A human emotion recognition method based on inertial sensor

Abstract

The present invention discloses a human emotion recognition method based on an inertial sensor, which specifically relates to the field of emotion recognition. The mapping relationship between human body movements and human emotions is established through a literature reference method and a fuzzy comprehensive evaluation method; a subject is stimulated to generate emotions through a video, and an inertial sensor worn on the subject is used to collect emotion action data; the emotion action data is segmented and sampled using a sliding window method to obtain emotion action data samples; the features of the emotion action data samples are extracted through a time-frequency domain analysis method; the feature vector is reduced in dimension using a principal component analysis method; model training and parameter optimization are performed using training data to obtain a weighted kernel support vector machine model; and human emotion recognition is performed using a weighted and support vector machine model. The present invention can effectively overcome the influence of a single kernel function used in a support vector machine model on the recognition effect, and can effectively improve the accuracy of human emotion recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of emotion recognition, and more specifically, to a human emotion recognition method based on inertial sensors. Background Art

[0002] At present, artificial intelligence and human-computer interaction are showing a trend of rapid growth and comprehensive expansion. Human emotions will cause certain changes in human speech, facial expressions, and body movements, which has led to the development of speech emotion recognition, facial emotion recognition, and action emotion recognition. In speech emotion recognition, voice signals are used as input. Although voice signals are easy to collect and intuitive to display, they have certain defects. For example, for the standard of voice input, everyone's accent and pronunciation habits are different, and there is no fixed standard for machines to understand these inputs. In addition, environmental noise processing is a recognized technical problem. In the field of facial expression and body movement emotion recognition, the most commonly used method is the image recognition method based on vision. Although this method is convenient and intuitive, it is relatively complex. It requires high-specification hardware equipment to process images, which will occupy a large amount of memory resources. In addition, factors such as lighting and occlusion in the shooting scene will also affect the recognition effect.

[0003] In recent years, with the development of microcomputer hardware, high-performance physical sensors have made great breakthroughs and are widely used in fields such as health detection, communication positioning, and entertainment games. The sensor monitoring system for motion recognition is called the human sensor network. The sensor is fixed to one or several parts of the human body to collect information. Compared with speech processing and image processing technology, sensors have many advantages in some aspects. The cost of sensors is low and they are more easily accepted by people in daily life. In addition, it is relatively small and can be worn on the body for a long time without affecting the normal activities of the monitored person. Users can also turn on or off the data collection function according to their living habits and ideas, without excessively invading personal privacy. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a human emotion recognition method based on an inertial sensor.

[0005] To achieve the above object, the present invention provides the following technical solution: a human emotion recognition method based on inertial sensor, the specific recognition steps are as follows:

[0006] Step S1: Establish the mapping relationship between human body movements and human emotions through literature reference method and fuzzy comprehensive evaluation method;

[0007] Step S2: Stimulate the subject to generate emotion and make action through the video, and collect the emotion and action data of m subjects using n inertial sensor nodes worn on the subject;

[0008] Step S3: using a sliding window method to perform segmentation sampling processing on the emotion action data collected by the n inertial sensor nodes worn by each subject, so as to obtain all emotion action data samples;

[0009] Step S4: extracting the features of the emotion action data sample by time-frequency domain analysis to obtain a feature vector representing human emotion;

[0010] Step S5: using principal component analysis to reduce the dimension of the feature vector and convert it into a comprehensive feature vector with as few features as possible and which can best represent human emotions;

[0011] Step S6: dividing the comprehensive feature vector converted in step S5 into training data and test data, and using the training data to perform model training and parameter optimization to obtain a weighted kernel support vector machine model;

[0012] Step S7: Perform human emotion recognition based on the test data and the weighted kernel support vector machine model in step S6.

[0013] In a preferred embodiment, the step S1 specifically comprises:

[0014] S1.1: Using the literature reference method, refer to multiple emotional action datasets provided by relevant literature, list human action labels and human emotion labels, and preliminarily establish the mapping relationship between human action and human emotion;

[0015] S1.2: Using the fuzzy comprehensive evaluation method, the action labels and emotion labels obtained in step S1.1 are evaluated in turn by multiple experts, and correlation scores are given to further determine the mapping relationship between human body movements and human emotions.

