Lower limb motion state recognition model training method and lower limb motion state recognition method

Abstract

The embodiment of the present disclosure provides a lower limb motion state recognition model training method and a lower limb motion state recognition method. The model training comprises: determining the sorting order of the binary classification models corresponding to various lower limb motion states, and sequentially training the binary classification models according to the sorting order to obtain model parameters and current classification results of each sample data; and training a result decision model according to the binary classification results obtained by processing the sample data by the binary classification models corresponding to all lower limb motion states and the lower limb motion state labels of the sample data, and determining the model parameters of the result decision model. Through the reference of the previous binary classification result output result, the correct result of the subsequent binary classification model can be realized as much as possible, the model accuracy and robustness of the subsequent binary classification model are improved, and then the accuracy and robustness of the entire lower limb motion state recognition model are improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of motion feature recognition, in particular to a lower limb motion state recognition model training method and a lower limb motion state recognition method. BACKGROUND

[0002] With the wide application of exoskeleton robots in military, industry, medical and other fields, human lower limb motion state recognition has become an important research field in order to achieve more coordinated and compliant assistance of human motion. Due to the working reliability, easy accessibility, low weight, low power consumption and low noise of inertial measurement unit itself, motion state recognition based on inertial measurement unit measurement data is one of the mainstream technologies in the industry.

[0003] At present, the industry has proposed a variety of lower limb motion state recognition methods based on inertial measurement unit measurement signals, including machine learning model recognition methods based on support vector machines, random forests, and deep learning methods such as convolutional neural networks and recurrent neural networks. However, these methods still have problems such as poor accuracy and robustness in dealing with specific types of lower limb motion state recognition (such as going upstairs and downstairs), and motion state recognition switching is not smooth. SUMMARY

[0004] In order to solve the problem of poor recognition accuracy of existing lower limb motion state recognition for certain specific lower limb motion states, the embodiments of the present disclosure provide a new lower limb motion state recognition model training method and a lower limb motion state recognition method

[0005] The embodiments of the present disclosure provide a lower limb motion state recognition model training method, the lower limb motion state recognition model comprising a variety of binary classification models corresponding to various lower limb motion states and a result decision model; the method comprises,

[0006] Based on the inertial measurement signals collected at various lower limb motion states, corresponding motion features are determined, the lower limb motion state is taken as the label of the sample data, and the corresponding motion features and the lower limb motion state form the sample data, and the sample data set is constructed by using the sample data;

[0007] The sorting order of the binary classification models corresponding to various lower limb motion states is determined, and the binary classification models corresponding to each lower limb motion state are trained according to the sorting order in the following S1-S2 in turn until the model parameters of the binary classification models corresponding to all lower limb motion states and the binary classification results obtained by processing the sample data are obtained:

[0008] S1: obtaining the preceding binary classification result obtained by processing the motion features by the binary classification model corresponding to the lower limb motion state sorted in the front;

[0009] S2: taking the motion features in the sample data corresponding to the current target lower limb motion state and the corresponding previous binary classification result as input features of positive samples, taking the motion features in the sample data corresponding to non-current target lower limb motion state and the corresponding previous binary classification result as input features of negative samples, training the binary classification model corresponding to the current target lower limb motion state, obtaining the model parameters of the binary classification model corresponding to the current target lower limb motion state and the current classification result of each sample data;

[0010] The result decision model is trained according to the binary classification results obtained by processing the sample data according to the binary classification models corresponding to all lower limb motion states and the lower limb motion state labels of the sample data, and the model parameters of the result decision model are determined.

[0011] Optionally, the method further comprises:

[0012] For the sample data of various lower limb motion states, the correlation scores of each motion feature and the label are determined, and a preset number of selected relevant features with higher correlation with the label of the sample data are selected from the motion features according to the correlation scores;

[0013] For the sample data of various lower limb motion states, a preliminary regression analysis is performed based on the corresponding selected relevant features and the label, and a regression analysis accuracy result corresponding to each motion state is determined.

[0014] The regression analysis accuracy result from best to worst is taken as the sorting order of the binary classification models corresponding to various lower limb motion states.

[0015] Optionally, the method further comprises:

[0016] The inertial measurement signals collected during various lower limb motion states are sampled according to a preset sliding time window and a sliding window frequency, respectively, to obtain a sorted kinematic feature sequence in the sliding time window and a sorting identifier of the kinematic feature sequence.

[0017] The motion features are determined based on the kinematic feature sequence, and the sorting identifier of the kinematic feature sequence is taken as the sorting identifier of the sample data.

[0018] In the case that the lower limb motion state sorted in the front is a continuous state, the method further comprises: according to the sorting identifier of each sample data, obtaining a previous historical binary classification result obtained by processing the previous sample data according to the binary classification model corresponding to the lower limb motion state sorted in the front.

[0019] The input features of the motion features in the sample data corresponding to the current target lower limb movement state and the corresponding previous binary classification results as positive samples include: the motion features in the sample data corresponding to the current target lower limb movement state, the corresponding previous binary classification results, and the corresponding previous historical binary classification results as input features of positive samples.

[0020] The input features of the motion features in the sample data corresponding to the current target lower limb movement state and the corresponding previous binary classification results as positive samples include: the motion features in the sample data corresponding to the current target lower limb movement state, the corresponding previous binary classification results, and the corresponding previous historical binary classification results as input features of positive samples.

[0021] Optionally, before training the binary classification model corresponding to the current target lower limb movement state, the method further comprises:

[0022] Correlation analysis is performed on the labels of the sample data and a motion feature based on a machine learning model to determine first correlation scores of various motion features with the current target lower limb movement state;

[0023] A plurality of target samples are randomly selected from a sample data set, and N nearest positive samples in a motion feature dimension closest to each of the target samples in positive samples are determined, and N nearest negative samples in the motion feature dimension closest to each of the target samples in negative samples are determined;

[0024] First average distances of the target samples and the N nearest positive samples in the motion feature dimension are calculated, second average distances of the target samples and the N nearest negative samples in the motion feature dimension are calculated, a distance difference value of the first average distance and the second average distance is calculated, and a second correlation score of the motion feature is determined according to the distance difference value;

[0025] The first correlation score and the second correlation score corresponding to each motion feature are weighted and summed to obtain a comprehensive score corresponding to each motion feature;

[0026] Based on the comprehensive score corresponding to each motion feature, a target motion feature used to represent the current target lower limb movement state is determined.

[0027] The input features of the motion features in the sample data corresponding to the current target lower limb movement state and the corresponding previous binary classification results as positive samples include: the motion features in the sample data corresponding to the current target lower limb movement state, the corresponding previous binary classification results, and the corresponding previous historical binary classification results as input features of positive samples.

[0028] The target motion feature in the sample data corresponding to the current target lower limb motion state and the previous two classification results corresponding to the target motion feature are taken as input features of positive samples, and the target motion feature in the sample data corresponding to a non-current target lower limb motion state and the previous two classification results corresponding to the target motion feature are taken as input features of negative samples, and the two classification model corresponding to the current target lower limb motion state is trained.

[0029] Optionally, the second correlation score of the one motion feature is determined according to the distance difference value, and the second correlation score of the one motion feature comprises:

[0030] In the case that the distance difference value is a positive value, the initial weight is reduced to obtain a corresponding reduced weight; and

[0031] In the case that the distance difference value is a negative value, the initial weight is increased to obtain a corresponding increased weight.

