A method for determining pedestrian foot zero speed based on support vector machine algorithm

Abstract

A method for determining the zero speed of the foot of a single-gait pedestrian based on an SVM algorithm is proposed, which includes: obtaining the output values ​​of a gyroscope and an accelerometer as raw data; preprocessing the raw data; designing a zero-speed determination classifier based on an SVM; training the SVM zero-speed determination classifier, continuously optimizing the SVM parameters, and obtaining a structural modeling SVM classifier for zero-speed determination; sending the data to be tested into the structural modeling SVM classifier for zero-speed determination to obtain a zero-speed determination result. The present invention also proposes a method for determining the zero speed of the foot of a single-gait pedestrian based on an SVM algorithm.

Description

Technical Field

[0001] The present invention relates to speed determination of positioning navigation, and in particular to a method for determining zero speed of multi-gait pedestrian feet based on a support vector machine algorithm. Background Art

[0002] For individual navigation based on foot-bound INS (inertial navigation system), the most important error constraint method is zero-speed correction. The zero-speed correction algorithm (ZUPT) is the core of the pedestrian navigation algorithm based on foot-bound INS. For low-cost micro-gyroscopes and micro-accelerometers, the inertial sensor errors such as zero bias, random drift, scale factor nonlinearity, and orthogonal error are relatively large. Without any assistance, the INS calculation will accumulate very quickly. The current mainstream judgment algorithm is to set several judgment thresholds based on the modulus of the output values ​​of the three-axis gyroscope and the three-axis accelerometer, and judge whether the foot is still on the ground by comparing the magnitude relationship between the modulus and the judgment threshold. Because the threshold setting of this algorithm is relatively absolute, for different step modes, different step frequencies, different sole materials, and different ground environments, the judgment threshold needs to be changed to obtain a better zero-speed judgment result. The algorithm has poor adaptability and a high misjudgment rate. Therefore, this zero-speed detection algorithm based on foot-bound INS is difficult to adapt to the needs of individual navigation under war conditions. Under modern warfare conditions, the ground environment is complex and the gait patterns of individual soldiers are more diverse. Zero-speed judgment of multiple gait patterns in complex environments has become a difficulty in individual soldier navigation. Therefore, it is necessary to study a highly robust zero-speed judgment algorithm that can adapt to individual soldier navigation in complex environments.

[0003] Support vector machine (SVM) is a machine learning method based on statistical learning theory developed in the 1990s. The algorithm mainly seeks to minimize statistical risk to improve the generalization ability of the learning machine, minimize empirical risk and confidence range, and thus achieve the purpose of obtaining good statistical laws even when the statistical sample size is small. The SVM classifier can achieve a balance between generalizability and empirical error, and is currently a popular classifier. The SVM algorithm can automatically remove some noise data, making it more robust, and it is widely used in gait pattern recognition. Therefore, the robustness of the SVM algorithm is very important for the zero-speed judgment of foot-bound INS.

[0004] Due to the many shortcomings of the currently commonly used zero-speed detection algorithm based on foot-bound INS and the robustness characteristics of the SVM algorithm, the present invention proposes a pedestrian foot zero-speed determination method based on the support vector machine algorithm to meet the needs of individual navigation under complex conditions. Summary of the invention

[0005] In order to solve the problems existing in the prior art, one aspect of the present invention proposes a method for zero-speed determination of the foot of a single-gait pedestrian based on an SVM algorithm, which includes: obtaining the output values ​​of a gyroscope and an accelerometer as raw data; preprocessing the raw data; designing a zero-speed determination classifier based on SVM; training the SVM zero-speed determination classifier, continuously optimizing the SVM parameters, and obtaining a structural modeling SVM classifier for zero-speed determination; sending the data to be tested into the structural modeling SVM classifier for zero-speed determination to obtain a zero-speed determination result.

[0006] Preferably, the single-gait pedestrian foot zero-speed determination method also includes: performing a walking mode zero-speed determination experiment and a running mode zero-speed determination experiment on the structural modeling SVM classifier; the walking mode zero-speed determination experiment includes: a normal walking mode zero-speed determination experiment, a fast walking mode zero-speed determination experiment, and a multi-step walking mode zero-speed determination experiment; the running mode zero-speed determination experiment includes: a normal-step running mode zero-speed determination experiment, a fast running mode zero-speed determination experiment, and a multi-step running mode zero-speed determination experiment.

[0007] Preferably, the design of the zero-speed determination classifier based on SVM includes: selecting the characteristic value of the SVM classifier; setting the determination threshold and generating a sample set; determining the sample set according to the determination criterion to obtain a zero-speed determination result vector; and setting SVM parameters.

[0008] Preferably, the characteristic values ​​of the SVM classifier are: gyro output modulus, accelerometer output modulus, gyro output modulus mean square error, accelerometer output modulus mean square error; wherein the sampling time interval Δt=0.01s, wherein,

[0009] The gyro output modulus is:

[0010] Mod_w'=(w x 2 +w y 2 +w z 2 ) 1 / 2 Δt

[0011] The output modulus value of the addition table is:

[0012] Mod_a'=|(a x 2 +a y 2 +a z 2 ) 1 / 2 -g0|Δt

[0013] The standard deviation of the gyro output modulus is:

[0014]

[0015] The standard deviation of the accelerometer output modulus is:

[0016]

[0017] Preferably, the determination thresholds are respectively: Thu_w, Thu_a, Thu_a_σ, Thu_w_σ, and the four characteristic values ​​and the four determination thresholds together constitute a determination criterion. Under this determination criterion, the zero speed determination result is as follows:

[0018]

[0019] Among them, Stance_d is 0, indicating a zero-speed state, and the foot is stationary on the ground; Stance_d is 1, indicating a non-zero-speed state, and the foot is in a kicking motion state.

