by obtaining the football motion recognition initial model Model OP , perform preliminary football action recognition and obtain initial football action recognition data;

[0133] The soccer action recognition initial data specifically includes plantar pressure prediction data, electromyography prediction data, posture movement prediction data, weight distribution prediction data and gait parameter data;

[0134] The plantar pressure prediction data is used to represent the predicted pressure value at the sensor position at the tth time step;

[0135] The electromyography prediction data is used to represent the electrical activity prediction signal in the muscle part at the t-th time step;

[0136] The posture motion prediction data is used to represent the joint angle, acceleration and angular velocity prediction data at the tth time step;

[0137] The weight distribution prediction data is used to represent the weight distribution prediction data of the key parts of the movement at the t-th time step;

[0138] The gait parameter data is used to represent the predicted values ​​of the step length, step frequency and gait cycle at the tth time step during the football movement.

[0139] By performing the above operations, in view of the fact that in the existing football motion recognition, the data of football motion changes frequently and in the actual motion recognition, the data source must consider the real-time motion process, and thus the continuity of the results must be guaranteed, and the traditional time series data prediction model bidirectional long short-term memory neural network generally uses a nonlinear activation function for activation at the output. This method improves the accuracy but affects the continuity of the final result, resulting in a technical problem that the availability of the prediction result is reduced. This solution creatively uses the output activation removal of the bidirectional long short-term memory neural network combined with the self-attention mechanism, optimizes the feature capture capability by combining the self-attention mechanism, and improves the continuity of the prediction result by removing the activation function of the output layer, and improves the overall availability of the method.

[0140] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the action strategy optimization is used to optimize the action strategy for the prediction model, specifically based on the football action recognition initial model Model OP The football action recognition optimization model Model is obtained by using the integrated differential evolution algorithm combined with reinforcement learning to optimize the model parameters and action strategy. PR And action strategy optimization model Model CP , specifically including the following steps:

[0141] Step S41: constructing an integrated differential evolution algorithm, including the following steps:

[0142] Step S411: Initialize population parameters, specifically by randomly generating an initial population to represent candidate solutions of model parameters and action strategies. The calculation formula for randomly generating an initial population is:

[0143] ;

[0144] In the formula, P is the initial population, X i is the i-th initial population individual, (0) is the initialization identifier, which is used to indicate that the number of iterations is 0, i is the population individual index, N is the total number of population individuals, is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy;

[0145] Step S412: Calculate individual fitness, specifically calculate the fitness function value of each population individual, the calculation formula is:

[0146] ;

[0147] In the formula, F(X i ) is the individual fitness of the i-th initial population individual, acc(·) is the accuracy evaluation function, eff(·) is the effectiveness evaluation function, and the specific calculation values ​​of the accuracy evaluation function and the effectiveness evaluation function are specifically calculated by applying the model parameter candidate solution and the action strategy candidate solution to the football action recognition initial model Model OP , after prediction, we get is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy;

[0148] Step S413: Differential evolution variant individual generation, for each population individual, randomly selects three different population individuals as material population individuals, and generates variant individuals, the calculation formula is:

[0149] ;

[0150] Where V i is the i-th mutant individual, i is the population individual index, X a is the number of individuals X in the i-th initial population i The first material population individual, FV is the variation factor, specifically a random number in the range of [0,2], X b It is an individual of the second material population, X c It is an individual of the third material population;

[0151] Step S414: integrated differential evolution crossover, specifically generating crossover candidate individuals based on the mutated individuals, the calculation formula is:

[0152] ;

[0153] Where U i is the i-th crossover candidate individual, V i is the ith mutant individual, X i is the i-th individual of the initial population, CR is the crossover probability, and j is a random number in the range [0,1];

[0154] Step S415: integrated differential evolution selection, specifically, re-calculating the variant individuals and crossover candidate individuals according to the operation in step S412, and selecting individuals with higher fitness values ​​as new initial population individuals for iterative calculation;