[0032] The average of all reduced weights and the increased weight is taken as the second correlation score of the one motion feature.

[0033] Optionally, the target motion feature used for representing the current target lower limb motion state is determined based on the comprehensive score corresponding to each motion feature, and the target motion feature used for representing the current target lower limb motion state comprises:

[0034] The comprehensive score corresponding to each motion feature is determined based on the comprehensive score corresponding to each motion feature.

[0035] The correlation of the selected motion feature is determined, and the irrelevant motion feature and the relevant motion feature are determined, and the selected relevant motion feature is determined in the relevant motion feature.

[0036] The irrelevant motion feature and the selected relevant motion feature are taken as the target motion feature representing the target lower limb motion state; or

[0037] The correlation of each motion feature is analyzed, and the irrelevant motion feature and the relevant motion feature are determined, and the selected relevant motion feature is determined in the relevant motion feature.

[0038] The target motion feature is determined in the irrelevant motion feature and the selected relevant motion feature according to the comprehensive score of the irrelevant motion feature and the selected relevant motion feature.

[0039] Optionally, the first correlation score of each motion feature and the current target lower limb motion state is determined by using a machine learning model to perform correlation analysis based on the label of the sample data and one motion feature, and the first correlation score of each motion feature and the current target lower limb motion state comprises:

[0040] Correlation analysis is performed on the label and the motion feature based on the sample data by using a logistic regression model, a feature coefficient corresponding to the motion feature is determined in the logistic regression model, and an absolute value of the feature coefficient is taken as a first correlation score of the motion feature and the current target lower limb motion state; or

[0041] Correlation analysis is performed on the label and the motion feature based on the sample data by using a Gaussian naive Bayes model, and standard deviations under the current target lower limb motion state and the non-current target lower limb motion state are respectively determined.

[0042] An average value of the standard deviations of the current lower limb state and the non-current lower limb state corresponding to the motion feature is taken as the first correlation score of the motion feature and the current target lower limb motion state.

[0043] Optionally, the training of the binary classification model corresponding to the current target lower limb motion state is performed to obtain model parameters of the binary classification model corresponding to the current target lower limb motion state and a current classification result of each sample data, including: training various types of binary classification models corresponding to the current target lower limb motion state respectively to obtain model parameters of the various types of binary classification models, and determining model quality scores of the various types of binary classification models.

[0044] The binary classification model with the optimal model quality score is selected as the selected binary classification model corresponding to the current target lower limb motion state, and a current classification result corresponding to the selected binary classification model is obtained.

[0045] In a second aspect, the embodiments of the present disclosure provide a lower limb motion state recognition method, including:

[0046] Based on the real-time collected lower limb motion inertial measurement signal, a corresponding motion feature is determined.

[0047] According to a predetermined sorting order of the lower limb motion state binary classification models, the motion feature is processed by the binary classification models corresponding to the lower limb motion states in sequence, to obtain corresponding binary classification results, wherein a binary classification model corresponding to a subsequent lower limb motion state processes the motion feature and a previous binary classification result to obtain a subsequent binary classification result, and the previous binary classification result is a classification result obtained by processing the motion feature by a previous lower limb motion state binary classification model.

[0048] All the binary classification results corresponding to the lower limb motion states are input into a result decision model to obtain a recognized lower limb motion state.

[0049] The lower limb motion state binary classification models and the result decision model are obtained by using the lower limb motion state recognition model training method.

[0050] In a third aspect, the embodiments of the present disclosure provide a computing device, comprising a processor and a memory, the memory being configured to store a computer program; the computer program, when loaded by the processor, causes the processor to execute the lower limb motion state recognition model training method and / or the lower limb motion state recognition method.

[0051] The lower limb motion state recognition model training method provided by the embodiments of the present disclosure determines the sorting order of the binary classification model corresponding to each lower limb motion state before training the binary classification model in the lower limb motion state recognition model, and then trains each binary classification model in sequence (cascade) according to the sorting order, and takes the binary classification result obtained by processing sample data by the preceding binary classification model as input when training the subsequent binary classification model, so that the training of the preceding binary classification model has an impact on the training of the subsequent binary classification model. By referring to the output result of the preceding binary classification result, the subsequent binary classification model can be made to output correct results as much as possible, and the model accuracy and robustness of the subsequent binary classification model are improved, and thus the accuracy and robustness of the entire lower limb motion state recognition model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure, together with the description.

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, those skilled in the art can obtain other drawings from these drawings without creative labor, and

[0054] Figure 1 is a flowchart of the lower limb motion state recognition model training method provided by the embodiments of the present disclosure;

[0055] Figure 2 is a flowchart of determining motion features based on inertial measurement signals according to some embodiments;

[0056] Figure 3 is a flowchart of determining the sorting order of the binary classification model corresponding to each lower limb motion state according to some embodiments;

[0057] Figure 4 is a flowchart of determining target motion features according to some embodiments of the present disclosure;

[0058] Figure 5 is a flowchart of the lower limb motion state recognition method provided by the embodiments of the present disclosure;

[0059] Figure 6 is a structural schematic diagram of a computing device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0060] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes, and are not intended to limit the scope of protection of the present disclosure.

[0061] The term "comprising" and variations thereof as used herein are open-ended, that is "including but not limited to". The term "based on" is "based, at least in part, on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Related terms such as "one embodiment", "an embodiment", "some embodiments", "various embodiments", "other embodiments", "another embodiment" and "at least one embodiment" mean "one, at least one, or more than one embodiment". In this document, relational terms such as "first" and "second", and the like can be used solely to distinguish one entity or action from another entity or action, without necessarily requiring or implying any actual such relationship or order between such entities or actions.

[0062] To solve the problem that the existing lower limb motion state recognition model has poor recognition accuracy and robustness for specific types of lower limb motion states, an embodiment of the present disclosure provides a new lower limb motion state recognition model training method. The lower limb motion state recognition model training method provided by the present disclosure is executed by a computing device.

[0063] It should be noted that the lower limb motion state recognition model provided by an embodiment of the present disclosure includes a variety of lower limb motion state corresponding binary classification model and result decision model.

[0064] Figure 1 is a flowchart of the lower limb motion state recognition model training method provided by an embodiment of the present disclosure. As shown in Figure 1 The lower limb motion state recognition model training method provided by the present disclosure includes S110-S170.

[0065] S110: Based on the inertial measurement signals collected at various lower limb motion states, the corresponding motion features are determined respectively, the lower limb motion state is taken as the label of the sample data, and the corresponding motion features are taken as the sample data, and the sample data set is constructed by using the sample data.

[0066] Because the human lower limb movement is the coordinated movement of the thigh and the shank, the inertial measurement unit used in actual collection and movement recognition is mostly two. Two inertial measurement units are respectively fixedly arranged relative to the thigh and the shank, and collect the movement states of the thigh and the shank respectively to obtain corresponding inertial measurement signals.

[0067] In the embodiments of the present disclosure, after obtaining the inertial measurement signals collected by the inertial measurement units fixed on the thigh and the shank, the computing device determines a plurality of movement features based on the inertial measurement signals. The following analyzes how to determine the plurality of movement features based on the inertial measurement signals in the embodiments of the present disclosure.