[0020] Preferably, the setting of SVM parameters includes: dividing the gait into two states: zero speed and non-zero speed, setting the parameters of the SVM classifier to 2 categories, and setting the SVM related parameters as follows after repeated verification: number of SVM categories: 2 categories, H0: zero speed state, H1: non-zero speed state; SVM label value: H0=0, H1=1; SVM type: C-SVM; SVM kernel function selection: radial basis kernel function, as shown in the following formula:

[0021]

[0022] The gamma function setting in the kernel function is: g = 2.8; SVM penalty coefficient: C = 6.2.

[0023] Preferably, the parameters of the structural modeling SVM classifier are as follows: number of support vectors: nSV=894; the constant term of the decision function is:

[0024] b=-0.197718

[0025] The SVM support vector is:

[0026]

[0027] The support vector coefficients are:

[0028] W=[w1 w2 w3 ... w 894 ]

[0029] The SVM high-dimensional feature space classification decision function becomes:

[0030]

[0031] The single-gait SVM zero-speed classifier judgment formula is:

[0032]

[0033] Since the label values ​​of the SVM zero-speed classifier are set to 1 and 0, the label values ​​of the SVM classification results are converted. j The judgment label is d j , the determination formula is as follows:

[0034]

[0035] According to another aspect of the present invention, a method for determining zero speed of a multi-gait pedestrian foot based on a support vector machine algorithm is provided, comprising:

[0036] Obtain gyro and accelerometer output values ​​as raw data; preprocess the raw data; design a multi-gait zero-speed determination classifier based on SVM, wherein the design of the multi-gait zero-speed determination classifier further includes: selecting eigenvalues ​​of the SVM classifier, wherein the eigenvalues ​​are 10, including: gyro output modulus, accelerometer output modulus, gyro output modulus mean square error, accelerometer output modulus mean square error, gyro output modulus mean, accelerometer output modulus mean, gyro modulus maximum, gyro modulus minimum, accelerometer output modulus maximum, accelerometer output modulus minimum;

[0037] Parameters of the multi-gait zero-speed determination classifier are set and trained to obtain a multi-gait structure SVM classifier for multi-gait zero-speed determination; based on the multi-gait structure SVM classifier, five groups of data of normal walking, fast walking, normal running, and fast running are respectively merged to obtain five groups of feature data with different step frequencies and different steps, the first group of data is used to perform multi-gait SVM learning and training, and then the five groups of data are classified using the learned multi-gait SVM classifier, and compared with the results obtained using the original traditional zero-speed determination algorithm.

[0038] Preferably, the mean value of the gyro output modulus is:

[0039]

[0040] The mean value of the accelerometer output modulus is:

[0041]

[0042] The maximum and minimum values ​​of the gyro module are:

[0043]

[0044] The maximum and minimum values ​​of the accelerometer output modulus are:

[0045]

[0046] The 10 types of feature values ​​are used to form a 10-dimensional feature vector to perform learning, training and classification on the multi-gait zero-speed determination classifier. The 10-dimensional feature vector is:

[0047] [Mod_w Mod_aσ Mod_w σ Mod_a W_mean A_mean Max_w Min_w Max_aMin_a].

[0048] Preferably, based on the LIBSVM software package, the multi-gait SVM classifier is trained using a sample data set containing the 10-dimensional feature vector to obtain a support vector supporting multi-gait zero-speed classification, and the parameters of the multi-gait SVM classifier are set as follows:

[0049] The multi-gait zero-speed judgment SVM classifier is divided into 4 categories: H0: walking zero speed, H0 = 0; H1: walking non-zero speed, H1 = 1; H2: running zero speed, H2 = 2; H3: running non-zero speed, H3 = 3; SVM type: C-SVM; SVM penalty coefficient: C = 1.8;

[0050] The training of the multi-gait zero-speed judgment SVM classifier includes: selecting the feature vectors corresponding to H0~H1, H0~H2, H0~H3, H1~H2; H1~H3, H2~H3 as sample training sets during training, and 6 trained SVM classifiers can be obtained through training. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Various embodiments or examples ("examples") of the present disclosure are disclosed in the following detailed description and accompanying drawings. The drawings are not necessarily drawn to scale. In general, unless otherwise specified in the claims, the operations of the disclosed methods may be performed in any order.

