A gesture recognition method based on improved tabu search optimization BP neural network

Abstract

This invention discloses a gesture recognition method based on an improved tabu search-optimized BP neural network, comprising: S1 initializing neural network parameters, determining network input and output parameters, and tabu search-related parameters; S2 generating neighborhood boundary solutions according to randomly generated operators; S3 calculating the point with the minimum mean square error within the neighborhood; S4 moving the operator to the optimal point and regenerating the neighborhood until the mean square error of the operator is the minimum value within the operator's neighborhood; S5 storing the operator in a tabu table, randomly jumping to generate initial solutions, and continuing to search for local extrema until the tabu table is full or the search ends after a set number of iterations; otherwise, returning to step 2; S6 cyclically passing the local extrema points in the tabu table to the BP neural network training model; and S7 calculating the mean square error of each extrema point to obtain the optimal gesture recognition network model. This invention can overcome the local extrema problem, accelerate the convergence speed of the algorithm, and significantly improve the accuracy of gesture recognition.

Description

Technical Field

[0001] This invention belongs to the field of gesture recognition algorithms, specifically relating to a method for recognizing gestures using a BP neural network optimized by an improved tabu search. Background Technology

[0002] With the rapid development of technology, human-computer interaction has become a new direction for development in many fields. Gesture recognition, as an important part of human-computer interaction, plays an irreplaceable role in fields such as medical devices, virtual games, and individual combat.

[0003] Currently, gesture recognition both domestically and internationally follows two directions: image recognition-based and sensor-based. Image recognition-based methods primarily use cameras to acquire images, but these are easily affected by lighting conditions and occlusion, resulting in poor real-time tracking and recognition performance, and therefore have not been widely adopted. Sensor-based gesture recognition and control, on the other hand, do not have these limitations. In recent years, researchers both domestically and internationally have mostly used electromyography (EMG) sensors to acquire EMG signals for gesture recognition. However, the accuracy rate of gesture recognition using EMG sensors is not very high. This is because the EMG signals from the human skin are weak and easily affected by external interference, and algorithms struggle to process complex EMG signals, leading to low recognition accuracy.

[0004] Gesture recognition is currently used in many fields. Although multi-sensor gesture recognition has many advantages, its development is limited by the low accuracy of the algorithm. Summary of the Invention

[0005] The purpose of this invention is to improve the accuracy of multi-sensor gesture recognition and help BP neural networks overcome problems such as being prone to getting trapped in local extrema.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] S1. Initialize the neural network parameters, determine the network input and output parameters, and tabu search related parameters;

[0008] S2. Generate the neighborhood boundary solution according to the randomly generated operator;

[0009] S3. Calculate the point with the minimum mean square error within the neighborhood;

[0010] S4. Move the operator to the optimal point and continue to generate the neighborhood until the mean square error of the operator is the minimum value in the operator's neighborhood.

[0011] S5. Store the operator in the tabu list, randomly jump to generate an initial solution, and continue to search for local extreme points until the tabu list is full or the search ends after a set number of iterations; otherwise, go to step 2.

[0012] S6. Pass the local extrema in the tabu list to the BP neural network training model in a loop;

[0013] S7. Train the local extrema in the taboo table in sequence, select the point with the minimum mean square error, which is the global minimum, and store the weights and thresholds of the BP neural network when the error is minimum to obtain the BP neural network model for gesture recognition.

[0014] The significant advantages of this invention compared to existing technologies are:

[0015] (1) The improved tabu search optimization BP neural network algorithm of the present invention uses the boundary solution of the operator neighborhood in the tabu search part. This eliminates the need to set the generation rules of the operator neighborhood and does not require complex calculations in the neighborhood. The operator neighborhood only needs to find the approximate location of the local optimum. This can solve the problem that the neighborhood structure is not easy to determine and reduce the amount of computation to improve the efficiency of the algorithm.

[0016] (2) The improved tabu search optimization BP neural network algorithm of the present invention uses a jump criterion in the tabu search part, that is, to determine whether the local optimum has been tabu. If it has been stored in the tabu list, it jumps a fixed number of steps in the original direction. If it has not been tabu, it jumps randomly. This criterion improves the global search capability of tabu search, can traverse all areas, and improves the accuracy of the algorithm's gesture recognition.