[0068] Figure 2 The flowchart in FIG. 1 is a flowchart of determining movement features based on inertial measurement signals provided by some embodiments. As shown in FIG. 1, in some embodiments, the computing device can determine at least part of the plurality of movement features by using S111-S112 as follows. Figure 2

[0069] S111: sampling the inertial measurement signals collected in various lower limb movement states according to a preset sliding time window and a sliding window frequency respectively to obtain a kinematic feature sequence sorted in the sliding time window.

[0070] The inertial measurement unit includes an accelerometer and a gyroscope, wherein the accelerometer can measure three-axis acceleration, and the gyroscope can measure three-axis angular velocity. After obtaining the original three-axis acceleration and the original three-axis angular velocity measured by one inertial measurement unit, the computing device can obtain three-axis acceleration and three-axis angular velocity in a reference coordinate system through data processing, and after obtaining the three-axis acceleration and the three-axis angular velocity, the corresponding roll angle, pitch angle and yaw angle are obtained through Euler angle calculation. Because the foregoing technical content is a common means in the art, the foregoing content will not be analyzed here.

[0071] As analyzed before, in the case of obtaining the kinematic features of the thigh and the shank by using two inertial measurement units respectively, 18 kinematic features can be obtained based on the inertial measurement signals at one time point, which are three-axis acceleration, three-axis angular velocity and roll angle, pitch angle and yaw angle obtained through attitude solution of the thigh and the shank respectively. Because the inertial measurement signals can be affected by the background noise and the environmental noise of the inertial measurement unit, a low-pass filter (such as a fourth-order Butterworth low-pass filter) can be used to smooth the inertial measurement signals in actual application, and the 18 kinematic features are determined by using the smoothed inertial measurement signals.

[0072] ​In practical applications, the motion state cannot be determined by the kinematic characteristics of a single time point, but by the kinematic characteristics of a sufficient number of time points. Therefore, the inertial measurement signals collected for various lower limb motion states need to be sampled using a preset sliding time window, and the kinematic characteristics in the sliding time window are determined according to the foregoing method, and the kinematic characteristics are sorted according to the time mark to obtain a sorted kinematic characteristic sequence.

[0073] In order to fully obtain a sorted kinematic characteristic sequence of a complete lower limb motion state, it is necessary to ensure that the size of the aforementioned preset sliding time window is greater than the lower limb motion cycle period. Considering that the normal walking frequency of a typical human body is 0.6-1.0 Hz (corresponding to 0.6 s for one step, and the lower limb motion cycle period is 1.2 s-2.0 s), the preset sliding time window can be set to 2 s to obtain a sorted kinematic characteristic sequence in a 2 s sliding time window.

[0074] It should be noted that, in order to obtain the inertial measurement signals and kinematic characteristics representing sufficient motion details, the correlation measurement unit should be sampled at a high frequency. For example, in a specific application, the sampling frequency of the inertial measurement unit is set to 50 Hz. In addition, in order to obtain as many sorted kinematic characteristic sequences as possible for a motion state, the sliding window step of the preset sliding time window cannot be too small. In an embodiment, the sliding step of the preset sliding time window is set to 0.02 s, that is, the preset sliding time window completes 50 sliding windows in 1 s, and 50 sorted kinematic characteristic sequences are obtained.

[0075] S112: Determine at least part of the motion characteristics based on the various kinematic characteristic sequences.

[0076] After obtaining a sorted kinematic characteristic sequence, at least one of the following parameters can be calculated as the corresponding motion characteristics for a single sorted kinematic characteristic sequence: mean value, maximum value, minimum value, median value, variance, peak-to-peak difference, signal energy, skewness, kurtosis, maximum slope, root mean square, absolute value mean, pulse factor, root amplitude, margin index, waveform factor, peak factor, vector sum amplitude, and absolute value area.

[0077] The mean represents the average value of all data in a kinematic feature sequence; the maximum represents the maximum value of all data in a kinematic feature sequence; the minimum represents the minimum value of all data in a kinematic feature sequence; the median represents the median value of all data in a kinematic feature sequence; the variance represents the dispersion degree of data in a kinematic feature sequence; the peak-to-peak difference represents the difference between the maximum value and the minimum value in a kinematic feature sequence; the signal energy represents the sum of squares of all data in a kinematic feature sequence; the skewness represents the asymmetry degree of all data in a kinematic feature sequence; the kurtosis represents the peak degree of distribution of all data in a kinematic feature sequence; the maximum slope represents the ratio of the difference between the maximum value and the minimum value to the corresponding time interval in a kinematic feature sequence; the root mean square is used to represent the effective amplitude of data in a kinematic feature sequence; the absolute value mean is the mean of the absolute values of all data in a kinematic feature sequence; the pulse factor is the ratio of the maximum value to the root mean square in a kinematic feature sequence; the square root amplitude is the square root of the root mean square value and the square root of the sequence length in a kinematic feature sequence; the margin index is the ratio of the maximum value to the square root amplitude in a kinematic feature sequence; the waveform factor represents the peak characteristic of data in a kinematic feature sequence, which is the ratio of the peak value to the root mean square; the wave crest factor represents the peak characteristic of data in a kinematic feature sequence, which is the ratio of the peak value to the mean value; and the vector sum amplitude is the amplitude of the direct sum of all values in a kinematic feature sequence. The absolute value area is the sum of the absolute values of all data in a kinematic feature sequence.

[0078] As analyzed above, the inertial measurement signal includes the thigh inertial measurement signal collected by the inertial measurement unit located on the thigh, and includes the shank inertial measurement signal collected by the inertial measurement unit located on the shank. The computing device can determine the thigh kinematic feature sequence and the shank kinematic feature sequence in the sliding time window respectively, and determine the thigh motion feature in the corresponding dimension based on the thigh kinematic feature sequence and the shank motion feature in the corresponding dimension based on the shank kinematic feature sequence by using the foregoing method.

[0079] Correspondingly, the computing device can also determine part of the motion feature according to S1121-S1122 as follows.

[0080] S1121: calculating the ratio of the thigh motion feature and the shank motion feature in the corresponding dimension, and / or the signal energy of the thigh motion feature and the shank motion feature in the corresponding dimension.

[0081] S1122: taking the ratio and / or the signal energy as at least part of the motion feature.

[0082] After obtaining the thigh action features and the shank action features in a preset sliding time window for a certain kinematic dimension, the ratio of the thigh action features and the shank action features can be calculated, or the signal energy of the thigh action features and the shank action features corresponding to the dimension can be calculated. The ratio represents the ratio of the thigh action features and the shank action features in the same kinematic dimension, and the signal energy is the sum of the squares of the same kinematic features.

[0083] In an embodiment, in order to identify whether filtering the inertial measurement signals affects the correlation between the kinematic features and the lower limb movement states, the present scheme calculates the corresponding kinematic features based on the inertial measurement signals before and after filtering, respectively, and determines the corresponding kinematic features according to the foregoing method.

[0084] After obtaining the plurality of kinematic features, the computing device then associates the plurality of kinematic features with the corresponding lower limb movement state labels to form sample data, and constructs a sample data set based on the sample data.

