Pose Estimation Method Based on Quantile-Weighted Robust Support Vector Machine

Abstract

A posture estimation method based on a quantile weighted robust support vector machine, including: 1) constructing a posture estimation data set based on sensing data; 2) processing the human body posture estimation data set. The "one-vs-rest" strategy is adopted to process the multi-class learning problem, that is, dividing the k-class classification problem into k binary classification learning sub-problems; 3) using the quantile method to screen out outliers, and then using the kNN method to calculate the weight values; 4) designing an intra-class weight matrix and an inter-class weight matrix respectively according to the imbalance degree and noise information; 5) constructing a robust support vector machine optimization model based on quantile weighting; 6) solving the linear equations according to the optimization problem; 7) obtaining the optimal robust support vector machine posture class hyperplane; 8) constructing a posture estimation decision function; 9) predicting the human body posture class according to the sensing data of the handheld device.

Description

Technical Field

[0001] The present invention belongs to the field of human body posture estimation for handheld devices, and mainly aims at the problem of class-imbalanced posture estimation in a noisy scenario. Specifically, it relates to a posture estimation method based on Quantile Weighted Robust Support Vector Machine (QWRSVM). Background Art

[0002] Human body posture estimation has high research value and broad application background. Its main application fields include: human-computer interaction, motion analysis, smart home, intelligent security, patient monitoring system, and virtual reality. In the era of mobile Internet, intelligent monitoring is widely applied in various public places, and the daily behaviors of the public can be monitored, which makes human activity recognition very promising in the field of security. In addition, with the gradual aging of the Chinese population, it is of great significance to give timely warnings and assistance for (elderly / patient) dangerous behaviors.

[0003] The present invention uses sensors in intelligent handheld devices as the data source for posture estimation, and is used to estimate six common postures: walking, going upstairs, going downstairs, sitting down, standing, and lying down. Due to the influence of factors such as the working environment, external interference jitter, electronic components, and circuit structure on the sensors, the collected data is contaminated by noise or outliers. For example, in the process of using an acceleration sensor, the pressure-sensitive material is affected by temperature and its own material. When the surrounding working temperature is too high, the polarization effect is strong, resulting in an increase in the output charge quantity and an increase in sensitivity. When the temperature is too low, the output charge quantity decreases and the sensitivity decreases, which further leads to abnormal response values during the detection process, generating noise and outliers. In addition, when the temperature changes suddenly, the pyroelectric effect of the pressure-sensitive material will cause the pressure-sensitive material to output a low-frequency clutter, which will also generate abnormal values and noise. Therefore, it poses a great challenge to construct a robust posture estimation method for the problem of class-imbalanced noisy posture estimation.

[0004] In the field of machine learning, there are two difficult problems for human pose estimation: Problem 1, most algorithms assume that the data is clean and free of noise pollution. However, there is inevitable noise in the data acquisition process, which will lead to the degradation of algorithm performance. Problem 2, there is an imbalance problem in the categories of human poses, that is, the number of normal data is much larger than that of abnormal data, which will cause traditional algorithms to ignore minority classes during the learning process. To alleviate the above problems, scholars in this field have given some solutions in the following literatures: [1]. Ye, Q., et al., Weighted Twin Support Vector Machines with Local Information and its application. Neural Networks, 2012. 35: p. 31 - 39. (abbreviated as WLTSVM) [2]. Shao, Y., et al., An efficient weighted Lagrangian twin support vector machine for imbalanced data classification. Pattern Recognition, 2014. 47(9): p. 3158 - 3167. (abbreviated as WLTSVM - CIL) [3]. Xu, Y., K - nearest neighbor - based weighted multi - class twin support vector machine. Neurocomputing, 2016. 205: p. 430 - 438. (abbreviated as KWMTSVM) [4]. Tanveer, M., A. Sharma and P. N. Suganthan, Least squares KNN - based weighted multi - class twin SVM. Neurocomputing, 2021. 459: p. 454 - 464. (abbreviated as LS - KWMTSVM)

[0005] Among them, in reference [1], by constructing two graphs, the within - class graph and the between - class graph, the within - class compactness and between - class separability of the data are respectively modeled. The purpose is to make the two generated hyperplanes most adaptable to the high - density relevant points and far away from other points, and in this way, the marginal noise can be filtered out. In reference [2], aiming at the imbalance problem, it is solved by reconstructing the training points and adopting appropriate weights. In reference [3], aiming at the fact that previous multi - classification methods ignored the local information of samples, a weight matrix is introduced based on kNN to make full use of the local information. In reference [4], on the basis of KWMTSVM, the inequality constraint is changed to an equality constraint, further reducing the time complexity.