[0155] Step S416: iterative solution, specifically setting a maximum number of iterations and performing iterative solution until a convergence condition is reached or the maximum number of iterations is reached, and the optimal population individual is used to represent the optimal individual model parameters and the optimal individual football action strategy;

[0156] Step S42: Reinforcement learning strategy optimization, specifically, taking the football motion problem as the basic environment of reinforcement learning, and adopting the Q learning optimization strategy to perform reinforcement learning training, so as to obtain the reinforcement learning environment and the reinforcement learning optimization strategy;

[0157] The calculation formula of the reinforcement learning basic environment is:

[0158] ;

[0159] Where Environment is the reinforcement learning environment, S is the state space, A is the action space, R is the reward function, and T is the state transfer function;

[0160] The calculation formula for the reinforcement learning training using the Q learning optimization strategy is:

[0161] ;

[0162] Where Q(·) is the Q-learning optimization strategy function, is an assignment operator. Left is the updated reinforcement learning optimization policy value, the assignment operator Right , is the reinforcement learning optimization strategy value before updating, where s is the state parameter, a is the action parameter, is the learning rate parameter, the value range is [0,1], r is the immediate reward parameter, is a discount factor used to measure the importance of reward parameters, and its value range is [0,1]. is the parameter corresponding to the next state All next action parameters in The maximum Q learning optimization strategy value that can be obtained, is the next state parameter, is the next action parameter;

[0163] Step S43: Constructing reinforcement learning convergence judgment conditions, specifically setting a convergence threshold, and performing difference calculation on the fitness values ​​in iterative training. When the absolute value of the difference between the fitness function value of the previous iteration and the fitness function value of the current iteration is less than the set convergence threshold, it indicates that the convergence condition is met, otherwise it indicates that the convergence condition is not met, and iterative training continues;

[0164] Step S44: Calculate the optimal strategy output, which is used to combine the integrated differential evolution algorithm and the reinforcement learning strategy optimization to calculate the optimal solution for reinforcement learning. The calculation formula is:

[0165] ;

[0166] In the formula, is the optimal solution predicted by the reinforcement learning model, which is used to represent the optimal model parameters and the optimal action strategy. argmax is the function for finding the maximum value of the objective function. i is the i-th individual of the initial population, P is the initial population, is the total number of iterations, F(X i ) is the individual fitness of the i-th individual in the initial population;

[0167] Step S45: Optimizing the action recognition model, specifically, optimizing the initial model Model of the football action recognition by obtaining the optimal solution predicted by the reinforcement learning model OP Model optimization, and obtain the football action recognition optimization model Model PR ;

[0168] Step S46: Action strategy optimization model training, specifically, through the reinforcement learning strategy optimization, the construction of reinforcement learning convergence judgment conditions and the calculation of the optimal strategy output, the football action strategy optimization model training is performed to obtain the action strategy optimization model Model CP ;

[0169] Step S47: Action strategy optimization, specifically using the soccer action recognition optimization model Model PR , based on the original data of football motion recognition, the football motion recognition sensor data is predicted, and the action strategy is used to optimize the model ModelCP , based on the football motion recognition sensor numerical data, a football motion optimization strategy is predicted.

[0170] By performing the above operations, in order to address the technical problem that in the existing football motion recognition, a simple sensor motion recognition method only predicts the result of the sensor motion, and cannot provide intuitive support for motion recognition and even the entire process of football motion, thereby reducing the referenceability of the method itself, this solution creatively adopts an integrated differential evolution algorithm combined with reinforcement learning, which not only optimizes the initial prediction model, but also further trains the motion strategy optimization model, so that it can adaptively adapt the corresponding action plan reference and support for the results of football motion recognition, making the output of the entire method more intuitive and reliable.