[0085] In some embodiments, when the sliding window sampling is performed according to the preset sliding time window and the sliding window frequency, the sorted kinematic feature sequence in the sliding time window is obtained, and the sorted identification of the kinematic feature sequence is also determined. When the kinematic features are determined based on the kinematic feature sequence, and the sample data is determined based on the kinematic features, the computing device also determines the sorted identification of the kinematic feature sequence as the sorted identification of the sample data. The role of the sorted identification will be analyzed later.

[0086] S120: Determine the sorted order of the binary classification models corresponding to the various lower limb movement states.

[0087] In the embodiments of the present disclosure, the binary classification models corresponding to the various lower limb movement states are not trained independently, but are trained in a cascade manner. The classification results obtained by the preceding binary classification model are used for training the binary classification model corresponding to the subsequent lower limb movement state.

[0088] In order to realize the cascade training, it is necessary to determine the sorted order of the binary classification models corresponding to the various lower limb movement states, so as to realize the training of the binary classification models corresponding to the various lower limb movement states in the sorted order.

[0089] In some embodiments, the sorted order of the lower limb movement states determined by artificial determination can be input into the computing device to determine the sorted order of the binary classification models corresponding to the various lower limb movement states.

[0090] In some other embodiments, the computing device can determine the sorted order of the binary classification models corresponding to the various lower limb movement states by a correlation analysis method. Figure 3 Some embodiments determine the sorted order of the binary classification models corresponding to the various lower limb movement states. As shown in FIG. 6, the method comprises the following steps:Figure 3 As shown, in some embodiments, the method of determining the ranking order includes S121-S123.

[0091] S121: For sample data of various lower limb movement states, determine the correlation score of each movement feature and the sample data label, and select a preset number of selected relevant features with higher correlation with the sample data label from the movement features according to the correlation score.

[0092] As previously analyzed, the number of movement features determined in the foregoing is large, and if all the aforementioned movement features are directly used for movement correlation analysis, it will cause excessive data operation amount. The ranking order of the binary classification model corresponding to the binary lower limb movement state is only an operation performed from a macroscopic dimension, and does not need to consider the problem of detailed accuracy. In order to reduce the operation amount, in the embodiments of the present disclosure, for sample data of various lower limb movement states, a small amount of sample data is used to analyze the correlation between each movement feature and the sample data label, and a preset number of selected relevant features with higher correlation with the sample data label are determined.

[0093] It should be noted that when performing S121 operation, for sample data of a specific lower limb movement state, the sample data corresponding to the lower limb movement state is used as positive sample, and sample data of other lower limb movement states is used as negative sample, and correlation analysis is performed.

[0094] In specific implementation, the foregoing operation can be performed by using an existing correlation analysis method such as Pearson correlation coefficient method, to obtain a preset number of selected relevant features.

[0095] S122: For sample data of various lower limb movement states, based on the corresponding selected relevant features and labels, perform preliminary regression analysis on various lower limb movement states to determine the regression analysis accuracy result corresponding to each movement state.

[0096] In specific implementation, after obtaining the selected relevant features corresponding to various lower limb movement states, a method such as logistic regression can be used to perform regression analysis, to obtain a corresponding regression analysis model, and the verification data in the sample data is used to evaluate the regression analysis model corresponding to each movement state, to determine the corresponding regression analysis accuracy result. The regression analysis accuracy result can be represented in percentage.

[0097] S123: The regression analysis accuracy result from best to worst is used as the ranking order of the binary classification model corresponding to various lower limb movement states.

[0098] After obtaining the regression analysis accuracy results by using S122, the regression analysis accuracy results are compared and ranked from good to bad (specifically, ranked from large to small) to determine the ranking order, and the ranking order can be used as the ranking order of the binary classification models corresponding to the lower limb movement states.

[0099] It should be noted here that the ranking order is only a preliminary ranking, and the ranking order is not absolutely correct.

[0100] In one specific embodiment, the lower limb movement states include running, jumping, squatting, climbing stairs, walking, and descending stairs. The ranking order determined by using S121-S123 is: squatting, running, jumping, climbing stairs, walking, and descending stairs. According to the ranking order, the numbers corresponding to the lower limb movement states are 1-6, respectively.

[0101] After performing S120, S130-S160 can be performed subsequently until the binary classification models corresponding to all lower limb movement states and the binary classification results obtained by using the binary classification models to process the sample data are determined.

[0102] S130: Obtain the previous binary classification result obtained by using the binary classification model corresponding to the lower limb movement state ranked first to process the movement features.

[0103] S140: Take the movement features in the sample data corresponding to the current target lower limb movement state i and the corresponding previous binary classification result as the input features of the positive sample, take the movement features in the sample data corresponding to the non-current target lower limb movement state and the corresponding previous binary classification result as the input features of the negative sample, train the binary classification model corresponding to the current target lower limb movement state i, and obtain the model parameters of the binary classification model corresponding to the current target lower limb movement state and the current classification result of each sample data.

[0104] As the name implies, the binary classification model is a model that implements a binary classification task. The labels of the sample data in the sample data set are the labels of various lower limb movement states, which cannot be directly used for training the binary classification model corresponding to the target lower limb movement state i. Therefore, in the embodiments of the present disclosure, the sample data corresponding to the current target lower limb movement state i and the corresponding previous binary classification result are taken as the positive sample, the sample data corresponding to the non-current target lower limb movement state and the corresponding previous binary classification result are taken as the negative sample, the binary classification model corresponding to the current target lower limb movement state i is trained, and the model parameters of the binary classification model corresponding to the current target lower limb movement state and the current classification result of each sample data are obtained.

[0105] It is noted here that the input data for the training of the binary classification model corresponding to the current target lower limb movement state i includes not only the movement features in the sample data, but also the previous binary classification result. The ordering sequence of the binary classification models corresponding to each lower limb movement state mentioned in the foregoing directly determines the previous binary classification result corresponding to the current target lower limb movement state i, which in turn affects the model parameters of the current target lower limb movement state i and the current classification result state of each sample data (the influence here is a relative influence, not an absolute influence). In actual application, because the previous binary classification result generally has high accuracy, using it for the training of the binary classification model corresponding to the current target lower limb movement state i can improve the model precision of the binary classification model corresponding to the current target lower limb movement state i.

[0106] It should also be noted that in the case where the current target lower limb movement state i is the first lower limb movement state in the ordering sequence, it does not have a corresponding previous binary classification result, and there is no need to use the so-called previous binary classification result as model input.

[0107] S150: Determine whether the current target lower limb movement state i is the last lower limb movement state in the ordering sequence; if yes, execute S160; if no, execute S170.

[0108] After determining the binary classification model corresponding to the current target lower limb movement state i, the computing device then determines whether the current target lower limb movement state i is the last lower limb movement state in the ordering sequence. If yes, it is determined that the binary classification models corresponding to all lower limb movement states have been trained, and therefore S170 can be executed; if no, the binary classification models corresponding to the lower limb movement states in the ordering sequence need to be trained, and therefore S160 and the steps S130-S150 executed in the subsequent loop are executed.

[0109] S160: Set i=i+1, and execute S130.

[0110] S170: Train the result decision model according to the binary classification results obtained by processing the sample data by using the binary classification models corresponding to all lower limb movement states and the lower limb movement state labels of the sample data, and determine the model parameters of the result decision model.