[0006] Although the methods in the above-mentioned literature have alleviated the noise problem and imbalance problem of SVM to a certain extent, the following challenges still exist:

[0007] (1) In WLTSVM, LS-KWMTSVM, and KWMTSVM, weights are assigned to samples only by calculating the number of kNNs of each sample, ignoring the differences in local distances of samples.

[0008] (2) In WLTSVM, the problem of between-class imbalance is not considered, resulting in the model being biased towards the minority class.

[0009] (3) In WLTSVM_CIL, weight construction is achieved by finding each other's kNNs to undersample the majority-class sample points, and this method highly depends on the selection of the k value.

[0010] (4) In addition, WLTSVM_CIL only considers the number of positive and negative class samples when designing weights and does not consider the local information between neighbors. Summary of the Invention

[0011] The present invention aims to overcome the above-mentioned disadvantages of the prior art and proposes a human pose estimation method based on a quantile-weighted robust support vector machine. The present invention mainly focuses on the problem of noise pose estimation with class imbalance and can be applied to a patient monitoring system to identify and warn of dangerous behaviors of patients, so as to provide timely treatment.

[0012] To this end, the present invention will focus on designing a quantile-weighted robust SVM method to solve the problem of class-imbalanced pose estimation in a noise scenario, specifically as follows:

[0013] (1) By combining the intra-class and inter-class local information, the method of the present invention can learn multi-class information more fully. Secondly, considering the distance differences between neighbor pairs of sample points, different weights are assigned according to the different distances between neighbor pairs to reflect the influence of local information.

[0014] (2) The present invention measures the distance between a sample and its surrounding samples based on the kNN method and assigns different weights to the samples according to the distance information. Using the quartile method, samples greater than the upper outlier cut-off point are defined as outliers (noise / outliers), and the weights are set to 0 in the method to alleviate the influence of noise.

[0015] (3) The present invention not only constrains the weights of intra-class sample points but also constrains the weights of inter-class sample points. Secondly, by weighted screening of inter-class samples, it is beneficial to shorten the training time of the method and deal with the problem of class imbalance.

[0016] In order to facilitate better understanding and description, the present invention formulates a unified symbol representation rule, in which scalars are represented by lowercase or uppercase letters without bold, such as: x, X. Vectors are represented by lowercase letters with bold, such as: x. Matrices are represented by uppercase letters with bold, such as: X.

[0017] The posture estimation method based on quantile weighted robust support vector machine of the present invention has the following overall process: Figure 1 The specific steps are described as follows:

[0018] Step 1: Construction of human posture estimation dataset. First, carry a smartphone and perform six activities in sequence: walking, walking upstairs, walking downstairs, sitting, standing, and lying down. Secondly, based on the smartphone sensor data, manually mark the posture category. Thirdly, preprocess the sensor signals (accelerometer and gyroscope) by applying filters, and then sample them with a sliding window of fixed width of 2.56 seconds of sensor time series. The sensor acceleration signal with gravity and human motion components is separated into human acceleration and gravity using a Butterworth low-pass filter. It is assumed that gravity has only low-frequency components, so a filter with a cutoff frequency of 0.3Hz is used. By calculating variables from the time domain and frequency domain, feature vectors can be obtained from each window. And provide data with these feature engineering. Then, the feature x processed by feature engineering is used i , and the manually labeled categories are used as output y i Finally, the data is normalized to obtain a supervised dataset where x i ∈R n is the sample feature of dimension n, y i is the output corresponding to the sample, and m is the sample size.