[0016] In a preferred embodiment, the step S3 specifically comprises:

[0017] S3.1: Use the sliding window method to divide the emotion action data collected by the n inertial sensor nodes worn by each subject into multiple emotion action data segments of equal length, and each emotion action data segment is used as a sample;

[0018] S3.2: The sliding window method has two parameters: window width and overlap rate. Let the window width be ω and the overlap rate between every two adjacent windows be μ, so that a total of t windows are obtained, where t = INT(V / μω)-1, V represents the length of the emotion action data, INT is the rounding function, and it is rounded down to the nearest integer;

[0019] S3.3: For the jth sensor node, the emotional action data segment in the Tth window consists of three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data, which are represented as

[0020] In a preferred embodiment, the specific steps of step S4 include: manually extracting the time domain features and frequency domain features of the emotional action data segment in each window through the time-frequency domain analysis method, and extracting the mean, variance, kurtosis, energy, and frequency domain entropy of the three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data respectively, and the obtained feature vector dimension is M = 5 features × 9 axes × n sensors.

[0021] In a preferred embodiment, the specific steps of step S5 include: using the principal component analysis algorithm to perform feature dimensionality reduction processing on the M-dimensional feature vector to convert it into N comprehensive feature vectors that best represent human emotions, where N<<M, F N represents the set of N-dimensional comprehensive feature vectors, F M Represents a set of M-dimensional feature vectors.

[0022] In a preferred embodiment, the step S7 specifically comprises:

[0023] The weighted kernel model constructs a weighted kernel function by linearly convexly combining multiple basic kernel functions. Its specific form is:

[0024] K(x,z)=a1K1(x,z)+α2K2(x,z)+…+α h K h (x,z)+…+α k K k (x,z)

[0025] Among them, K(x,z) is the weighted kernel function, K k (x,z) is the basic kernel function, including the linear kernel function, polynomial kernel function, Gaussian kernel function and Sigmoid kernel function, k is the number of basic kernel functions, α h is the weight coefficient corresponding to the hth basic kernel function, h = 1, 2, ..., k, and the weight coefficient satisfies the following conditions: a1 + α2 + ... + α h +…+α k =1;

[0026] By defining the fuzzy function f(θ h ) to get the model recognition rate θ h The membership degree is used to calculate the weight coefficient α corresponding to the h-th basic kernel function. h, the fuzzy function f(θ h ) is:

[0027]

[0028] Among them, a is a constant, θ h is the model recognition rate obtained when the support vector machine model uses the hth basic kernel function for recognition, then the weight coefficient α corresponding to the hth basic kernel function h for:

[0029]

[0030] The support vector machine model uses a weighted kernel function to establish a weighted kernel support vector machine model to classify and recognize human emotions.

[0031] Technical effects and advantages of the present invention:

[0032] The present invention uses literature reference method and fuzzy comprehensive evaluation method to establish the mapping relationship between human body movements and human emotions, and constructs an emotional movement data set; at the emotion recognition level, a weighted support vector machine model is proposed. The model mainly includes two parts. First, the kernel function in the support vector machine model is linearly convexly combined to construct a multi-kernel support vector machine model, and then the fuzzy function is used to improve the weight calculation method of multiple kernel functions in the support vector machine model to construct a weighted kernel support vector machine model. This can effectively overcome the influence of the support vector machine model using a single kernel function on the recognition effect, and can effectively improve the recognition accuracy of the human emotion recognition method. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart for implementing the present invention.

[0034] Figure 2 It is a schematic diagram of sliding window segmentation of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The present invention provides Figure 1 A human emotion recognition method based on inertial sensor is shown, and the specific recognition steps are as follows:

[0037] Step S1: Establish the mapping relationship between human body movements and human emotions through literature reference method and fuzzy comprehensive evaluation method. The specific steps include:

[0038] Using the literature reference method, we refer to multiple emotional action data sets provided by relevant literature, list human action labels and human emotion labels, and preliminarily establish the mapping relationship between human action and human emotion;

[0039] Using the fuzzy comprehensive evaluation method, the action labels and emotion labels obtained in step S1.1 are evaluated by multiple experts in turn, and correlation scores are given to further determine the mapping relationship between human actions and human emotions;

[0040] Step S2: Stimulate the subject to generate emotion and make action through the video, and collect the emotion and action data of m subjects using n inertial sensor nodes worn on the subject;