[0171] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the adaptive feedback adjustment is used to combine the action strategy optimization and the sensor action recognition model to perform football action recognition, and in actual use, the action feedback and adjustment of adaptive football training are performed, specifically using the football action recognition optimization model Model PR And action strategy optimization model Model CP ,According to the real-time collected original data of football motion recognition, sensor motion recognition and motion strategy optimization prediction are performed to obtain an adaptive football motion reference solution;

[0172] The adaptive soccer motion reference scheme specifically includes a visualized plantar pressure trend reference image, a visualized muscle activity gait phase image, a visualized posture change reference image and weight distribution control reference data.

[0173] Embodiment 7, see Figure 1 and Figure 2 , this embodiment is based on the above embodiment, and a football motion recognition system based on deep learning provided by the present invention includes a data measurement and acquisition module, an optimization data processing module, a sensor motion recognition module, a motion strategy optimization module and an adaptive feedback adjustment module;

[0174] The data measurement and acquisition module is used for data measurement and acquisition, and obtains the original data of football motion recognition through data measurement and acquisition, and sends the original data of football motion recognition to the optimization data processing module, the action strategy optimization module and the adaptive feedback adjustment module;

[0175] The optimized data processing module is used for optimizing data processing, obtaining optimized football motion data through optimizing data processing, and sending the optimized football motion data to the sensor motion recognition module;

[0176] The sensing action recognition module is used for sensing action recognition, obtaining an initial model for football action recognition through sensing action recognition, and sending the initial model for football action recognition to the action strategy optimization module;

[0177] The action strategy optimization module is used for action strategy optimization, and obtains a football action recognition optimization model and an action strategy optimization model through action strategy optimization, and sends the football action recognition optimization model and the action strategy optimization model to the adaptive feedback adjustment module;

[0178] The adaptive feedback adjustment module is used for adaptive feedback adjustment, and an adaptive football motion reference scheme is obtained through adaptive feedback adjustment.

[0179] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0180] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0181] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A football motion recognition method based on deep learning, Features: The method comprises the following steps: Step S1: data measurement and collection, data measurement is performed through sensors and inertial measurement devices, and original data of football motion recognition is collected; Step S2: optimizing data processing, performing data enhancement and time window enhancement on the original data of football motion recognition to obtain optimized football motion data; Step S3: sensor action recognition, which is used to construct a deep learning model for football motion recognition applied to sensor data, specifically, based on the optimized football motion data, an output activation removal bidirectional long short-term memory neural network combined with a self-attention mechanism is used to construct a prediction model to obtain an initial model for football motion recognition; Step S4: action strategy optimization, which is used to optimize the action strategy for the prediction model, specifically, based on the football action recognition initial model and the football action recognition initial data, an integrated differential evolution algorithm combined with reinforcement learning is used to optimize model parameters and action strategy, so as to obtain a football action recognition optimization model and an action strategy optimization model, which specifically includes the following steps: Step S41: construct an integrated differential evolution algorithm; Step S42: optimize reinforcement learning strategies; Step S43: construct reinforcement learning convergence judgment conditions; Step S44: calculate the optimal strategy output; Step S45: optimize the action recognition model; Step S46: train the action strategy optimization model; Step S47: optimize the action strategy; Step S5: Adaptive feedback adjustment to obtain an adaptive football motion reference solution.

2. A football motion recognition method based on deep learning according to claim 1, Features: In step S3, the output activation of the self-attention mechanism is combined to remove the long short-term memory neural network, which specifically includes a bidirectional long short-term memory layer, an attention layer, and an optimized output layer; The bidirectional long short-term memory layer is used as a basic model for data prediction; The attention layer is used to optimize the learning effect of the gait features of football movements; The optimized output layer is used to remove the nonlinear activation function of the output layer and perform continuity optimization of the prediction results.