[0019] Step 2: Multi-classification modeling for human pose estimation. The “one-to-rest” strategy is used to handle the multi-classification learning problem, that is, the S-class classification problem is divided into S binary classification learning sub-problems. For the s-th class s=1,...S, the data set is divided into two non-repeating parts: the data sample matrix A belonging to the s-th class s And the data sample matrix of all remaining categories Among them, the matrix Contains l s s-th class sample points; contain The remaining sample points are used. s To represent the human body pose estimation data sample point index belonging to class s, represents the sample x of class s i ,i∈I S .

[0020] Step 3: Calculate the within-class weight value and noise estimate using kNN weighted sum and quartiles respectively. Denote the within-class weight value as s = 1, ..., S, i ∈ I s representing the within-class weight of the i-th sample in the s-th class. The within-class weight process of the present invention is as Figure 2 shown.

[0021] (3.1) Calculate the kNN distance of each sample x i within the class, and then sort them in ascending order and denote as n (s) , and denote the distance value as s = 1, ..., S, i, j ∈ I s , k = 1, ..., K.

[0022]

[0023] (3.2) Calculate the kNN distance of each category of data upper quartile and lower quartile Then, calculate the interquartile range IQD (s) :

[0024]

[0025] (3.3) Calculate the upper outlier cut-off point in different categories respectively, denoted as t (s) , s = 1, ..., S representing the upper outlier cut-off point of the s-th class.

[0026]

[0027] where β is a constant.

[0028] (3.4) According to the property of the upper outlier cut-off point, the closer the kNN value of a sample is to the upper outlier cut-off point value of the class, the farther the sample point is from the distribution center of the class. Therefore, this sample should be assigned a smaller within-class weight. When the kNN value is greater than the upper outlier cut-off point value, the sample is regarded as an outlier and its weight is assigned 0. And the smaller the kNN value, the closer the sample point is to the sample point center and the more important it is to the class, and a larger weight should be assigned but not greater than 1. Based on the above analysis, for the s-th class: The within-class weight of the present invention is constructed as follows:

[0029]

[0030] Step 4: In order to use the local geometric structure of the data for modeling, calculate the between-class weight of each category according to the kNN method. The between-class weight process of the present invention is as Figure 3 shown.

[0031] (4.1) Adopt the strategy of "one pair with the rest" to construct an inter-class graph for each class, denoted as That is, for the \(s^{th}\) class (\(s = 1,\cdots,S\)), the original classes are divided into two non-overlapping parts. The data samples belonging to the \(s^{th}\) class are regarded as the intra-class part \(A\) s , and the remaining samples that do not belong to the \(s^{th}\) class are regarded as the inter-class part

[0032]

[0033] where \(x\) i belongs to other classes except the \(s^{th}\) class, \(x\) j belongs to the \(s^{th}\) class, \(x\) i , \(x\) j is the kNN relationship.

[0034] (4.2) Use the Euclidean metric to measure the distance between any pair of kNNs. The present invention aims to extract as many possible support vectors residing in the samples of another class as possible. Then, the inter-class weight value is redefined, denoted as and the mean value of the distance between the sample and the kNN \(s = 1,\cdots,S, i\in I\) s is used as the weight value.

[0035]

[0036] Step 5: Construct the QWRSVM classifier. The present invention aims to find an optimal hyperplane \(f\) s (x):

[0037] \(f\) s (x) = \(w\) s T \(x + b\) s = 0, \(s = 1,\cdots,S\) (8)

[0038] where \(f\) s (x) is the proximal hyperplane of class \(s\). \(w\) s and \(b\) s are the weight and bias of the corresponding hyperplane, and \(x\) is the feature of the sample.

[0039] To solve the overfitting problem and the matrix singular value problem caused by empirical risk minimization in practice. The present invention proposes a structural risk minimization version by introducing an additional regularization term Then, the original problem of QWRSVM is obtained:

[0040]

[0041] where \(\xi\) is a non-negative slack variable; \(c_1, c_2\) are penalty parameters.