[0041] Specifically, firstly, a movie video is watched to stimulate the subject to produce the emotion and make corresponding movements. N inertial sensor nodes are fixed on n parts of the subject's body respectively. Each sensor node contains a three-axis angular velocity meter, a three-axis gyroscope and a three-axis magnetometer. The collected emotional movement data include three-axis acceleration data, three-axis angular velocity speed and three-axis magnetic field data. The inertial sensor node uses short-range wireless transmission technology to transmit the collected emotional movement data to the host computer for data storage and processing. Then, the n inertial sensors are used to collect the emotional movement data of m subjects in sequence. For the jth (j=1,2,…,n) sensor node, the collected three-axis acceleration data can be expressed as The three-axis angular velocity data can be expressed as The three-axis magnetic field data can be expressed as The three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data collected by the j-th sensor node can constitute the emotional motion data S j =(a j ,g j ,m j );

[0042] Step S3: Use the sliding window method to segment and sample the emotion action data collected by the n inertial sensor nodes worn by each subject to obtain all emotion action data samples. The specific steps include:

[0043] like Figure 2 The emotional action data S obtained in step S2 is shown j The sliding window method is used to divide the emotion action data collected by the n inertial sensor nodes worn by each subject into multiple emotion action data segments of equal length, and each emotion action data segment is used as a sample;

[0044] The sliding window method has two parameters: window width and overlap rate. Let the window width be ω and the overlap rate between every two adjacent windows be μ, so that a total of t windows are obtained, where t = INT(V / μω)-1, V represents the length of the emotion action data, INT is the rounding function, and it is rounded down to the nearest integer;

[0045] For the jth sensor node, the emotional action data segment in the Tth window consists of three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data, which are respectively expressed as

[0046] Step S4: extracting the features of the emotion action data sample by time-frequency domain analysis to obtain a feature vector representing human emotion, the specific steps comprising: manually extracting the time domain features and frequency domain features of the emotion action data segment in each window by time-frequency domain analysis, respectively extracting the mean, variance, kurtosis, energy, and frequency domain entropy of the three-axis acceleration data, the three-axis angular velocity data, and the three-axis magnetic field data, and obtaining a feature vector dimension of M=5 features×9 axes×n sensors;

[0047] Specifically, for the mean values ​​of the three-axis acceleration data, angular velocity data, and magnetic field data in the Tth sliding window, the formulas are: Where l = {x, y, z} represents the three-axis direction, represents the mean value of the three-axis acceleration data of the T-th sliding window of the j-th sensor node, represents the three-axis acceleration data of the Tth sliding window of the jth sensor node, represents the mean of the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the mean value of the three-axis magnetic field data of the T-th sliding window of the j-th sensor node, represents the three-axis magnetic field data of the Tth sliding window of the jth sensor node, i = 1, 2, ..., ω;

[0048] For the variance of the three-axis acceleration data, angular velocity data, and magnetic field data in the Tth sliding window, the formulas are Where l = {x, y, z} represents the three-axis direction, represents the variance of the three-axis acceleration data of the T-th sliding window of the j-th sensor node, represents the variance of the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the variance of the three-axis magnetic field data of the T-th sliding window of the j-th sensor node;

[0049] For the kurtosis of the triaxial acceleration data, angular velocity data, and magnetic field data in the Tth sliding window, the formulas are: Where l = {x, y, z} represents the three-axis direction, represents the kurtosis of the triaxial acceleration data of the Tth sliding window of the jth sensor node, represents the kurtosis of the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the kurtosis of the three-axis magnetic field data of the T-th sliding window of the j-th sensor node;

[0050] For the energy of the three-axis acceleration data, angular velocity data, and magnetic field data in the Tth sliding window, the formulas are Where l = {x, y, z} represents the three-axis direction, represents the energy of the three-axis acceleration data of the T-th sliding window of the j-th sensor node, represents the energy of the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the energy of the three-axis magnetic field data of the T-th sliding window of the j-th sensor node;

[0051] For the frequency domain entropy of the three-axis acceleration data, angular velocity data, and magnetic field data in the Tth sliding window, the formulas are: Where l = {x, y, z} represents the three-axis direction, represents the frequency domain entropy of the triaxial acceleration data of the Tth sliding window of the jth sensor node, represents the frequency domain entropy of the three-axis angular velocity data of the T-th sliding window of the j-th sensor node, represents the frequency domain entropy of the three-axis magnetic field data of the T-th sliding window of the j-th sensor node;