3. A football motion recognition method based on deep learning according to claim 2, Features: In step S3, the step of using the output activation removal long short-term memory neural network combined with the self-attention mechanism to construct a prediction model to obtain an initial model for football action recognition includes: Step S31: constructing a bidirectional long short-term memory layer, specifically by respectively constructing a forward long short-term memory subnet and a backward long short-term memory subnet, performing bidirectional hidden state calculation and merging, and obtaining a merged hidden state output; Step S32: construct an attention layer, specifically including the following steps: Step S321: calculating the attention score; step S322: normalizing the attention weight; step S323: calculating the context vector; step S324: constructing the attention layer, specifically, by calculating the attention score, normalizing the attention weight and calculating the context vector, optimizing the performance of the model in capturing features from the time series data, and constructing the attention layer; Step S33: constructing an optimized output layer to enhance the temporal continuity of football motion recognition, specifically by removing the nonlinear activation function in the output gate of the traditional bidirectional long short-term memory subnet to optimize the output layer; Step S34: Sensor action recognition model training, specifically, by constructing a bidirectional long short-term memory layer, constructing an attention layer, and constructing an optimized output layer, the sensor action recognition model is trained to obtain a football action recognition initial model Model OP ; Step S35: football motion recognition, specifically by obtaining the football motion recognition initial model Model OP , perform preliminary football action recognition and obtain initial football action recognition data; The initial data for football action recognition specifically includes plantar pressure prediction data, electromyography prediction data, posture movement prediction data, weight distribution prediction data and gait parameter data.

4. A method for recognizing football movements based on deep learning according to claim 3, Features: In step S41, the construction of the integrated differential evolution algorithm includes the following steps: Step S411: Initialize population parameters, specifically by randomly generating an initial population to represent candidate solutions of model parameters and action strategies. The calculation formula for randomly generating an initial population is: ； In the formula, P is the initial population, X i is the i-th initial population individual, (0) is the initialization identifier, which is used to indicate that the number of iterations is 0, i is the population individual index, N is the total number of population individuals, is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy; Step S412: Calculate individual fitness, specifically calculate the fitness function value of each population individual, the calculation formula is: ； In the formula, F(X i ) is the individual fitness of the i-th initial population individual, acc(·) is the accuracy evaluation function, eff(·) is the effectiveness evaluation function, and the specific calculation values ​​of the accuracy evaluation function and the effectiveness evaluation function are specifically calculated by applying the model parameter candidate solution and the action strategy candidate solution to the football action recognition initial model Model OP , after prediction, we get is the candidate solution of the i-th model parameter, is the candidate solution of the i-th action strategy; Step S413: Differential evolution variant individual generation, for each population individual, randomly selects three different population individuals as material population individuals, and generates variant individuals, the calculation formula is: ； Where V i is the i-th mutant individual, i is the population individual index, X a is the number of individuals X in the i-th initial population i The first material population individual, FV is the variation factor, specifically a random number in the range of [0,2], X b It is an individual of the second material population, X c It is an individual of the third material population; Step S414: integrated differential evolution crossover, specifically generating crossover candidate individuals based on the mutated individuals, the calculation formula is: ； Where U i is the i-th crossover candidate individual, V i is the ith mutant individual, X i is the i-th individual of the initial population, CR is the crossover probability, and j is a random number in the range [0,1]; Step S415: integrated differential evolution selection, specifically, re-calculating the variant individuals and crossover candidate individuals according to the operation in step S412, and selecting individuals with higher fitness values ​​as new initial population individuals for iterative calculation; Step S416: iterative solution, specifically setting a maximum number of iterations and performing iterative solution until a convergence condition is reached or the maximum number of iterations is reached. The optimal population individual is used to represent the optimal individual model parameters and the optimal individual football action strategy.