[0111] After all the binary classification models corresponding to the lower limb movement states are trained, the binary classification results obtained by processing the sample data by the binary classification models corresponding to the lower limb movement states can be obtained. Then, the binary classification results are sorted according to the same sorting order, and the binary classification result vectors of the sample data can be obtained. The binary classification result vectors are used as the input of the result decision model, and the lower limb movement state labels of the sample data are used as the output of the result absolute model. The result decision model is trained, and the model parameters of the result decision model can be obtained.

[0112] After the binary classification models trained according to the sorting order and the result decision model are obtained, the binary classification models and the result model are combined, and the lower limb movement state recognition model is obtained. How to use the lower limb movement state recognition model will be analyzed later.

[0113] As analyzed above, before the binary classification models in the lower limb movement state recognition model are trained, the sorting order of the binary classification models corresponding to the lower limb movement states is determined, and then the binary classification models are trained one by one (concatenation) according to the sorting order. The binary classification results obtained by processing the sample data by the preceding binary classification models are used as the input when the subsequent binary classification models are trained, so that the training of the preceding binary classification models has an impact on the training of the subsequent binary classification models. Since a lower limb movement state sample data has only one lower limb movement state label, by referring to the output results of the preceding binary classification results, the subsequent binary classification models can output correct results as much as possible, and the model accuracy and robustness of the subsequent binary classification models are improved, and thus the accuracy and robustness of the entire lower limb movement state recognition model are improved.

[0114] As analyzed above, in some embodiments, the sorting identifiers of the sample data can also be obtained. In the case where the sorting identifiers are obtained, in some embodiments, the following S180 can also be performed while the S130 is performed.

[0115] S180: In the case where the lower limb movement state sorted in the front is a continuous movement state, the preceding historical binary classification results obtained by processing the preceding sample data by the binary classification model corresponding to the lower limb movement state sorted in the front are obtained according to the sorting identifiers of the sample data.

[0116] For example, if the current lower limb movement state is jumping, the preceding lower limb movement states include running, and running is a continuous movement state, the historical binary classification results obtained by processing the historical sample data before the sample data by the binary classification model corresponding to running can be obtained.

[0117] In the case where the S180 is performed, the S140 is specifically the following S141. In the case where the S180 is performed, the S140 is specifically the following S141.

[0118] S141: The motion features in the sample data corresponding to the current target lower limb motion state, the corresponding previous binary classification result, and the corresponding previous historical binary classification result are taken as the input features of the positive sample; the motion features in the sample data corresponding to the non-current target lower limb motion state, the corresponding previous binary classification result, and the corresponding previous historical binary classification result are taken as the input features of the negative sample.

[0119] By taking the historical binary classification result obtained from the historical sample data before processing the sample data of the previous continuous motion state as the input to train the current target lower limb motion state, some continuous lower limb motion states can be effectively identified, and thus the classification of the current lower limb motion state for the aforementioned continuous lower limb motion state is more accurate, and the recognition accuracy of the binary classification model corresponding to the current lower limb motion state is improved.

[0120] In the foregoing scheme, all motion features are used to train the binary classification model corresponding to the current lower limb motion state. As previously analyzed, the number of all motion features is too large, causing the model of the binary classification model to be too complex, and overfitting problem may occur. To solve this problem, in specific implementation, the motion features with high correlation with the current lower limb motion state can be selected as target motion features from all motion features, and only the target motion features and the previous binary classification result are used to train the binary classification model corresponding to the current lower limb motion state.

[0121] Figure 4 is a method flowchart for determining target motion features adopted by some embodiments of the present disclosure. As shown in Figure 4 in some embodiments, the method for determining target motion features includes S210-S250 as follows.

[0122] S210: A machine learning model is used to perform correlation analysis based on the labels of sample data and a motion feature, to determine a first correlation score of various motion features with the current target lower limb motion state.

[0123] Before determining the first correlation score of various motion features with the current target lower limb motion state, the data with the current target lower limb motion label as the positive sample and the data with the non-current target lower limb motion label as the negative sample need to be used for correlation analysis.

[0124] In specific implementation, the computing device performs the aforementioned correlation analysis by using a predetermined machine learning model to obtain the first correlation score. The machine learning model can be any one of a logistic regression model, a latent Dirichlet allocation model, a K-nearest neighbors algorithm model, a decision tree model, a Gaussian naive Bayes analysis model, an extreme gradient boosting tree model, a light gradient boosting machine model, a gradient boosting-based decision tree model, a random forest model, and a generalized additive model, and the present disclosure is not particularly limited thereto.

[0125] In some embodiments, the computing device performs correlation analysis on the label of the sample data and one motion feature by using a logistic regression model to determine a first correlation score of one motion feature with the current target lower limb movement state, specifically including the following S211-S212.

[0126] S211: Correlation analysis is performed based on the label of the sample data and one motion feature respectively by using a logistic regression model to determine a feature coefficient of one motion feature in the logistic regression model.

[0127] The basic form of the logistic regression model is where X f represents one motion feature, w f is a feature coefficient, which represents the influence of one motion feature X f on the label of the sample data, and w f represents the influence degree of one motion feature on the output result.

[0128] S212: The absolute value of the feature coefficient is taken as the first correlation score of one motion feature with the current target lower limb movement state.

[0129] As previously analyzed, the greater the absolute value of w f is, the greater the influence of one motion feature on the output result is, and therefore the absolute value of the feature coefficient can be taken as the first correlation score of one motion feature with the current target lower limb movement state, that is, MLscore(X f ) = |w f |.

[0130] In some other embodiments, the computing device performs correlation analysis on the label of the sample data and one motion feature by using a Gaussian naive Bayes model to determine a first correlation score of one motion feature with the current target lower limb movement state, specifically including the following S213-S214.

[0131] S213: Correlation analysis is performed based on the label of the sample data and one motion feature by using a Gaussian naive Bayes model to determine the standard deviation under the current target lower limb movement state and the non-current target lower limb movement state respectively.

[0132] S214: Obtain the average of the standard deviations of the current target lower limb state and the non-current target lower limb motion state corresponding to one motion feature as the first correlation score of one motion feature and the current target lower limb motion state.

[0133] For one motion feature X f , the conditional probability of the given lower limb motion state C k is a Gaussian distribution , where μ k,f is the mean of the motion feature X k in the category C f , and δ k,f is the standard deviation of the motion feature X k in the category C f . Because the importance of the motion feature X f for the classification of the lower limb motion state C k can be measured by the change of the conditional probability, the importance of the motion feature for determining the lower limb motion state can be represented by the standard deviation of the conditional probability P(X f |C k ). Therefore, the first correlation score can be determined by obtaining the average of the standard deviations of the motion features.

[0134] In some embodiments, at least two machine learning models can be used to analyze the correlation between one motion feature and the labels of the sample data, respectively, to obtain the corresponding correlation scores, and then the corresponding correlation scores are weighted and summed to obtain the corresponding first correlation score.

[0135] S220: Randomly select multiple target samples in the sample data set, and determine the N nearest positive samples in one motion feature dimension of each target sample in the positive samples, and determine the N nearest negative samples in one motion feature dimension of each target sample in the negative samples.

[0136] Through the numerical distribution characteristics of a large number of motion features with significant correlation with the label and the numerical analysis of motion features with no significant correlation with the label, it is found that the numerical characteristics of the same class samples (same class positive samples, same class negative samples) of the motion features with significant correlation with the label are relatively close, and the numerical characteristics of the different class samples are far away, while the same class features and different class features of the motion features with no significant correlation with the label do not have the above characteristics.