[0052] In the attached figure:

[0053] Figure 1 A flow chart of a method for determining zero speed based on a single gait in SVM according to the present invention is shown;

[0054] Figure 2 Comparison of zero-speed determination results for the second group of normal walking data;

[0055] Figure 3 Comparison of zero-speed determination results for the second group of fast walking data;

[0056] Figure 4 Comparison of zero-speed determination results for the second group of multi-cadence walking data;

[0057] Figure 5 Comparison of zero-speed determination results for the second group of normal running data;

[0058] Figure 6 Comparison of zero-speed determination results for the second group of fast running data;

[0059] Figure 7 Comparison of zero-speed determination results for the second group of multi-cadence running data;

[0060] Figure 8 To compare the classification results of running data using normal walking SVM;

[0061] Fig. 9 SVM classification for multiple gait patterns;

[0062] Fig.10 This is a comparison of the zero-speed determination results of multiple gaits in group 1;

[0063] Fig.11 This is a comparison of the zero-speed determination results of multiple gaits in group 2;

[0064] Fig.12 This is a comparison of the zero-speed determination results of multiple gaits in group 3;

[0065] Fig.13 This is a comparison of the zero-speed determination results of multiple gaits in Group 4;

[0066] Fig.14 This is a comparison of the zero-speed determination results of multiple gaits in Group 5;

[0067] Fig.15 It is the pedestrian navigation map of the Navigation Engineering Center of Tsinghua University;

[0068] Fig.16 This is a comparison chart between the SVM classification results and the original adaptive classification results;

[0069] Fig.17 The zero speed determination results of the three algorithms are shown below;

[0070] Fig.18 Comparison of walking trajectories of the three algorithms. DETAILED DESCRIPTION

[0071] Before explaining one or more embodiments of the present disclosure in detail, it should be understood that the embodiments are not limited to the construction details in their specific applications, and the steps or methods proposed in the following embodiments or drawings. The system and method of the present invention will be described in detail below with reference to the drawings.

[0072] 1. Single-gait zero-speed determination method based on SVM

[0073] In the field of pedestrian navigation, SVM is mainly used for the classification of gait patterns. The present invention proposes to use the SVM classifier for the zero-speed and non-zero-speed classification in the foot-bound INS. For individual navigation based on the foot-bound INS, the purpose of zero-speed determination is to perform zero-speed correction (ZUPT), so it is necessary to accurately determine the zero-speed and non-zero-speed moments. The zero-speed determination problem of the foot-bound INS is converted into a zero-speed and non-zero-speed classification problem, which is divided into two categories, one is the grounded static state (zero speed), and the other is the kicking motion state (non-zero speed). The zero-speed classification during zero-speed determination is a real-time classification. It is necessary to determine the category of the walking state at the current moment in real time, and the system decides whether to use the ZUPT algorithm. The flow chart of the single-gait zero-speed determination method based on SVM is shown below. Figure 1 As shown below. Figure 1 , describe the technical solution of the present invention in detail.

[0074] (I) Obtaining gyro and accelerometer output values ​​as raw data

[0075] In order to avoid adding extra sensors to the individual navigation of the foot-worn INS, the present invention performs zero-speed determination based on the output values ​​of the gyroscope and accelerometer. Therefore, firstly, the output values ​​of the gyroscope and accelerometer are collected as the original sample data required for the learning and training of the SVM classifier, and the sample data consists of a feature value and a determination result; the feature value of the sample data adopts the feature value in the existing zero-speed determination algorithm, the determination threshold is manually adjusted, and the existing zero-speed determination algorithm is used to perform zero-speed determination, and the determination result is obtained as the standard determination result of the sample set.

[0076] (II) Preprocessing of raw data

[0077] Since the output modulus of the gyroscope and accelerometer is large, it is not conducive to SVM learning and training, so the original data needs to be preprocessed. In order to facilitate the learning and training of SVM, the original data is uniformly multiplied by the sampling time interval Δt=0.01s.

[0078] (III) Design of zero speed judgment classifier based on SVM

[0079] (1) Select the feature value of the SVM classifier

[0080] According to the actual experimental situation, the single-gait SVM classifier selects four eigenvalues, namely, the gyro output modulus, the accelerometer output modulus, the gyro output modulus mean square error, and the accelerometer output modulus mean square error, as the SVM classifier eigenvalues. The SVM classifier eigenvalue is multiplied by the sampling time interval Δt = 0.01s. The SVM eigenvalue vector is calculated as follows:

[0081] Gyro output modulus:

[0082] Mod_w'=(w x 2 +wy 2 +w z 2 ) 1 / 2 Δt (1)

[0083] Add table output modulus value:

[0084] Mod_a'=|(a x 2 +a y 2 +a z 2 ) 1 / 2 -g0|Δt (2)

[0085] Gyro output modulus standard deviation:

[0086]

[0087] Accelerometer output modulus standard deviation:

[0088]

[0089] (2) Set the judgment threshold and generate a sample set

[0090] Furthermore, four decision thresholds are set for the above four feature vectors, namely: Thu_w, Thu_a, Thu_a_σ, Thu_w_σ. The four feature vectors and the four decision thresholds together constitute the decision criterion. Under this decision criterion, the zero speed decision result is as follows:

[0091]

[0092] Among them, Stance_d is 0, indicating a zero-speed state, and the foot is stationary on the ground; Stance_d is 1, indicating a non-zero-speed state, and the foot is in a kicking motion state.