[0052] According to the above method, the 45-dimensional features of the three-axis acceleration data, three-axis angular velocity data and three-axis magnetic field data in the T-th sliding window are obtained. These features can constitute a 45-dimensional feature vector:

[0053]

[0054] Step S5: Use the principal component analysis method to reduce the dimension of the feature vector and convert it into a comprehensive feature vector with the smallest possible number and the best representation of human emotions (due to the possible correlation and redundancy between the time domain and frequency domain features extracted in step S4, the obtained feature vector has a certain correlation, so feature selection is required to simplify the data structure, improve the operation speed and accuracy of the recognition algorithm, and obtain the smallest possible number of comprehensive feature vectors from the feature vector that best represent human emotions). The specific steps include: using the principal component analysis algorithm to reduce the dimension of the M-dimensional feature vector and convert it into N comprehensive feature vectors that best represent human emotions, where N<<M, F N represents the set of N-dimensional comprehensive feature vectors, F M Represents a set of M-dimensional feature vectors;

[0055] Step S6: dividing the comprehensive feature vector converted in step S5 into training data and test data, and using the training data to perform model training and parameter optimization to obtain a weighted kernel support vector machine model;

[0056] Specifically, the reduced-dimensional comprehensive feature vector obtained in step S5 is divided into training data and test data, and the parameters are optimized by, for example, leave-one-out cross-validation. Specifically, the training data is evenly divided into B parts, of which B-1 parts are used for model training and the remaining 1 part is used for verification. A total of B model trainings are performed until all the data are used for verification. Then, the model is trained and the parameters are optimized using the training data, so that a classification model for each human emotion can be obtained. When evaluating the model, the accuracy is selected as the factor for model evaluation, and the parameter obtained in the B training with the highest accuracy is the optimized parameter.

[0057] Step S7: Perform human emotion recognition based on the test data and the weighted kernel support vector machine model in step S6. The specific steps include:

[0058] The weighted kernel model constructs a weighted kernel function by linearly convexly combining multiple basic kernel functions. Its specific form is:

[0059] K(x,z)=a1K1(x,z)+α2K2(x,z)+…+α h K h (x,z)+…+α k K k (x,z)

[0060] Among them, K(x,z) is the weighted kernel function, K k (x,z) is the basic kernel function, including the linear kernel function, polynomial kernel function, Gaussian kernel function and Sigmoid kernel function, k is the number of basic kernel functions, αh is the weight coefficient corresponding to the hth basic kernel function, h = 1, 2, ..., k, and the weight coefficient satisfies the following conditions: a1 + α2 + ... + α h +…+α k =1;

[0061] By defining the fuzzy function f(θ h ) to get the model recognition rate θ h The membership degree is used to calculate the weight coefficient α corresponding to the h-th basic kernel function. h , the fuzzy function f(θ h ) is:

[0062]

[0063] Among them, a is a constant, θ h is the model recognition rate obtained when the support vector machine model uses the hth basic kernel function for recognition, then the weight coefficient α corresponding to the hth basic kernel function h for:

[0064]

[0065] The support vector machine model uses a weighted kernel function to establish a weighted kernel support vector machine model to classify and recognize human emotions.

[0066] The specific implementation method is:

[0067] For example, referring to the emotional action dataset provided in the study of Klaus R. Scherer and Heinerelglgring, human action labels and human emotion labels are listed, and anger, boredom, excitement, tension, pain and sleepiness are initially determined to be marked as emotional labels. Human action labels include waving hands, clapping, covering the face with both hands, rubbing hands, waving fists and stretching; then fuzzy comprehensive evaluation is used to further determine the mapping relationship between human actions and human emotions; human action labels and human emotion labels are evaluated by 20 experts in turn; combined with the reference dataset and fuzzy comprehensive evaluation, we determined that the mapping relationship between human actions and human emotions is sleepiness (stretching), boredom (waving hands), excitement (clapping), tension (rubbing hands), anger (waving fists), pain (covering the face with both hands);

[0068] A total of 9 subjects were recruited to participate in the data collection of this study, including 5 males and 4 females, ranging in age from 22 to 26 years old. All subjects were in good health. Six emotional videos were used to stimulate the subjects to produce 6 corresponding emotions and make corresponding movements. Four inertial sensor nodes were fixed on the left wrist, right wrist, left forearm and right forearm of each subject. The sampling frequency of the inertial sensor nodes was set to 50Hz, and the duration of each movement was specified as two minutes. During the data collection process, the subjects made corresponding movements according to their habits.