[0137] ​Based on the above finding, the embodiments of the present disclosure can determine the relevance score of each motion feature and the label by whether the same-class features and the different-class features have typical clustering characteristics. Specifically, to make the finally determined relevance score typical and reliable, a plurality of target samples need to be randomly selected in the sample data set, and then for each target sample, M nearest samples in the positive sample closest to the target sample in a motion feature dimension and N nearest samples in the negative sample closest to the target sample in a motion feature dimension are determined respectively. The foregoing M and N can be the same or different, and the embodiments of the present disclosure do not make any limitation. In actual application, M and N can be 1 or a positive integer greater than 1.

[0138] S230: calculating a first average distance of the target sample and the N nearest positive samples in a motion feature dimension, and calculating a second average distance of the target sample and the N nearest negative samples in a motion feature dimension, calculating a distance difference of the first average distance and the second average distance, and determining a second relevance score of a motion feature according to the distance difference.

[0139] The first average distance of the target sample and the M nearest positive samples in a motion feature dimension is calculated by calculating the distance of the M nearest samples in a motion feature dimension and the target sample in a motion feature dimension, and then taking the average of the foregoing distances to obtain the first average distance.

[0140] Similarly, the first average distance of the target sample and the M nearest negative samples in a motion feature dimension is calculated by calculating the distance of the M nearest samples in a motion feature dimension and the target sample in a motion feature dimension, and then taking the average of the foregoing distances to obtain the first average distance.

[0141] After obtaining the first average distance and the second average distance, the distance difference of the first average distance and the second average distance can be calculated. After obtaining the distance difference, the computing device can determine the second relevance score of a motion feature based on the distance difference corresponding to all target samples according to the following method:

[0142] In the case that the distance difference is positive, the initial weight is reduced to obtain the corresponding reduced weight; and in the case that the distance difference is negative, the initial weight is increased to obtain the corresponding increased weight; and the average of all reduced weights and increased weights is taken as the second relevance score of a motion feature. The foregoing initial weight can be 0 or a value greater than 0.

[0143] S240: performing weighted summation on the first relevance score and the second relevance score corresponding to each motion feature to obtain a comprehensive score corresponding to each motion feature.

[0144] The first correlation score and the second correlation score of each motion feature are determined by the aforementioned method based on correlation analysis of a machine learning model and the method based on feature proximity analysis, respectively. Then, the first correlation score and the second correlation score can be weighted and summed to obtain a comprehensive score corresponding to each motion feature.

[0145] In some embodiments, in order to achieve better data comparison and thus make the weight setting simpler, the computing device can normalize the first correlation score and the second correlation score respectively to obtain a first normalized score and a second normalized score, and then weighted and sum the first normalized score and the second normalized score to obtain a comprehensive score corresponding to each motion feature.

[0146] Specifically, the computing device can use the formula to calculate the comprehensive score W mix,f , wherein w RFscore and w MLscore are predetermined weights, RFscore f is the first correlation score of the fth motion feature, RFscore min is the minimum value of the first correlation scores of all motion features, RFscore max is the maximum value of the first correlation scores of all motion features, MLscore f is the second correlation score of the fth motion feature, MLscore min is the minimum value of the second correlation scores of all motion features, and MLscore max is the maximum value of the second correlation scores of all motion features.

[0147] S250: Based on the comprehensive score corresponding to each motion feature, determine a target motion feature for representing the current target lower limb motion state.

[0148] After obtaining the comprehensive score corresponding to each motion feature, the computing device can sort the comprehensive scores in size to determine the sorting order corresponding to each motion feature. After determining the sorting order, the target motion feature for representing the current target lower limb motion state can be determined according to the sorting order and a predetermined selection criterion. For example, in some applications, the predetermined selection criterion is to select the L motion features in the front, and then the L motion features in the front are taken as the target lower limb motion state.

[0149] As analyzed above, the embodiment of the disclosure determines the first correlation score of each motion feature based on the machine learning method, determines the second correlation score of each motion feature based on the clustering method, and then determines the comprehensive score of each motion feature by weighting and summing the first correlation score and the second correlation score, and determines the target motion feature representing the target lower limb motion state according to the comprehensive score of each motion feature. The foregoing method determines the comprehensive score of each motion feature by comprehensively determining the comprehensive score of each motion feature by different correlation analysis methods, so that the score result of the comprehensive score is more in line with the actual situation, and then the finally selected target motion feature is more reasonable.

[0150] In other embodiments, the target motion feature representing the current target lower limb motion state can also be directly determined by the foregoing method of determining the first correlation score or the second correlation score.

[0151] In actual application, some motion features have correlation (for example, linear positive correlation). If the foregoing motion features with correlation are all used for model training, not only the number of motion features is too large during subsequent model training, but also the fitting may be poor during model training. Therefore, during the process of determining the target motion feature based on the comprehensive score of the motion feature, some motion features with correlation also need to be removed. In some embodiments, the method of determining the target motion feature based on the comprehensive score can include S261-S263 as follows.

[0152] S261: determining a candidate motion feature based on the comprehensive score corresponding to each motion feature.

[0153] Determining the candidate motion feature according to the comprehensive score corresponding to each motion feature can be sorting the comprehensive scores of the motion features, and selecting the motion features with the top pre-set number of comprehensive scores as the candidate motion features.

[0154] S262: performing pairwise correlation analysis on the candidate motion features, determining irrelevant motion features and relevant motion features, and determining selected relevant motion features in the relevant motion features.

[0155] In specific implementation, the computing device can perform pairwise correlation analysis on the candidate motion features by using a method such as Pearson correlation coefficient analysis, to obtain pairwise correlation coefficients of the candidate motion features, and specifically using The Pearson correlation coefficient r is calculated, where A and B are feature vectors constructed by using one kind of motion feature in the sample data in the sample data set, and N is the length of the feature vector.

[0156] After the two-by-two correlation coefficients of each pair of the to-be-selected motion features are calculated, the two-by-two correlation coefficients can be compared with a pre-set correlation threshold. If the two-by-two correlation coefficient is greater than the correlation threshold, the two to-be-selected motion features are determined as correlated motion features, otherwise, the two to-be-selected motion features are determined as uncorrelated features. In the case that the two to-be-selected motion features are determined as correlated motion features, the to-be-selected motion feature with a better comprehensive score can be selected as the selected correlated motion feature.

[0157] S263: After the uncorrelated motion features and the selected correlated motion features are determined, the two can be used as target motion features.

[0158] In some embodiments, the method of determining target motion features based on comprehensive scores can include S264-S265 as follows.

[0159] S264: Based on the two-by-two correlation analysis of each motion feature, uncorrelated motion features and correlated motion features are determined, and a selected correlated motion feature is determined among the correlated motion features.

[0160] Unlike the foregoing S261-S263, in this embodiment, the two-by-two correlation analysis of each motion feature is first performed to determine uncorrelated motion features and a selected correlated motion feature among the correlated motion features. In specific implementation, the same method as the foregoing S162 can be used to perform the correlation analysis to determine the uncorrelated motion features and the correlated motion features, and then the uncorrelated motion features and the selected correlated motion feature are obtained.