[0093] A set of sample data of single-gait individual navigation is collected, and a set of sample data is obtained by calculation using formulas (1) to (5). Let the four eigenvalue variables be x, and the sample matrix composed of n sample data be X, which are respectively:

[0094] x=[x1 x2 x3 x4]=[Mod_w' Mod_a' σ' Mod_w σ' Mod_a ] (6)

[0095]

[0096] (3) The sample set is judged according to the judgment criteria to obtain the zero speed judgment result vector

[0097] According to the judgment criteria of formula (5), the sample data is judged to obtain a set of judgment results, which are used as the standard zero-speed judgment results. Let the zero-speed judgment result be d i , then the zero speed determination result vector can be formed by the determination results:

[0098] zup=[d1 d2 ... d n ] T (8)

[0099] (4) Setting of SVM parameters

[0100] SVM is used to design a zero-speed judgment classifier. After the number of classifications of the classifier is determined and the sample data to be learned is prepared, the open source library package can be used to design the parameters, train and classify the SVM. LIBSVM is a set of SVM program libraries developed by Professor Lin Zhiren in 2001. The library has few input parameters and is easy to expand. It can easily classify or regress data, so it has become the most widely used SVM program library in China. The present invention will use the LIBSVM library to design a zero-speed judgment SVM classifier for foot-bound INS. Based on LIBSVM, the complex learning and training process of SVM can be completed. Designers do not need to worry about the complex mathematical implementation process of the SVM classifier, and can focus on more creative work such as feature vector selection and parameter optimization to improve the design efficiency of the SVM classifier.

[0101] Based on the SVM single gait zero speed judgment algorithm, the gait is divided into two states: zero speed and non-zero speed. The parameters of the SVM classifier are set to 2 categories. After repeated verification, the SVM related parameters are set as follows:

[0102] Number of SVM categories: 2 categories, H0: zero speed state, H1: non-zero speed state;

[0103] SVM label value: H0=0, H1=1;

[0104] SVM type: C-SVM

[0105] SVM kernel function selection: radial basis kernel function, as shown in formula (9).

[0106] Radial Basis Fuction (RBF):

[0107]

[0108] Gamma function setting in kernel function: g = 2.8;

[0109] SVM penalty coefficient: C = 6.2

[0110] (IV) Training the SVM classifier and continuously optimizing the SVM parameters to obtain a structural modeling SVM classifier for zero speed determination

[0111] The prepared single-gait zero-speed judgment sample data is used to train the set SVM classifier. The trained support vector and model parameters are saved in the structure variable model. The specific parameters are as follows:

[0112] Number of support vectors: nSV = 894;

[0113] The constant term of the decision function (SVM parameter: -model.rho) is:

[0114] b=-0.197718(10)

[0115] The SVM support vector (SVM parameter: model.SVs) is:

[0116]

[0117] The support vector coefficient (SVM parameter: model.sv_coef) is:

[0118] W=[w1 w2 w3 ... w 894 ](12)

[0119] (5) sending the data to be tested into the structural modeling SVM classifier for zero speed determination to obtain the zero speed determination result;

[0120] Use the support vectors and parameters obtained by SVM training to perform zero-speed classification. Suppose the vector to be classified is x j , combining the SVM classification decision formula (13) and the radial basis kernel function formula (9) to obtain the single-step SVM zero-speed classifier decision formula (14)

[0121] The SVM high-dimensional feature space classification decision function becomes:

[0122]

[0123] The single-gait SVM zero-speed classifier judgment formula (14) is:

[0124] (14)

[0125] Since the label values ​​of the SVM zero-speed classifier are set to 1 and 0, the label values ​​of the SVM classification results are converted. j The judgment label is d j , as shown in the determination formula (15).

[0126]

[0127] 2. Experiment on zero-speed judgment of walking mode based on SVM

[0128] (1) Zero speed determination experiment in normal walking mode

[0129] Collect 5 groups of normal walking data, first use the original judgment algorithm to judge the zero speed, and get the judgment results of zero speed and non-zero speed, which are called "original judgment results" here. The support vector and model parameters trained by the present invention are the results obtained by training the first group of walking data. Use the trained SVM classifier to classify the 5 groups of data respectively, and get the SVM classification results. The classification results of the second group of data are compared as follows Figure 2 shown. Figure 2 The vertical axis is the judgment result, 1 indicates non-zero speed state, 0 indicates zero speed state, and the horizontal axis indicates the sampling point. Table 1 is a statistical table of classification accuracy. It can be seen that the classification accuracy of the five groups of data is more than 97%, and the trajectory of the original judgment result and the SVM classification result are almost the same. The SVM classification result has no noise, which shows that the SVM classifier has good fault tolerance and robustness, and can meet the zero speed judgment requirements of foot-mounted INS.

[0130] Table 1 Zero speed determination results in normal walking state

[0131]

[0132] (2) Zero speed determination experiment in fast walking mode

[0133] Five groups of fast walking data were collected, and the traditional zero-speed judgment algorithm was used to manually adjust the threshold to make a zero-speed judgment on the fast walking state, and the judgment result was used as a comparison benchmark. The first group of data was then used to train the SVM classifier, and then the five groups of data were classified for zero-speed judgment, and the classification accuracy was observed and compared. Table 2 shows the classification results of the five groups of fast walking data. The classification accuracy rate reached more than 97%, and the SVM classifier had a high accuracy for zero-speed judgment of normal walking and fast walking. The SVM classification results of the second group of data are shown in Table 2. Figure 3 shown.