[0017] (3) The improved tabu search optimization BP neural network algorithm of the present invention selects to expand the tabu table in the tabu search part to store all the local optima found. This not only solves the problem of poor local search capability of tabu search, but also creates good conditions for the subsequent BP algorithm search, which can ensure that the optimal point is obtained and improve the accuracy of gesture recognition. Attached Figure Description

[0018] Figure 1 This is a flowchart of the algorithm for improving tabu search and optimizing BP neural networks according to the present invention.

[0019] Figure 2 This is a schematic diagram of the operator domain structure of the present invention.

[0020] Figure 3 This is a schematic diagram illustrating movement within the scope of this invention.

[0021] Figure 4 This is a schematic diagram illustrating the change in random jump error in the field of this invention.

[0022] Figure 5 It is the mean square error of each stage of the optimized neural network.

[0023] Figure 6 It is the classification result of the optimized neural network on the test set.

[0024] Figure 7 This is a graph showing the changes in the training parameters of the optimized neural network.

[0025] Figure 8 It represents the correlation between the various stages of the optimized neural network. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] S1. Initialize neural network parameters, determine network input and output parameters, and tabu search related parameters.

[0028] The input to the network model is the electromyographic and inertial features of the hand gesture. Weights and thresholds are initialized, and the parameters for tabu search are determined. The backpropagation (BP) algorithm parameters are initialized; this experiment uses a 78-input, 1-output model. The number of nodes in the hidden layer is determined using an empirical formula. The empirical formula for determining the number of neurons in the hidden layer is:

[0029]

[0030] Where z represents the number of nodes in the hidden layer, x represents the number of nodes in the input layer, y represents the number of nodes in the output layer, and β represents a constant between 1 and 10.

[0031] After multiple experiments, 14 hidden layers were selected, the learning rate η was set to 0.001, and the expected error was set to 0.00001.

[0032] S2. Generate neighborhood boundary solutions according to the randomly generated operators.

[0033] This experiment uses tabu search to optimize the BP neural network. More precisely, tabu search in this experiment only serves to find the approximate location of local optima. Therefore, operator boundary solutions are used, i.e., a critical value is set for the neighborhood range, such as... Figure 2 As shown, after multiple experiments, the neighborhood range was taken from -0.003 to +0.003. That is, each feature only calculates the boundary value of the neighborhood 0.003 before and after the operator. If the value of each feature is less than the boundary value of its range, it is written into the tabu table.

[0034] S3. Function evaluation calculates the mean square error and selects the point with the minimum mean square error in the neighborhood.

[0035] Evaluation involves assessing the boundary solutions within the neighborhood of the solution, then selecting the necessary boundary solutions to exchange with the operator, thus achieving neighborhood shifting. The mean squared error (MSE) is calculated as follows:

[0036]

[0037] Let m be the number of samples in the training set, and Y i Let y represent the predicted value output by the neural network model for the i-th sample. iLet represent the true value of the training set when the i-th sample is generated.

[0038] S4. Move the operator to the optimal point and continue generating the neighborhood until the minimum point is evaluated as the operator;

[0039] The move domain is the process of transforming the current operator to the previously evaluated boundary solution. The initial solution finds a better neighbor solution within its neighborhood; the operator moves, and the neighborhood structure is recalculated, as follows: Figure 3 As shown, the right boundary value calculated in the neighborhood of operator 1 is better. The operator moves to its position and becomes operator 2, generating a new neighborhood boundary value. The search is repeated until the mean square error of the operator's evaluation function is minimized.

[0040] S5. Store the operator in the tabu list, randomly jump to generate the initial solution, and continue to search for local extreme points until the tabu list is full or the search ends after a set number of iterations.