[0042] For the convenience of description, the present invention transforms the vector form of the above samples into the following matrix form

[0043]

[0044] Among them, Equation (10) represents the optimization problem of the proximal hyperplane of class s. Q s is a diagonal matrix, and the values on its main diagonal are the within-class weight values of class s, Equation (4), D s is a diagonal matrix, and the values on its main diagonal are the between-class weight values of class s, Equation (7), which are defined as follows:

[0045]

[0046] used to control the influence of samples within or between classes; e s1 , e s2 is a unit vector of appropriate dimension.

[0047] Analyzing the optimization problem (10), the first term of the objective function to be minimized is the regularization term. In order to achieve the minimum structural risk, that is, to balance the model complexity and the model accuracy and avoid overfitting; the second term is to make the sample points of the s-th class closer to the hyperplane of this class, that is, to make the samples gather as close as possible to the hyperplane of this class; the last term requires that the distances from the sample points of other classes to the hyperplane are at least 1, and the slack variable is the measurement error when the constraint conditions are not satisfied, that is, it aims to maximize the separation of the two types of sample points.

[0048] Step Six: Solve the optimization problem in Step Five.

[0049] (6.1) Taking the optimization problem (10) as an example, first substitute the first constraint condition into the objective function, and then substitute the second constraint condition to obtain the Lagrangian function of the objective function:

[0050]

[0051] (6.2) Take the gradient of Equation (10) with respect to w s , b s and set the gradient equal to 0 to get

[0052]

[0053] (6.3) According to Equation (13) and Equation (14), solve to get

[0054]

[0055] Among them, H = [A e s1 , G = [B e s2, I is the identity matrix of appropriate dimension. Thus, the training phase of the QWRSVM method is completed.

[0056] Step 7: Classifier pose estimation prediction phase. For a new data sample, i.e., the handheld device sensor data x, the category of the new sample is determined by calculating the distance from x to each non-parallel proximal hyperplane. Each non-proximal hyperplane represents a pose (walking, walking upstairs, walking downstairs, sitting, standing, lying down). The category of the hyperplane with the closest distance is the category of the new sample. The decision function is as follows:

[0057]

[0058] where |·| represents the absolute value, and s = 1,..., S represents the s-th category.

[0059] In the present invention, for the problem of noise pose estimation with class imbalance, it can be applied to application scenarios such as patient or elderly care for identifying abnormal pose warnings, and a QWRSVM method is proposed. When the dataset has features such as noise / outliers and class imbalance learning, the QWRSVM method performs more excellently.

[0060] The advantages of the present invention are:

[0061] (1) By combining the local information within and between classes, the method of the present invention can learn multi-class information more fully. Secondly, considering the distance difference between neighbor pairs of sample points, different weights are assigned according to the different distances of neighbor pairs to reflect the influence of local information.

[0062] (2) The present invention measures the distance between a sample and its surrounding samples based on the kNN method, and assigns different weights to the samples according to the distance information. Using the quartile method, the samples greater than the upper outlier cut-off point are defined as outliers (noise / outliers), and the weights are set to 0 in the method to mitigate the influence of noise.

[0063] (3) The present invention not only performs weight constraints on the sample points within a class, but also on the sample points between classes. Secondly, by weighted screening of the sample points between classes, it is beneficial to shorten the training time of the method and deal with the problem of class imbalance. Brief Description of the Drawings

[0064] Figure 1 is the overall algorithm flowchart of the present invention.

[0065] Figure 2 is the flowchart of the within-class weight algorithm of the present invention.

[0066] Figure 3 is the flowchart of the between-class weight algorithm of the present invention. Detailed Embodiments

[0067] The following presents the preferred embodiments of the present invention in conjunction with the accompanying drawings to elaborate in detail on the technical solutions of the present invention. The embodiments described herein are only for explaining the present invention and do not limit the present invention. The specific steps are as follows:

[0068] Step 1: Construction of the human pose estimation dataset. First, carry a smartphone and perform one of six activities (walking, walking upstairs, walking downstairs, sitting, standing, lying down) in sequence, while recording the video of each activity. Second, based on the sensor data collected by the smartphone, manually label the category of the movement according to the data obtained from the video and the sensor. Third, perform preprocessing by applying the accelerometer and gyroscope, and then sample with a sliding window of a fixed width of 2.56-second sensor time series. Use a Butterworth low-pass filter to separate the sensor acceleration signal with gravity and human movement components into human acceleration and gravity. Assuming that gravity only has low-frequency components, a filter with a cut-off frequency of 0.3 Hz is used. By calculating variables from the time domain and frequency domain, a feature vector can be obtained from each window. And provide the data with these feature engineering. Then, use the feature x after feature engineering i , and at the same time use the manually labeled category as the output y i . Finally, normalize the data to obtain a supervised dataset where x i ∈R n is the sample feature of dimension n, y i is the corresponding output of the sample, and m is the sample size.

[0069] Step 2: Process the human pose estimation dataset. Adopt the "one-vs-rest" strategy to handle the multi-classification learning problem, that is, divide the S-class classification problem into S binary classification learning sub-problems. For the s-th class, s = 1,..., S, divide the dataset into two non-overlapping parts: the data sample matrix A belonging to the s-th class s and the data sample matrix of all remaining classes where the matrix contains l s sample points of the s-th class; contains sample points of the remaining classes; use I s to represent the index of the human activity data sample points belonging to the s-th class, representing the sample x of the s-th class i , i ∈ I S .

[0070] Step 3: Use kNN weighted sum and quartiles to calculate the within-class weight value and noise estimation respectively. Denote the within-class weight value as s = 1,..., S, i ∈ I sDenote the within-class weight of the \(i\)-th sample in the \(s\)-th class. The within-class weight process of the present invention is as follows Figure 2 shown.

[0071] (3.1) Calculate the kNN distance of each sample \(x\) in the \(s\)-th class i within the class, and then sort them in ascending order and denote them as \(n\) (s) , and denote the distance values as \(s = 1,\cdots,S, i,j\in I\) s , \(k = 1,\cdots,K\).

[0072]

[0073] (3.2) Calculate the upper quartile and lower quartile of the kNN distances of each category of data Then, calculate the interquartile range IQD (s) :

[0074]

[0075] (3.3) Calculate the upper outlier cut-off point \(t\) (s) , \(s = 1,\cdots,S\) represents the upper outlier cut-off point of the \(s\)-th class.

[0076]

[0077] where \(\beta\) is a constant.

[0078] (3.4) According to the property of the upper outlier cut-off point, the closer the kNN value of a sample is to the upper outlier cut-off point value of the class, the farther the sample point is from the distribution center of the class. Therefore, this sample should be assigned a smaller within-class weight. When the kNN value is greater than the upper outlier cut-off point value, the sample is regarded as an outlier and its weight is assigned to 0. And the smaller the kNN value, the closer the sample point is to the sample center, the more important it is to the class, and a larger weight should be assigned but not greater than 1. Based on the above analysis, for the \(s\)-th class: The within-class weight of the present invention is constructed as follows:

[0079]

[0080] Step four: In order to use the local geometric structure of the data for modeling, according to the kNN method, calculate the between-class weights of each category respectively. The between-class weight process of the present invention is as follows Figure 3 shown.

[0081] (4.1) Adopt the "one-vs-rest" strategy to construct an between-class graph for each class and denote it as That is, for each category \(s = 1,\ldots,S\), the original categories are divided into two non - overlapping parts. The data samples belonging to the \(s\) - th class are regarded as the intra - class part \(A\). s , and the remaining samples that do not belong to the \(s\) - th class are regarded as the inter - class part

[0082]

[0083] Here, \(x\) i belongs to other categories except the \(s\) - th class, \(x\) j belongs to the \(s\) - th class, \(x\) i , \(x\) j is a kNN relationship.

[0084] (4.2) The Euclidean metric is used to measure the distance between any pair of kNNs. The present invention aims to extract as many possible support vectors residing in samples of another class as possible. Then, the inter - class weight value is re - defined and denoted as and the mean value of the distance between the sample and the kNN \(s = 1,\ldots,S,i\in I\) s is used as the weight value.