[0134] Table 2 Results of zero speed determination in fast walking state

[0135]

[0136] (3) Zero-speed determination experiment of multi-step walking mode

[0137] In practical applications, it is impossible for people to maintain a constant cadence when walking. Sometimes they walk slowly, and sometimes they walk quickly. Therefore, the zero-speed gait determination algorithm based on SVM must adapt to different cadences. In essence, the faster the cadence, the more significant the dynamic characteristics of the foot when landing, that is, the larger the output value of the inertial sensor, and the larger the difference between the characteristic values ​​of slow walking and fast walking, so retraining is needed. The 5 groups of normal walking data and the 5 groups of fast walking data are combined one by one, and the first group is used for SVM training. Then the 5 groups of data are classified separately to observe the accuracy of the classification results, as shown in Table 3. The SVM classification results of the second group of data are shown in Table 3. Figure 4 shown.

[0138] Table 3 Zero speed determination results of multi-step frequency walking state

[0139]

[0140] 3. Experiment on zero speed determination of running mode based on SVM

[0141] (1) Zero speed determination experiment in normal running mode

[0142] For the running gait pattern, the foot touches the ground for a very short time during running. The mean square error of the gyro and accelerometer modulus values ​​is used when performing zero-speed judgment according to formula (5). Since the mean square error needs to be calculated by using a sliding window to intercept a period of data for calculation, theoretically, part of the zero-speed state time may be lost. Therefore, when performing the running state zero-speed judgment calculation, only the gyro and accelerometer output modulus values ​​are used to perform zero-speed judgment. For the 5 groups of original data measured by normal pace running, according to the SVM parameters set by the present invention, the first group is used for SVM training, and the 5 groups of data are reclassified using the trained SVM classifier, and compared with the zero-speed judgment results obtained by the original method.

[0143] Table 4 shows the accuracy statistics. It can be seen that the zero-speed determination results of normal running state obtained by using SVM classifier are basically consistent with the original determination results, and the average accuracy rate reaches more than 98%. The SVM classifier can complete the zero-speed determination of normal running state. Figure 5 shown.

[0144] Table 4 Zero speed determination results in normal running state

[0145]

[0146] (2) Zero speed determination experiment in fast running mode

[0147] In a combat environment, forced march mode or fast running mode often occurs. At this time, the gait of individual navigation is more intense than normal running. The output values ​​of the gyroscope and accelerometer are larger when the foot lands, and the time interval between landing is also very small. This experiment collected 5 groups of fast running data, and used the first group of fast running data for SVM learning and training to test the effect of the SVM classifier. Table 5 shows the results of the fast running experiment. Statistics show that the average accuracy rate during fast running is more than 99%, and the SVM classifier can achieve zero speed judgment during fast running. The SVM classification results of the second group of data are shown in Figure 5. Figure 6 shown.

[0148] Table 5 Results of zero speed determination in fast running state

[0149]

[0150]

[0151] (3) Zero-speed determination experiment of multi-cadence running mode

[0152] In order to verify the effect of the SVM classifier under multi-cadence running state, 5 groups of normal pace running data and 5 groups of fast running data were combined respectively to obtain 5 groups of multi-cadence running data. The first group was used for SVM training and learning, and then the 5 groups of mixed pace data were classified respectively using the learned SVM classifier to observe the classification effect. Table 6 shows the results of multi-cadence running state determination. Statistics show that the accuracy of zero-speed determination in multi-cadence mode is over 98%, which is consistent with the accuracy of single-cadence running mode, showing the good robustness and adaptability of the SVM classifier. The SVM classification results of the second group of data are shown in Table 6. Figure 7 shown.

[0153] Table 6 Zero speed determination results of multi-step running state

[0154]

[0155] 4. Multi-gait zero-speed determination method based on SVM

[0156] (I) Problem of zero-speed determination of single-gait SVM

[0157] In a combat environment, a soldier often needs to change his gait mode, and the modes such as walking, fast walking, running, and forced marching and running are constantly switched, so the zero-speed judgment algorithm of the single-soldier navigation must be able to adapt to the multiple gait modes of the single-soldier navigation. The running state and the walking state are two completely different gait modes. Compared with the walking and running states, the output module value, the mean square error of the module value, the average value, the maximum and minimum values ​​and other characteristic values ​​of the micro-inertial sensor when the foot lands on the ground are all quite different. The SVM classifier trained during normal walking cannot accurately classify the zero-speed state during running. The SVM classifier trained by the normal walking data of the present invention is used to classify the characteristic data during normal running and fast running. The classification results are shown in Table 7. It can be seen that the accuracy of zero-speed classification is greatly reduced, and the accuracy of zero-speed judgment during fast running is less than 80%, indicating that the faster the step frequency, the lower the accuracy. Figure 8 This is a comparison chart of the results of using the normal walking SVM classifier to classify the running data. It can be seen that the classification results and the original judgment results no longer overlap, but there are more misjudgment states. Therefore, the normal walking SVM classifier cannot be used for zero speed judgment when running.