[0041] Following step four, if the mean square error of the boundary values ​​within the operator's neighborhood is greater than the operator's evaluation error, then the operator can be considered a local optimum. Next, it is determined whether the local optimum among the candidate solutions satisfies the pardon criterion. If it does, the tabu is removed, and a fixed number of steps are taken to escape this local optimum. If it does not satisfy the criterion, the tabu step size is updated, and a random jump is performed. This jump is called a shake in heuristic algorithms and is an effective way to escape local optima. Because it is impossible to determine the quality of the search region, random jumps are generally chosen. Similar to randomly constructing an initial solution, a new solution is generated for subsequent neighborhood searches. The tabu search ends when the tabu list is full or after a set number of iterations.

[0042] S6. Pass the local extrema in the tabu list to the BP neural network training model in a loop;

[0043] S7. Train the local extrema in the taboo table in sequence, select the point with the minimum mean square error, which is the global minimum, and store the weights and thresholds of the BP neural network when the error is minimum to obtain the BP neural network model for gesture recognition.

[0044] Update the weights and biases in the network model, and update the parameters from the input layer to the hidden layer:

[0045]

[0046] Where η represents the learning rate, k represents the number of steps to update weights and biases, and ω (k) This represents the weights from the input layer to the hidden layer at step k. denoted by , where represents the bias from the input layer to the hidden layer at step k, and MSE represents the mean squared error.

[0047] Parameter updates from hidden layer to output layer:

[0048]

[0049] Where v (k) This represents the weights from the k-th hidden layer to the output layer. This represents the bias from the k-th hidden layer to the output layer.

[0050] Once all data in the taboo table has been processed, the stored mean squared errors are compared, and the weight and threshold at the minimum error value are taken. These weights and thresholds are the optimal parameters at this point.

[0051] S8. Input gesture test set, using a trained BP neural network to recognize gestures.

[0052] The recognition performance of the algorithm of this invention can be further illustrated by the following simulation experiments:

[0053] (1) Network Model Performance (MSE)

[0054] Improved tabu search optimizes the network performance changes at each stage of the BP neural network, such as... Figure 5 As shown in the figure, the four dashed lines and one circle represent the MSE during training, the MSE during testing, the MSE in the actual experiment, and the optimal MSE, respectively. The circle is the optimal point for recognition. It can be seen from the figure that the network model reaches the global optimum in about 56 steps, with a mean square error of 0.000001, and the accuracy meets the experimental requirements.

[0055] (2) Test set identification results

[0056] The experiment selected 6 gestures, with 20 data points selected for each gesture, for a total of 120 gesture data points. These data points were input into the test set after the network model was trained. The optimized BP neural network classification results are shown below. Figure 6 As shown in the figure, circles represent expected values, stars represent predicted values, and squares below represent error values. The figure shows that the predicted values ​​closely match the expected values, while the error values ​​fluctuate little. This indicates that the optimized neural network has a high recognition accuracy, meeting the experimental requirements.

[0057] (3) Training parameters at each stage of the network

[0058] Gradient: The essence of BP neural network search is still gradient descent. The gradient change can express the search progress of the network model, such as... Figure 7 As shown, the gradient descent during the training process of the improved tabu search-optimized BP neural network is smooth and gradually maintains a decreasing trend until the end, indicating that the model search is smooth and the error gradually decreases until it meets the experimental accuracy requirements.

[0059] Algorithm Error Precision (mu): This experiment uses the trainlm algorithm to approximate the expected value. Higher algorithm precision indicates more accurate point searching. As shown in the figure, the optimized neural network model algorithm precision can reach 10. -7 The search points are more accurate.

[0060] Validation Checks: During BP neural network training, the mean squared error (MSE) is checked for consecutive increases. If the MSE increases more than 6 times consecutively, training will be forcibly stopped. Figure 8 As shown, the mean square error of the optimized neural network increases slightly until it reaches the experimental precision at step 56.

[0061] (4) Correlation

[0062] The model's predicted values ​​and the actual values ​​exhibit a linear regression fit; the closer the value is to 1, the higher the correlation, indicating a more accurate predictive model. For example... Figure 8 As shown in the figure, the top left represents the correlation between the predicted values ​​and the true values ​​in the training set, the top right represents the correlation between the predicted values ​​and the true values ​​in the validation set, the bottom left represents the correlation between the predicted values ​​and the true values ​​in the test set, and the bottom right represents the total correlation. It can be seen from the figure that the correlation of each stage of the optimized neural network model is basically 1, which shows that the gesture model fits the gesture data very well and fully interprets the internal relationship of the gesture data during the training process.