[0161] S265: According to the comprehensive scores of the uncorrelated motion features and the selected correlated motion features, the target motion features are determined among the uncorrelated motion features and the selected correlated motion features.

[0162] After the uncorrelated motion features and the selected correlated motion features are determined, the foregoing comprehensive scores of the motion features can be sorted to determine the sorting order, and the motion feature with the highest sorting order can be determined as the target motion feature.

[0163] In the foregoing S130, only the training of the binary classification model corresponding to the current target lower limb motion state i is mentioned, but the specific type of the binary classification model used is not mentioned. How to determine the specific type of the binary classification model corresponding to the current target lower limb state is analyzed below. In some embodiments, the computing device can use S131-S132 as follows to determine the binary classification model corresponding to the current target lower limb motion state.

[0164] S131: Various types of binary classification models corresponding to the current target lower limb motion state are trained respectively to obtain model parameters of the various types of binary classification models, and model quality scores of the various types of binary classification models are determined.

[0165] The binary classification model corresponding to the current target lower limb movement state can be a binary classification model such as a logistic regression model as previously analyzed. In actual application, the computing device can train the various types of binary classification models according to the training method of the corresponding binary classification model, and score the model quality of the various binary classification models after training ends by sampling the evaluation sample data. The aforementioned model quality score can be obtained by at least one of accuracy, precision, recall, F1 score, area under the receiver operating characteristic curve, training time, and the like.

[0166] S132: Select the binary classification model with the best model quality score as the selected binary classification model corresponding to the current target lower limb movement state.

[0167] After obtaining the model quality scores of the various types of binary classification models, the computing device can select the binary classification model with the best model quality score as the selected binary classification model corresponding to the current target lower limb state.

[0168] In actual application, in order to reduce the influence of target feature selection on the model quality score of each binary classification model, the corresponding binary classification model can be used to determine the corresponding target movement feature according to S210-S260 previously described, and then all the target movement features are used to train this type of binary classification model, and the corresponding target movement features are used to determine the model quality score of each binary classification model.

[0169] In addition to providing the lower limb movement state recognition model training method described above, the embodiments of the present disclosure also provide a lower limb movement state recognition method. Figure 5 is a flowchart of the lower limb movement state recognition method provided by the embodiments of the present disclosure. As shown in Figure 5 The lower limb movement state recognition method provided by the embodiments of the present disclosure includes S310-S330. The various binary classification models and result decision models used in the lower limb movement recognition method of the embodiments of the present disclosure are the models trained previously.

[0170] S310: Determine the corresponding movement feature based on the real-time collected lower limb movement inertial measurement signal.

[0171] In the embodiments of the present disclosure, the collected lower limb movement inertial measurement signal can be sampled by a sliding window using the steps in the model training method mentioned previously, the corresponding kinematic feature sequence is obtained, and the kinematic feature sequence is processed to obtain the movement feature. This will not be analyzed further, and the specific content can be referred to the analysis previously.

[0172] S320: According to the predetermined order of the lower limb motion state binary classification model, the motion features are processed by the corresponding binary classification model of each lower limb motion state in turn to obtain the corresponding binary classification result, wherein the binary classification model corresponding to the latter lower limb motion state processes the motion features and the former binary classification result to obtain the corresponding latter binary classification result, and the former binary classification result is the classification result obtained by processing the motion features by the former lower limb motion state binary classification model.

[0173] After obtaining the motion features, the computing device determines the order of the lower limb motion state binary classification model according to the foregoing, and processes the motion features by the corresponding binary classification model in turn to obtain all binary classification results. As in the operation in the model training process, the binary classification result obtained by processing by the former binary classification model and the motion features are taken as the input of the latter binary classification model, until the binary classification results of all binary classification models are obtained.

[0174] S330: The binary classification results corresponding to all lower limb motion states are input into the result decision model to obtain the recognized lower limb motion state.

[0175] After obtaining the binary classification results output by all binary classification models, the computing device then inputs all binary classification results into the result decision model to obtain the recognized lower limb motion state.

[0176] In some embodiments, for each lower limb motion binary classification model, the computing device does not use all motion features as model input features, but uses the selected motion features used in the foregoing model training method as model input to obtain the corresponding binary classification result.

[0177] In some embodiments, in the case where the lower limb motion state is a continuous motion state, the computing device will also use the historical binary classification result obtained by temporarily processing the former motion feature by the foregoing model training method, and input the historical binary classification result into the latter binary classification model to obtain the latter binary classification result.

[0178] The embodiments of the present disclosure also provide a computing device for implementing the foregoing method. Figure 6 is a structural schematic diagram of the computing device provided by the embodiments of the present disclosure. The following will be specifically referred to Figure 6 which shows a structural schematic diagram suitable for implementing the computing device 600 in the embodiments of the present disclosure. Figure 6 The computing device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present disclosure.

[0179] As Figure 6As shown, computing device 600 can include a processing device (e.g., central processing unit, graphics processing unit, etc.) 601 that can perform various appropriate actions and processes in accordance with programs stored in read only memory (ROM) 602 or loaded into random access memory (RAM) 603 from storage device 608. Various programs and data required by computing device 600 for operation are also stored in RAM 603. Processing device 601, ROM 602, and RAM 603 are connected to each other by bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0180] Generally, the following devices can be connected to I / O interface 605: input devices 606, including, for example, a touchscreen, a touchpad, a camera, a microphone, etc.; output devices 607, including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 608, including, for example, a magnetic tape, a hard disk, etc.; and communication devices 609. Communication devices 609 can allow computing device 600 to communicate with other devices wirelessly or via a wire to exchange data. Although Figure 6 Computing device 600 is shown with various devices, but it is understood that not all of the shown devices are required to be implemented or present. More or fewer devices can alternatively be implemented or present.

[0181] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication devices 609, or installed from storage devices 608, or installed from ROM 602. When the computer program is executed by processing device 601, the above-mentioned functions defined in the methods of embodiments of the present disclosure are performed.

[0182] It should be noted that the computer-readable medium described above in the present disclosure can be a computer-readable storage medium, a computer-readable signal medium, or, or any combination of the two.

[0183] A computer readable storage medium can be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a computer readable storage medium can include, but are not limited to, the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

[0184] A computer readable signal medium can include a propagated data signal with computer readable program code embodied therein, for use by or in connection with an instruction execution system, apparatus, or device. Examples of a computer readable signal medium include but are not limited to electronic, magnetic, optical, electromagnetic, infrared, or semiconductor

[0185] In some embodiments, clients, servers, and other components can communicate over a network using any known or future developed network protocol, including HTTP (Hypertext Transfer Protocol), and can be interconnected using any form or medium of digital data communication (for example, communication networks) including a local area network ("LAN"), a wide area network ("WAN"), the Internet, and peer-to-peer networks (for example, ad hoc peer-to-peer networks), as well as any currently known or future developed networks.

[0186] The computer readable medium described above can be included within the computing device; alternatively, the computer readable medium can exist as a separate entity with the computing device.

[0187] Computer program code for carrying out operations of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the testee's computer, partly on the testee's computer, as a stand-alone software package, partly on the testee's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the testee's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0188] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of present disclosure. In this regard, each block in the flow diagrams or block diagrams can represent a module, a procedure, or a part of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or in the reverse order, depending on the functionality involved. It is also noted that each block in the block diagrams and / or flow diagrams and combinations of blocks in the block diagrams and / or flow diagrams can be implemented by special-purpose hardware-based systems that perform the specified functions or operations, or combinations of special-purpose hardware and computer instructions.