[0158] Table 7 Classification results of normal walking SVM classifier for running data

[0159]

[0160] (II) Design of zero-speed judgment classifier for multiple gait patterns

[0161] When a soldier is walking on foot, the walking state and the running state are two states with relatively large differences. For the traditional zero-speed determination algorithm, the determination thresholds of each characteristic value when running are different from those when walking. In order to accurately determine the moment of landing at zero speed when running, the corresponding threshold should be adjusted according to the gait pattern. Therefore, the walking and running modes can be determined first, and then different determination thresholds can be set for zero-speed determination.

[0162] In the single gait mode, the SVM classifier can better determine the walking or running state at zero speed. However, the single gait SVM classifier cannot cross-classify the zero speed, which means that the parameters of the SVM classifiers in different gait modes are completely different. In order to improve the discrimination and accuracy of the SVM classifier, it is necessary to distinguish between the walking and running modes.

[0163] It is difficult to solve multi-classification problems with SVM. The classic SVM classifier can only solve binary classification problems, while in many practical applications, it is often necessary to solve multi-class classification problems. The present invention aims to further solve this problem by combining binary classifiers to solve the multi-classification problem of SVM classifiers.

[0164] In order to improve the accuracy of zero-speed judgment in multi-gait mode, the SVM classifier is used to distinguish between walking and running states, and to distinguish between zero and non-zero speed states in walking and running. The two-category (zero speed, non-zero speed) classification problem in single gait is transformed into a four-category (walking zero speed, walking non-zero speed, running zero speed, running non-zero speed) classification problem in multi-gait mode. Fig. 9 As shown. For the foot-strap INS, whether walking or running, it is only necessary to determine the zero speed state and the non-zero speed state, so classification is required in the end, and the walking zero speed and the running zero speed are uniformly classified as the zero speed state, and then the zero speed correction is performed. The classification results of the walking and running states can also be used for the adaptive control of the filter parameters of the multi-sensor assisted individual navigation to improve the filtering and navigation accuracy.

[0165] (1) SVM feature value selection and sample set generation in multi-gait mode

[0166] The eigenvalues ​​used by the SVM classifier in a single gait can determine the zero-speed and non-zero-speed states, but cannot distinguish between running and walking states, so it is necessary to increase the eigenvalues. After screening, the mean of the gyroscope and accelerometer modulus, the maximum and minimum modulus values ​​in the sliding time window, and other parameters are effective eigenvalues ​​for distinguishing between running and walking. The output values ​​of the gyroscope and accelerometer are relatively large, so they are multiplied by the sampling period Δt = 0.01s to become a small value, which is beneficial to improve the accuracy of SVM classification. In order to design an SVM classifier suitable for multi-gait modes, the following eigenvalues ​​are added:

[0167] Gyro output modulus mean:

[0168]

[0169] Accelerometer output modulus mean:

[0170]

[0171] Maximum and minimum gyroscope modulus values:

[0172]

[0173] The maximum and minimum values ​​of the accelerometer output modulus:

[0174]

[0175] The added 6 types of eigenvalues ​​are combined with the 4 types of eigenvalues ​​defined for single-gait zero-speed judgment in the present invention into a 10-dimensional feature vector as shown in formula (22). The multi-gait zero-speed judgment SVM classifier uses the 10-dimensional feature vector for learning, training and classification.

[0176] [Mod_w Mod_aσ Mod_w σMod_a W_mean A_mean Max_w Min_w Max_aMin_a] (twenty two)

[0178] The mixed gait sample data of walking and running are classified. In order to improve the classification accuracy of the sample data, different judgment thresholds are set for walking and running respectively by manually adjusting the judgment threshold, which can obtain a more accurate sample data set.

[0179] In order to utilize the existing relatively accurate zero-speed determination results, the 5 groups of normal walking and fast walking data and the 5 groups of normal running and fast running data collected above are combined into 5 groups of multi-step frequency and multi-gait mixed data, and the first group of mixed gait data is used to learn and train the SVM classifier.

[0180] (2) Setting and training of SVM classifier parameters in multi-gait mode

[0181] Based on the LIBSVM software package, the SVM classifier is trained using a sample data set containing 10-dimensional feature vectors to obtain support vectors that support multi-gait zero-speed classification. The SVM parameters are set as follows:

[0182] The multi-gait zero-speed judgment SVM classifier is divided into 4 categories:

[0183] H0: walking speed zero, H0=0;

[0184] H1: walking speed is not zero, H1=1;

[0185] H2: zero running speed, H2=2;

[0186] H3: Running speed is not zero, H3=3.

[0187] SVM type: C-SVM

[0188] SVM kernel function selection: radial basis kernel, as shown in formula (9).

[0189] The gamma function setting in the kernel function is: g=2.8.

[0190] SVM penalty coefficient: C=1.8.

[0191] The LIBSVM library uses the One-Versus-One (OVO) method for multi-classification SVM learning and classification. The approach is to design an SVM classifier between any two categories. Therefore, the multi-gait mode zero-speed judgment SVM classifier needs to design 6 SVM classifiers. The multi-gait zero-speed judgment SVM classifier is divided into four categories: H0, H1, H2, and H3. During training, the feature vectors corresponding to H0~H1, H0~H2, H0~H3, H1~H2; H1~H3, H2~H3 are selected as sample training sets. Through training, 6 trained SVM classifiers can be obtained. When classifying new gait data, the 6 classifiers are used to classify and test the data to be tested, and then voting is adopted. The category with the most votes is the classification of the current gait data to be tested. The LIBSVM library can directly process the multi-gait classification SVM. The processing process is similar to the design of the single-gait SVM classifier. The only difference is that the number of classifications is changed from 2 to 4.