[0085]

[0086] Step Five: Construct the QWRSVM classifier. The present invention aims to find an optimal hyperplane \(f\) s (x):

[0087] \(f\) s (x)=w s T x + b s = 0, \(s = 1,\ldots,S\) (8)

[0088] where, \(f\) s (x) is the proximal hyperplane of class \(s\). \(w\) s and \(b\) s are the weight and bias of the corresponding hyperplane, and \(x\) is the feature of the sample.

[0089] To solve the over - fitting problem and the matrix singular value problem brought about by empirical risk minimization in practice. The present invention proposes a version of structural risk minimization by introducing an additional regularization term Then, the original problem of QWRSVM is obtained:[[]]

[0090]

[0091] where, \(\xi\) is a non - negative slack variable; \(c_1,c_2\) are penalty parameters.

[0092] For the convenience of description, the present invention transforms the vector form of the above samples into the following matrix form:

[0093]

[0094] Among them, Equation (10) represents the optimization problem of the proximal hyperplane of class s. Q s is a diagonal matrix, and the values on its main diagonal are the within-class weight values of class s, Equation (4), D s is a diagonal matrix, and the values on its main diagonal are the between-class weight values of class s, Equation (7), which are defined as follows:

[0095]

[0096] used to control the influence of samples within or between classes; e s1 , e s2 is a unit vector of appropriate dimension.

[0097] Analyzing the optimization problem (10), the first term of the objective function to be minimized is the regularization term. To achieve the minimum structural risk, that is, to balance the model complexity and the model accuracy and avoid overfitting; the second term is to make the sample points of the s-th class closer to the hyperplane of this class, that is, to make the samples gather as close as possible to the hyperplane of this class; the last term requires that the distances from the sample points of other classes to the hyperplane are at least 1, and the slack variable is the measurement error when the constraint conditions are not satisfied, that is, it aims to maximize the separation of the two types of sample points.

[0098] Step Six: Solve the optimization problem in Step Five.

[0099] (6.1) Taking the optimization problem (10) as an example, first substitute the first constraint condition into the objective function, and then substitute the second constraint condition to obtain the Lagrangian function of the objective function:

[0100]

[0101] (6.2) Take the gradient of Equation (10) with respect to w s , b s and set the gradient equal to 0, we get

[0102]

[0103]

[0104] (6.3) According to Equation (13) and Equation (14), solve to get

[0105]

[0106] Among them, H = [A e s1, G = [B e s2 , where I is the identity matrix of appropriate dimension. Thus, the training phase of the QWRSVM method is completed.

[0107] Step Seven: Classifier pose estimation prediction phase. For a new data sample, i.e., the handheld device sensor data x, the class of the new sample is determined by calculating the distance from x to each non-parallel proximal hyperplane. Each non-proximal hyperplane represents a pose (walking, walking upstairs, walking downstairs, sitting, standing, lying down). The class of the hyperplane with the closest distance is the class of the new sample. The decision function is as follows:

[0108]

[0109] where |.| represents the absolute value, and s = 1,..., S represents the s-th class.

[0110] To demonstrate the robustness of the present invention against noise, the present invention adds random missing value noise with intensities of 0.03, 0.05, 0.10, 0.15, 0.20, and 0.30 to each row of the human pose estimation dataset used for experiments. The human pose estimation dataset contains a total of 561 features. Due to the large number of features, feature screening is performed, and 100 key features are selected. Finally, to ensure the fairness of the experimental results, the ten-fold cross-validation method is used. Since the present experiment uses an imbalanced dataset, G_mean is used as the imbalanced classification performance metric for the method, and the specific calculation is as follows:

[0111]

[0112] where TP represents the number of samples of the s-th class predicted as the s-th class, FN represents the number of samples of the s-th class predicted as other classes, FP represents the number of samples of other classes predicted as the s-th class, and TN represents the number of samples of other classes predicted as other classes.