[0189] The units described in the embodiments of the present disclosure can be implemented by software, or by hardware. In some cases, the names of the units do not constitute a limitation on the units themselves. The functions described above can be performed at least in part by one or more hardware logic components. For example, non-limiting examples of hardware logic components that can be used include field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SOCs), complex programmable logic devices (CPLDs), etc.

[0190] The foregoing is merely illustrative of the various ways and specific embodiments in which the disclosure can be carried out. Numerous modifications can be made to these specific embodiments and implementations without departing from the spirit and scope of the disclosure. It is, therefore, intended that this disclosure not be limited to the particular implementation described exclusively hereinabove but that all such alterations and modifications that come within the scope of the present disclosure are covered by the appended claims.

Claims

1. A method for training a lower limb motion state recognition model, characterized in that, The lower limb movement state recognition model includes a binary classification model and an outcome decision model corresponding to various lower limb movement states; the method includes... Based on the inertial measurement signals collected during various lower limb movement states, the corresponding motion features are determined respectively. The lower limb movement states are used as labels for the sample data, and the corresponding motion features are combined to form sample data. The sample data is then used to construct a sample dataset. The ranking order of binary classification models corresponding to various lower limb movement states is determined, and the binary classification models corresponding to each lower limb movement state are trained sequentially according to the ranking order as follows (S1-S2) until the model parameters of the binary classification models corresponding to all lower limb movement states are obtained and the binary classification results are obtained by processing the sample data respectively: S1: Obtain the binary classification result obtained by processing the motion features of the first-ranked lower limb motion state; S2: Use the motion features in the sample data corresponding to the current target lower limb movement state and the corresponding previous binary classification results as the input features of positive samples, and use the motion features in the sample data not corresponding to the current target lower limb movement state and the corresponding previous binary classification results as the input features of negative samples. Train the binary classification model corresponding to the current target lower limb movement state to obtain the model parameters of the binary classification model corresponding to the current target lower limb movement state and the current classification results of each sample data. The step of determining the ranking order of the binary classification models corresponding to various lower limb movement states includes: for sample data of various lower limb movement states, determining the correlation score between each movement feature and the label, and selecting a preset number of relevant features with high correlation to the label of the sample data based on the correlation score; for sample data of various lower limb movement states, performing preliminary regression analysis based on the corresponding selected relevant features and labels, and determining the regression analysis accuracy result corresponding to each movement state. The order of the regression analysis accuracy results from best to worst is used as the sorting order of the binary classification models corresponding to various lower limb movement states; The result decision model is trained based on the binary classification results obtained by processing the sample data according to the binary classification model corresponding to all lower limb movement states and the lower limb movement state labels of the sample data, and the model parameters of the result decision model are determined.

2. The training method of claim 1, wherein, The motion characteristics are determined based on the inertial measurement signals collected during various lower limb movement states, including: According to the preset sliding time window and sliding window frequency, the inertial measurement signals collected during various lower limb movement states are sampled by sliding window to obtain the sorted kinematic feature sequence within the sliding time window and the sorting identifier of the kinematic feature sequence. The motion features are determined based on the kinematic feature sequence, and the sorting identifier of the kinematic feature sequence is used as the sorting identifier of the sample data; When the lower limb movement state that is ranked first is a continuous state, the method further includes: obtaining the previous historical binary classification result obtained by the binary classification model processing the previous sample data according to the ranking identifier of each sample data; The method further comprises: The method further comprises:

3. Training method according to any of claims 1-2, characterized in that, The method further comprises: The method further comprises: a plurality of target samples are randomly selected from the sample data set, and a nearest N positive sample in one motion feature dimension and each of the target samples is determined from the positive samples, and a nearest N negative sample in the one motion feature dimension and each of the target samples is determined from the negative samples; calculating a first average distance of the target sample and the nearest N positive neighboring samples in the one motion feature dimension, and calculating a second average distance of the target sample and the nearest N negative neighboring samples in the one motion feature dimension, calculating a distance difference of the first average distance and the second average distance, and determining a second correlation score of the one motion feature according to the distance difference; The method further comprises: The method further comprises: The method further comprises: The method further comprises:

4. The training method of claim 3, wherein, The method further comprises: The method further comprises: The method further comprises: The method further comprises:

5. The training method of claim 3, wherein, The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further comprises: The method further correlation analysis is performed on the selected relevant motion features to determine the target motion feature. The target motion feature is determined from the irrelevant motion features and the selected relevant motion features according to a comprehensive score of the irrelevant motion features and the selected relevant motion features.

6. The training method of claim 3, wherein, The first correlation score of each motion feature with the current target lower limb motion state is determined by performing correlation analysis on the labels of the sample data and one motion feature based on the machine learning model, including: The feature coefficient corresponding to the one motion feature in the logistic regression model is determined by performing correlation analysis on the labels of the sample data and the one motion feature based on the logistic regression model, and the absolute value of the feature coefficient is taken as the first correlation score of the one motion feature with the current target lower limb motion state; or, The standard deviation under the current target lower limb motion state and the non-current target lower limb motion state is determined by performing correlation analysis on the labels of the sample data and the one motion feature based on the Gaussian Naive Bayes model. The average value of the standard deviations of the one motion feature under the current lower limb state and the non-current lower limb state is taken as the first correlation score of the one motion feature with the current target lower limb motion state.

7. The training method according to any one of claims 1 to 2, characterized in that, The model parameters of the binary classification model corresponding to the current target lower limb motion state and the current classification result of each sample data are obtained by training the binary classification model corresponding to the current target lower limb motion state, including training various types of binary classification models corresponding to the current target lower limb motion state to obtain the model parameters of the various types of binary classification models and determine the model quality score of the various types of binary classification models. The binary classification model with the optimal model quality score is selected as the selected binary classification model corresponding to the current target lower limb motion state, and the current classification result corresponding to the selected binary classification model is obtained.

8. A method of recognizing a motion state of a lower limb, characterized by, It includes: determining the corresponding motion features based on the real-time collected lower limb motion inertial measurement signals; processing the motion features by using the binary classification models corresponding to each lower limb motion state in turn according to the predetermined sorting order of the lower limb motion state binary classification models to obtain the corresponding binary classification results, wherein the binary classification model corresponding to the subsequent lower limb motion state processes the motion features and the previous binary classification result to obtain the corresponding subsequent binary classification result, and the previous binary classification result is the classification result obtained by processing the motion features by using the previous lower limb motion state binary classification model; inputting all the binary classification results corresponding to the lower limb motion states into the result decision model to obtain the recognized lower limb motion state. The training method of any one of claims 1-7 is used to obtain each lower limb motion state binary classification model and the result decision model.

9. A computing device, comprising: The lower limb motion state recognition model training method of any one of claims 1-7 and / or the lower limb motion state recognition method of claim 8 is executed by the processor when the computer program is loaded into the processor. The lower limb motion state recognition model training method of any one of claims 1-7 and / or the lower limb motion state recognition method of claim 8 is executed by the processor when the computer program is loaded into the processor.

Images