[0192] (III) Experiment on zero-speed determination of multiple gait patterns based on SVM

[0193] The five groups of data of normal walking, fast walking, normal running and fast running were merged respectively to obtain five groups of feature data with different step frequencies and different step patterns. The first group of data was used for SVM learning and training, and then the five groups of data were classified using the learned SVM classifier and compared with the results obtained by the original traditional zero-speed judgment algorithm.

[0194] The results of 5 groups of multi-gait zero-speed judgments are as follows Figure 10 to Figure 14 Table 8 is a statistical table of zero-speed determination accuracy. From the results of multi-gait and multi-step frequency zero-speed determination, it can be seen that the average accuracy of multi-gait classification is more than 98%, reaching the accuracy of single-gait SVM classifier. When SVM processes multi-classification problems, it converts multi-classification problems into multiple binary classification problems, so the total number of SVM support vectors is more than that of single-gait classification, and more zero-speed determination time is required.

[0195] In the case of multiple gaits and multiple steps, there are some misidentifications of running and walking states, which is inevitable in actual sports. The gyro and accelerometer output characteristics of jogging and fast walking are very similar, so it is easy to misjudge. Although gait patterns may cause misjudgment, for the ZUPT of the foot-bound INS, the main requirement is to accurately determine the zero speed, regardless of whether it is walking zero speed or running zero speed, so its actual zero speed determination accuracy is higher.

[0196] Table 8 Zero speed determination results of multi-gait and multi-step frequency data

[0197]

[0198]

[0199] 5. SVM-based zero-speed detection pedestrian navigation experiment

[0200] (I) Experimental preparation and experimental steps

[0201] An actual pedestrian navigation experiment was conducted in the office corridor of the Navigation Engineering Center of Tsinghua University. Fig.15 As shown, the starting point coincides with the end point, wherein segment B is in a jogging state and segment C is in a fast running state, and the others are in a normal uniform walking state.

[0202] (II) Experimental results

[0203] In order to verify the effectiveness of the new algorithm, the following three zero-speed judgment algorithms are used for comparison.

[0204] Algorithm 1: Use fixed threshold judgment method to perform zero speed judgment.

[0205] Algorithm 2: Adopting an adaptive threshold determination method, this method is difficult to implement in practice. The present invention adopts a method of manually determining the gait type of the original data afterwards, and sets different zero-speed determination thresholds according to the gait and step frequency. This method is relatively accurate, but difficult to implement in practice.

[0206] Algorithm 3: A zero-speed judgment algorithm based on SVM gait pattern recognition is used to learn the SVM classifier using known data, and then the trained SVM classifier is used to perform real-time zero-speed judgment in pedestrian navigation experiments.

[0207] The normal walking and running data obtained above are used to train the SVM classifier, and then the entire navigation experiment data is classified. The classification results are as follows: Fig.16 shown.

[0208] Number of supported vectors: 2096

[0209] Classification accuracy: 97.3625% (11591 / 11905)

[0210] Zero speed determination accuracy: 97.81%

[0211] The accuracy determination is a relative concept. By using the classification results to conduct pedestrian navigation experiments and observing the navigation errors using different zero-speed detection methods, a more accurate judgment can be obtained.

[0212] Using three different algorithms, the results of zero speed determination can be obtained respectively, such as Fig.17As shown in the figure, the zero-speed determination algorithm using Algorithm 1 cannot detect the zero-speed moment of landing when running, especially when running fast. The zero-speed determination results of Algorithm 2 and Algorithm 3 are relatively consistent, with an accuracy rate of 97.81%. However, the similarity between Algorithm 1 and Algorithm 2 is only 82.13%, and the similarity between Algorithm 1 and Algorithm 3 is 83.62%.

[0213] The above three algorithms are used to perform offline calculations of pedestrian navigation, and the trajectories obtained are as follows: Fig.18 As shown in the table, the coordinates of the starting point are (0,0), and ideally the coordinates of the end point are also (0,0), but due to the error, the actual end point coordinates do not coincide with the starting point coordinates, and the end point coordinates are the errors, as shown in the table. The horizontal walking distance of the foot-bound INS under the three algorithms is calculated, as shown in Table 9.

[0214] Table 9 Comparison of position error and total walking distance of three algorithms

[0215]

[0216] From the comparison results of the three algorithms, it can be seen that when using Algorithm 1, the error between the starting point and the end point reaches (4.92, -5.23), which is a relatively large error. The errors of Algorithm 2 and Algorithm 3 are relatively close, and the end point error is less than 2 meters. The total walking distance of Algorithm 2 and Algorithm 3 is also relatively close. The total walking distance of Algorithm 1 is about 11 meters different from that of Algorithm 2 and Algorithm 3, and the distance error is relatively large.