[0069] Since the emotional action collection time is relatively long and the amount of emotional action data collected is relatively large, it is difficult to obtain as much complete information about the emotional action data as possible. The sliding window method is used to segment the emotional action data collected by the four inertial sensor nodes worn by each subject to obtain multiple emotional action data segments of equal length, and each emotional action data segment is used as a sample;

[0070] The sliding window method has two parameters: window width and overlap rate. Let the window width be 100 and the overlap rate between every two adjacent windows be 50%. In this way, a total of t windows are obtained, where t = INT (V / 50) - 1, V represents the length of the emotion action data, INT is the rounding function, and it is rounded down to the nearest integer; for the emotion action data fragment in the Tth window of the jth sensor node, it consists of three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data, which are respectively expressed as

[0071] Through the time-frequency domain analysis method, the time domain features and frequency domain features of the emotional action data fragment in each window are manually extracted, and the mean, variance, kurtosis, energy, and frequency domain entropy of the three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data are extracted respectively. The dimension of the obtained feature vector is M = 5 features × 9 axes × 4 sensors; the feature vector of the T-th sliding window of the j-th sensor node can be expressed as:

[0072]

[0073] There may be correlation and redundancy between the extracted time domain and frequency domain features, so feature selection is needed to simplify the data structure and improve the operation speed and accuracy of the recognition algorithm; the principal component analysis method is used to combine similar features and dimensions to reduce the feature dimension of the 45-dimensional feature vector. We select features with a cumulative contribution rate greater than 95% and convert them into 6 comprehensive feature vectors that best represent human emotions, among which, F N represents a set of 6-dimensional comprehensive feature vectors, F MRepresents a set of 45-dimensional feature vectors;

[0074] The reduced-dimensional comprehensive feature vector obtained in step S5 is divided into training data and test data, and the parameters are optimized by, for example, leave-one-out cross-validation. Specifically, the training data is evenly divided into 8 parts, of which 7 parts are used for model training and the remaining 1 part is used for verification. The model training is performed 8 times in total until all the data are used for verification. Then, the model is trained and the parameters are optimized using the training data, so that a classification model for each human emotion can be obtained. When evaluating the model, the accuracy rate is selected as the factor for model evaluation, and the parameters obtained in the 8 training times with the highest accuracy rate are the optimized parameters. The average accuracy rate of the 8 tests of leave-one-out cross-validation is taken as the final emotion recognition accuracy rate, and the weighted and support vector machine model proposed in the present invention is used to classify and recognize the test data set, and the recognition accuracy rate is 98.41%;

[0075] Then, using the same training data and test data and the same training method, experiments were conducted using the support vector machine (SVM) algorithm, the hidden Markov model (HMM) algorithm, the BP neural network (BPNN) algorithm, the decision tree (DT) algorithm, and the random forest (RF) algorithm, respectively, and compared with the method provided in the embodiment of the present invention. The emotion recognition accuracy of each algorithm is shown in Table 1;

[0076] Table 1 Comparison of classification accuracy obtained using different human emotion recognition methods

[0077] method SVM HMM BPNN DT RF The present invention Average recognition rate 94.31％ 76.24％ 92.11％ 88.56％ 91.54％ 98.41％

[0078] The present invention provides a human emotion recognition method based on inertial sensors, establishes a mapping relationship between human body movements and human emotions by using a literature reference method and a fuzzy comprehensive evaluation method, and constructs an emotion movement data set; at the emotion recognition level, the present invention mainly proposes a weighted support vector machine model, which mainly includes two parts. First, a multi-kernel support vector machine model is constructed by linear convex combination of kernel functions in a support vector machine model, and then a weighted kernel support vector machine model is constructed by improving the weight calculation method of multiple kernel functions in the support vector machine model by using a fuzzy function. This can effectively overcome the influence of a single kernel function used in a support vector machine model on the recognition effect, and can effectively improve the recognition accuracy of the human emotion recognition method.