5. A method for recognizing football movements based on deep learning according to claim 4, Features: In step S42, the reinforcement learning strategy optimization is specifically to use the football motion problem as the reinforcement learning basic environment, and adopt the Q learning optimization strategy to perform reinforcement learning training to obtain the reinforcement learning environment and the reinforcement learning optimization strategy; In step S43, the reinforcement learning convergence judgment condition is constructed, specifically, a convergence threshold is set, and the fitness value in the iterative training is calculated by difference. When the absolute value of the difference between the fitness function value of the previous iteration and the fitness function value of the current iteration is less than the set convergence threshold, it indicates that the convergence condition is met, otherwise it indicates that the convergence condition is not met, and the iterative training continues; In step S44, the calculated optimal strategy output is used to combine the integrated differential evolution algorithm and the reinforcement learning strategy optimization to perform reinforcement learning optimal solution calculation; In step S45, the action recognition model is optimized, specifically by obtaining the optimal solution predicted by the reinforcement learning model, and performing the initial model model of the football action recognition. OP Model optimization, and obtain the football action recognition optimization model Model PR ; In step S46, the action strategy optimization model training is specifically to perform football action strategy optimization model training through the reinforcement learning strategy optimization, the construction of reinforcement learning convergence judgment conditions and the calculation of optimal strategy output, to obtain the action strategy optimization model Model CP ; In step S47, the action strategy optimization is specifically to use the football action recognition optimization model Model PR , based on the original data of football motion recognition, the football motion recognition sensor data is predicted, and the action strategy is used to optimize the model Model CP , based on the football motion recognition sensor numerical data, a football motion optimization strategy is predicted.

6. A method for recognizing football motions based on deep learning according to claim 5, Features: In step S5, the adaptive feedback adjustment is used to combine the action strategy optimization and the sensor action recognition model to perform football action recognition, and in actual use, the action feedback and adjustment of adaptive football training are performed, specifically using the football action recognition optimization model Model PR And action strategy optimization model Model CP ,According to the real-time collected original data of football motion recognition, sensor motion recognition and motion strategy optimization prediction are performed to obtain an adaptive football motion reference solution; The adaptive soccer motion reference scheme specifically includes a visualized plantar pressure trend reference image, a visualized muscle activity gait phase image, a visualized posture change reference image and weight distribution control reference data.

7. A method for recognizing football movements based on deep learning according to claim 6, Features: In step S1, the data measurement and collection is used to collect the basic data set required for football motion recognition, specifically, data measurement is performed through sensors and inertial measurement devices, and raw data for football motion recognition is collected; The sensors specifically include a plantar pressure sensor and an electromyography sensor; The inertial measurement device is specifically used to detect acceleration and angular velocity data during football movement; The football motion recognition raw data specifically includes plantar pressure data, electromyography data and posture motion data.

8. A football motion recognition system based on deep learning, used to implement a football motion recognition method based on deep learning as described in any one of claims 1 to 7, Features: It includes a data measurement and acquisition module, an optimization data processing module, a sensor action recognition module, an action strategy optimization module and an adaptive feedback adjustment module.

9. A football motion recognition system based on deep learning according to claim 8, Features: The data measurement and acquisition module is used for data measurement and acquisition, and obtains the original data of football motion recognition through data measurement and acquisition, and sends the original data of football motion recognition to the optimization data processing module, the action strategy optimization module and the adaptive feedback adjustment module; The optimized data processing module is used for optimizing data processing, obtaining optimized football motion data through optimizing data processing, and sending the optimized football motion data to the sensor motion recognition module; The sensing action recognition module is used for sensing action recognition, obtaining an initial model for football action recognition through sensing action recognition, and sending the initial model for football action recognition to the action strategy optimization module; The action strategy optimization module is used for action strategy optimization, and obtains a football action recognition optimization model and an action strategy optimization model through action strategy optimization, and sends the football action recognition optimization model and the action strategy optimization model to the adaptive feedback adjustment module; The adaptive feedback adjustment module is used for adaptive feedback adjustment, and an adaptive football motion reference scheme is obtained through adaptive feedback adjustment.

Images