[0113] The present invention is based on an intelligent handheld device, uses sensors such as accelerometers and gyroscopes, and then samples samples with a 2.56-second length sensor time series as a sliding window with a fixed width. The samples represent six poses: walking, going upstairs, going downstairs, sitting, standing, and lying down. To verify the robustness of the QWRSVM method against noise, a total of 6 groups of noise scenarios are designed. Specifically as follows: Gaussian noise pollution is added to the data, and the variances are mild (0.03 and 0.05), moderate (0.1 and 0.15), and severe noise (0.2 and 0.25). Table 1 records the performance comparison of pose estimation of the double-boundary support vector machine (TBSVM), weighted twin support vector machine (WLTSVM), and the QWRSVM method of the present invention in terms of the G-means imbalanced metric.

[0114]

[0115] Table 1 Comparison of Pose Estimation Performance of Each Method at Different Noise Intensities

[0116] The results in Table 1 show that as the noise intensity of the dataset increases, the performance of TBSVM and WLTSVM gradually decreases. For the QWRSVM method proposed in the present invention, the pose estimation performance is not significantly affected as the noise intensity increases. Although, in the mild noise scenarios (0.03 and 0.05), the performance of TBSVM is slightly higher than that of QWRSVM; however, in the moderate and severe noise scenarios, the pose estimation performance of QWRSVM exceeds that of TBSVM and WLTSVM. In summary, the QWRSVM method proposed in the present invention has good robustness in the pose estimation scenario for dealing with noise / outliers.

[0117] The above is the preferred embodiment of the present invention, and the protection scope of the present invention is not limited thereto. Any transformation or replacement conceived without creative work by any person skilled in the art within the technical scope disclosed by the present invention is covered by the protection scope of the present invention. The protection scope of the present invention shall be subject to the protection scope defined by the claims.