[0063] (6) Gesture classification accuracy

[0064] Because the test set was small and the error was difficult to detect, 10 healthy men were randomly selected, each wearing a device, and 120 gesture data points were collected. After processing, a total of 1200 gesture data points were obtained (200 for each of the 6 gestures). These 1200 test data points were input into various algorithms, including the most commonly used classification algorithms, K-Nearest Neighbors, SVM, and the traditional BP neural network. The recognition data is shown in Table 1. It can be seen that the average accuracy of the improved BP neural network after emergency search optimization reached 99.75%, the average accuracy of the traditional BP neural network was 94.67%, the average accuracy of the K-Nearest Neighbors algorithm was 92.82%, and the average accuracy of the Support Vector Machine algorithm was 97.92%. It can be seen that the optimized BP neural network had the highest recognition accuracy, and its accuracy was improved by an average of 5.08% compared with the traditional neural network.

[0065] Table 1. Gesture recognition accuracy on the test set

[0066]

[0067] This invention provides a gesture recognition method based on an improved tabu search optimized BP neural network. After verification, it shows a significant improvement in the stability of gesture recognition in multi-sensor scenarios and an improvement in accuracy compared to commonly used gesture classification algorithms.

Claims

1. A gesture recognition method based on an improved tabu search optimized BP neural network, characterized in that: Includes the following steps: S1. Initialize the neural network parameters, determine the network input and output parameters, and tabu search related parameters; S2. Generate neighborhood boundary solutions according to randomly generated operators: Select tabu search to optimize the BP neural network, randomly generate tabu search operators, generate operator neighborhoods based on the neighborhood range, and calculate the critical value of the neighborhood; S3. Select the optimal point in the neighborhood: Calculate the mean square error of the operator and the boundary solution, compare the mean square errors and select the point with the smallest error, which is the optimal point in the neighborhood. S4. Move the operator to the optimal point and continue to generate the neighborhood until the mean square error of the operator is the minimum value in the operator's neighborhood. S5. Store the operator in the tabu list, randomly jump to generate an initial solution, and continue to search for local extreme points until the tabu list is full or the search ends after a set number of iterations; otherwise, go to step 2. S6. Pass the local extrema in the tabu list to the BP neural network training model in a loop; S7. Train the local extrema in the taboo table in sequence, select the point with the minimum mean square error, which is the global minimum, and store the weights and thresholds of the BP neural network when the error is minimum to obtain the BP neural network model for gesture recognition. The calculation process for the weights and threshold of the BP neural network at the minimum value is as follows: Update the weights and biases in the network model, and update the parameters from the input layer to the hidden layer: in Let k represent the learning rate, and k represent the number of steps to update the weights and biases. This represents the weights from the input layer to the hidden layer at step k. This represents the bias from the input layer to the hidden layer at step k. Indicates mean square error; Parameter updates from hidden layer to output layer: in This represents the weights from the k-th hidden layer to the output layer. This represents the bias from the k-th hidden layer to the output layer.

2. The gesture recognition method based on an improved tabu search optimized BP neural network according to claim 1, characterized in that: The formulas for calculating the mean square error of the operator and the boundary solution are: Let m be the number of samples in the training set. This represents the predicted value output by the neural network model for the i-th sample. Let represent the true value of the training set when the i-th sample is generated.

3. The gesture recognition method based on an improved tabu search optimized BP neural network according to claim 1, characterized in that: The specific process of S5 is as follows: If the evaluation error of the boundary values ​​within the operator's neighborhood is greater than the operator's evaluation error, then the operator can be considered a local optimum. Next, it is determined whether the local optimum among the candidate solutions satisfies the pardon criterion. If the pardon criterion is satisfied, the tabu is removed, and the algorithm moves forward K steps to escape this local optimum. If the tabu is not satisfied, the tabu step size is updated, and a random jump is performed to generate a new solution for subsequent neighborhood search, until the tabu list is full or the algorithm has run a set number of times.

Images