[0079] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A human emotion recognition method based on inertial sensor, characterized in that: The specific identification steps are as follows: Step S1: Establish the mapping relationship between human actions and human emotions through literature reference method and fuzzy comprehensive evaluation method; S1.1: Using the literature reference method, refer to multiple emotional action datasets provided by relevant literature, list human action labels and human emotion labels, and preliminarily establish the mapping relationship between human action and human emotion; S1.2: Using the fuzzy comprehensive evaluation method, the action labels and emotion labels obtained in step S1.1 are evaluated by multiple experts in turn, and correlation scores are given to further determine the mapping relationship between human actions and human emotions; Step S2: Stimulate the subject to generate emotion and make action through the video, and collect the emotion and action data of m subjects using n inertial sensor nodes worn on the subject; Step S3: using a sliding window method to perform segmentation sampling processing on the emotion action data collected by the n inertial sensor nodes worn by each subject, so as to obtain all emotion action data samples; S3.1: Use the sliding window method to divide the emotion action data collected by the n inertial sensor nodes worn by each subject into multiple emotion action data segments of equal length, and each emotion action data segment is used as a sample; S3.2: The sliding window method has two parameters: window width and overlap rate. Let the window width be ω and the overlap rate between every two adjacent windows be μ, so that a total of t windows are obtained, where t = INT(V / μω)-1, V represents the length of the emotion action data, INT is the rounding function, and it is rounded down to the nearest integer; S3.3: For the jth sensor node, the emotional action data segment in the Tth window consists of three-axis acceleration data, three-axis angular velocity data, and three-axis magnetic field data, which are represented as j=1,2,…,n, T=1,2,…,t; Step S4: extracting the features of the emotion action data sample by time-frequency domain analysis to obtain a feature vector representing human emotion, the specific steps comprising: manually extracting the time domain features and frequency domain features of the emotion action data segment in each window by time-frequency domain analysis, respectively extracting the mean, variance, kurtosis, energy, and frequency domain entropy of the three-axis acceleration data, the three-axis angular velocity data, and the three-axis magnetic field data, and obtaining a feature vector dimension of M=5 features×9 axes×n sensors; Step S5: using principal component analysis to reduce the dimension of the feature vector and convert it into a comprehensive feature vector with as few features as possible and which can best represent human emotions; Step S6: dividing the comprehensive feature vector converted in step S5 into training data and test data, and using the training data to perform model training and parameter optimization to obtain a weighted kernel support vector machine model; Step S7: Perform human emotion recognition based on the test data and the weighted kernel support vector machine model in step S6.

2. The method for human emotion recognition based on inertial sensor according to claim 1, characterized in that: The specific steps of step S5 include: using the principal component analysis algorithm to perform feature dimensionality reduction processing on the M-dimensional feature vector, and convert it into N comprehensive feature vectors that best represent human emotions, where N<<M, F N represents the set of N-dimensional comprehensive feature vectors, F M Represents a set of M-dimensional feature vectors.

3. A human emotion recognition method based on an inertial sensor according to any one of claims 1-2, characterized in that: The specific steps of step S7 include: The weighted kernel model constructs a weighted kernel function by linearly convexly combining multiple basic kernel functions. Its specific form is: K(x,z)=a1K1(x,z)+α2K2(x,z)+…+α h K h (x,z)+…+α k K k (x,z) Among them, K(x,z) is the weighted kernel function, K k (x,z) is the basic kernel function, including the linear kernel function, polynomial kernel function, Gaussian kernel function and Sigmoid kernel function, k is the number of basic kernel functions, α h is the weight coefficient corresponding to the hth basic kernel function, h = 1, 2, ..., k, and the weight coefficient satisfies the following conditions: a1 + α2 + ... + α h +…+α k =1; By defining the fuzzy function f(θ h ) to get the model recognition rate θ h The membership degree is used to calculate the weight coefficient α corresponding to the h-th basic kernel function. h , the fuzzy function f(θ h ) is: Among them, a is a constant, θ h is the model recognition rate obtained when the support vector machine model uses the hth basic kernel function for recognition, then the weight coefficient α corresponding to the hth basic kernel function h for: The support vector machine model uses a weighted kernel function to establish a weighted kernel support vector machine model to classify and recognize human emotions.

Images