Claims

1. A pose estimation method based on quantile weighted robust support vector machine, comprising the following steps: Step 1: Construction of human posture estimation dataset; First, carry a smartphone and perform six activities in sequence, walking, walking upstairs, walking downstairs, sitting, standing, and lying down; Second, based on the smartphone sensor data, manually mark the posture category; Third, pre-process the sensor signal by applying a filter and then sample it in a fixed-width sliding window; Use a Butterworth low-pass filter to separate the sensor acceleration signal with gravity and human motion components into human acceleration and gravity; Assume that gravity only has low-frequency components, so a filter with a cut-off frequency of 0.3 Hz is used; feature vectors are obtained from each window by calculating variables in the time domain and frequency domain; and the data after such feature engineering are provided; then, the features x after feature engineering are used i , while the manually labeled category is used as the output y i ; finally, the data are normalized to obtain a supervised dataset where x i ∈R n is the sample feature of dimension n, y i is the corresponding output of the sample, and m is the sample size; Step 2: Multi-class modeling of human pose estimation; The "one-vs-rest" strategy is adopted to handle the multi-class learning problem, that is, dividing the classification problem of S classes into S binary classification learning sub-problems; For the s-th class, s = 1,..., S, the data set is divided into two non-overlapping parts: the data sample matrix A belonging to the s-th class s and the data sample matrix of all remaining classes where the matrix contains l s sample points of the s-th class; contains sample points of the remaining classes; Use I s to represent the index of the human pose estimation data sample points belonging to the s-th class, representing the sample x of the s-th class i , i ∈ I S ; Step 3: Calculate the within-class weight value and noise estimate using kNN weighted sum and quartiles respectively; denote the within-class weight value as which represents the within-class weight of the i-th sample in the s-th class; (3.1) Calculate the kNN distance of each sample x in the s-th class within the class, and then sort them from small to large. Denote the distance values as i (3.2) Calculate the kNN distances of various types of data the upper quartile and the lower quartile Then, calculate the interquartile range IQD (s) : (3.3) Calculate the upper outlier cut-off point separately for different categories, denoted as t (s) , where s = 1,..., S represents the upper outlier cut-off point of the s-th category; where β is a constant; (3.4) According to the properties of the upper outlier truncation point, the closer the kNN value of a sample is to the upper outlier truncation point value of this class, the farther the sample point is from the distribution center of this class; therefore, this sample should be assigned a smaller within-class weight; when the kNN value is greater than the upper outlier truncation point value, the sample is regarded as an outlier, and its weight is assigned 0; and the smaller the kNN value, the closer the sample point is to the sample point center, the more important it is for this class, and a larger weight should be assigned but not greater than 1; based on the above analysis, the within-class weight value for the s-th class is constructed as follows: Step four: In order to use the local geometric structure of the data for modeling, according to the kNN method, calculate the between-class weights of each category respectively; (4.1) Adopt the strategy of "one pair with the remainder" to construct an inter-class graph for each class, denoted as That is, for the s-th category, s = 1,..., S, the original categories are divided into two non-overlapping parts. The data samples belonging to the s-th category are regarded as the intra-class part A s , and the remaining samples that do not belong to the s-th category are regarded as the inter-class part The x here i belongs to other categories except the s-th category, x j belongs to the s-th category, x i , x j is a kNN relationship; (4.2) Use the Euclidean metric to measure the distance between any pair of kNNs; extract the possible support vectors residing in samples of another class; then define the inter-class weight value denoted as and use the mean of the distances between the samples and the kNNs as the weight value; Step 5: Construct a QWRSVM classifier; aim to find an optimal hyperplane f s (x) for each pose category: f s (x) = w s T x + b s = 0, s = 1,..., S (8) where f s (x) is the proximal hyperplane of class s; w s and b s are the weights and biases of the corresponding hyperplane, and x is the feature of the sample; To solve the overfitting problem and matrix singular value problem brought by empirical risk minimization in practice; by introducing an additional regularization term The structural risk minimization version is proposed; then, the primal problem of QWRSVM is obtained: where ξ is a non-negative slack variable; c1, c2 are penalty parameters; For the convenience of expression, convert the vector form of the above sample into the following matrix form Among them, Equation (10) represents the optimization problem of the proximal hyperplane of class s; Q s is a diagonal matrix, and the values on its main diagonal are the within-class weight values of class s, Equation (4), D s is a diagonal matrix, and the values on its main diagonal are the between-class weight values of class s, Equation (7), which are defined as follows: For controlling the influence of samples within or between classes; e s1 ,e s2 is a unit vector of appropriate dimension; Analyze the optimization problem (10). The first term of the objective function to be minimized is the regularization term. In order to achieve the minimum structural risk, that is, to balance the model complexity and model accuracy and avoid overfitting; the second term is to make the sample points of the s-th class closer to the hyperplane of this class, that is, to make the samples gather as close as possible to the hyperplane of this class; the last term requires that the distance from the sample points of other classes to the hyperplane is at least 1, and the slack variable is the measurement error when the constraint conditions are not met, that is, it aims to maximize the separation of the two types of sample points; Step six: Solve the optimization problem in step five; (6.1) In the optimization problem formula (10), first substitute the first constraint condition into the objective function, and then substitute the second constraint condition to obtain the Lagrangian function of the objective function: (6.2) Take the gradient of equation (10) with respect to w s , b s , set the gradient equal to 0, and we get (6.3) Solve according to Equation (13) and Equation (14). Obtain where H = [A e s1 , G = [B e s2 , and i is the identity matrix of appropriate dimension; thus, the training phase of the QWRSVM method is completed; Step seven: The classifier pose estimation prediction stage; for a new data sample, that is, the handheld device sensor data x, judge the category of the new sample by calculating the distance from x to each non-parallel proximal hyperplane; each non-proximal hyperplane represents a pose; the category of the hyperplane with the closest distance is the category of the new sample; the decision function is as follows: where |.| represents the absolute value, and s = 1,..., S represents the s-th category.

2. The pose estimation method based on the quantile weighted robust support vector machine according to claim 1, characterized in that: The sensor described in step one is an accelerometer and a gyroscope.

3. The pose estimation method based on the quantile weighted robust support vector machine according to claim 1, wherein: The fixed width of the sliding window described in step one is a sensor time series with a length of 2.56 seconds.

4. The pose estimation method based on quantile weighted robust support vector machine according to claim 1, wherein: The postures described in step seven include: walking, going upstairs, going downstairs, sitting down, standing, lying down.

Images