[0217] The present invention mainly studies a new zero-speed determination algorithm based on SVM gait pattern recognition. To verify the applicability of the new algorithm, 5 groups of data of normal walking state, fast walking state, normal running and fast running state were collected respectively, and then the general algorithm and the new algorithm were used to carry out zero-speed determination experiments respectively. In the single gait mode, a two-class SVM classifier was designed, and the accuracy of the SVM classifier in gait modes such as normal walking, fast walking, multi-frequency walking, normal running, fast running, and multi-step frequency running was verified respectively. The experimental results show that compared with the general algorithm, the accuracy reaches more than 97%. In order to improve the applicability of the new algorithm, according to the actual situation, a four-class SVM classifier was designed. By combining the collected data of various step frequencies and various gaits, multi-gait and multi-gait motion states are obtained. After the classifier is trained, the four-class SVM classifier can accurately separate the running and walking states, and can determine the respective zero-speed states, and the accuracy rate also reaches more than 97%.

[0218] In order to further verify the effect of the method of the present invention, a group of pedestrian navigation data with mixed gaits was collected. The experimental results show that the endpoint error of the navigation solution of the method of the present invention using the new algorithm is equivalent to or even higher than the accuracy of the manual adaptive zero-speed judgment algorithm, and is much higher than the accuracy of the ordinary algorithm.

[0219] The zero-speed determination algorithm based on SVM gait pattern recognition solves the problem of adaptive change of the determination threshold in the commonly used zero-speed determination algorithm, greatly improving the adaptability and accuracy of the zero-speed determination algorithm. The algorithm has moderate computational complexity and can be used for real-time calculation of pedestrian navigation, with high engineering practical value.

Claims

1. A method for determining zero speed of a single-gait pedestrian's foot based on a support vector machine algorithm, comprising: Get gyro and accelerometer output values ​​as raw data; Preprocess the raw data; Design a zero-speed determination classifier based on SVM; Training the SVM zero-speed determination classifier, continuously optimizing SVM parameters, and obtaining a structural modeling SVM classifier for zero-speed determination; The data to be tested is sent to the structural modeling SVM classifier for zero speed determination to obtain a zero speed determination result; The parameters of the structure modeling SVM classifier are as follows: Number of support vectors: nSV = 894; The constant term of the decision function is: b＝-0.197718 The SVM support vector is: The support vector coefficients are: <h2 style=";text-align:left;direction:ltr">W=[w1 w2 w3...w<h2 style=";text-align:left;direction:ltr"> 894 <h2 style=";text-align:left;direction:ltr"> ] The SVM high-dimensional feature space classification decision function becomes: The single-gait SVM zero-speed classifier judgment formula is: Since the label values ​​of the SVM zero-speed classifier are set to 1 and 0, the label values ​​of the SVM classification results are converted. j The judgment label is d j , the determination formula is as follows:

2. The method for determining zero speed of a single-gait pedestrian's foot according to claim 1, characterized in that: The single-gait pedestrian foot zero-speed determination method also includes: performing a walking mode zero-speed determination experiment and a running mode zero-speed determination experiment on the structural modeling SVM classifier; the walking mode zero-speed determination experiment includes: a normal walking mode zero-speed determination experiment, a fast walking mode zero-speed determination experiment, and a multi-step walking mode zero-speed determination experiment; the running mode zero-speed determination experiment includes: a normal-step running mode zero-speed determination experiment, a fast running mode zero-speed determination experiment, and a multi-step running mode zero-speed determination experiment.

3. The method for determining zero speed of a single-gait pedestrian's foot according to claim 1, characterized in that: The design of the zero-speed determination classifier based on SVM includes: Select the feature value of SVM classifier; Set the judgment threshold and generate a sample set; The sample set is judged according to the judgment criterion to obtain a zero-speed judgment result vector; Set the SVM parameters.

4. The method for determining zero speed of a single-gait pedestrian's foot according to claim 3, characterized in that: The characteristic values ​​of the SVM classifier are: gyro output modulus, accelerometer output modulus, gyro output modulus mean square error, accelerometer output modulus mean square error; wherein the sampling time interval Δt=0.01s, wherein, The gyro output modulus is: Mod_w'=(w x 2 +w y 2 +w z 2 ) 1 / 2 Δt Accelerometer output modulus: Mod_a'=|(a x 2 +a y 2 +a z 2 ) 1 / 2 -g0|Δt The mean square error of the gyro output modulus is: The mean square error of the accelerometer output modulus is:

5. The method for determining zero speed of a single-gait pedestrian's foot according to claim 4, characterized in that: The determination thresholds are respectively: Thu_w, Thu_a, Thu_a_σ, Thu_w_σ. The four characteristic values ​​and the four determination thresholds together constitute a determination criterion. Under this determination criterion, the zero speed determination result is as follows: Among them, Stance_d is 0, indicating a zero-speed state, and the foot is stationary on the ground; Stance_d is 1, indicating a non-zero-speed state, and the foot is in a kicking motion state.

6. The method for determining zero speed of a single-gait pedestrian's foot according to claim 3, characterized in that: The setting of SVM parameters includes: dividing the gait into two states of zero speed and non-zero speed, setting the parameters of the SVM classifier to two categories, and setting the SVM related parameters as follows after repeated verification: Number of SVM categories: 2 categories, H0: zero speed state, H1: non-zero speed state; SVM label value: H0=0, H1=1; SVM type: C-SVM; SVM kernel function selection: radial basis kernel function, as shown in the following formula: The gamma function setting in the kernel function is: g = 2.8; SVM penalty coefficient: C = 6.2.

Images