Cross-individual gait recognition method and system based on wearable inertial sensor

Abstract

The invention discloses a cross-individual gait recognition method and system based on a wearable inertial sensor. The system comprises the wearable inertial sensor, a signal processing module, a storage module and a man-machine interaction module. Wherein the wearable inertial sensor is arranged at the lower limb foot part of a user and is used for acquiring IMU data; the signal processing module is used for preprocessing the received IMU data, and obtaining to-be-identified IMU data after time window division of a sliding step length; according to a working mode selected by a user, constructing a data set by using the IMU data so as to train a gait recognition model, or inputting the IMU data into the trained gait recognition model, and outputting an action mode / type recognition result; the storage module is used for storing historical IMU data and action mode / type recognition results; and the man-machine interaction module is used for man-machine interaction of a user, selection of a system working mode and display and playback of an action mode / type recognition result.

Description

Technical Field

[0001] The present invention relates to a lower limb gait recognition method, and in particular to a cross-individual gait recognition method and system based on wearable inertial sensors. Background Art

[0002] An exoskeleton robot is a wearable robot that serves people. Its core function is to correctly understand people's autonomous movement intentions and provide appropriate assistance to complete daily movements and tasks.

[0003] Currently, the types of signal sources used to understand a person's voluntary movement intention are categorized into two main categories: typical physical signals, primarily those containing information such as force and posture, and bioelectrical signals such as electroencephalogram (EEG), electromyography (EMG), and electrocardiogram (ECG) generated by human movement. Decoding human movement intention using bioelectrical signals, such as EEG and EMG, has been a hot topic in recent years, but its widespread application has been hindered by technical difficulties and high costs. Force and posture signals acquired by force sensors and inertial sensors, on the other hand, are a mature and cost-effective technology, and are currently a widely used method for analyzing human movement intention.

[0004] Patent application publication number CN119523468A uses improved MCD transfer learning to achieve cross-individual gait recognition of electromyographic signals; however, this method still has shortcomings such as a large number of sensors, data redundancy, and poor cross-individual adaptability, which affect the timeliness and universality of exoskeleton interactive control. Summary of the Invention

[0005] The purpose of the present invention is to provide a cross-individual gait recognition method and system based on wearable inertial sensors, so as to solve the problems of large number of sensors and poor cross-individual adaptability when currently using wearable inertial sensors to analyze the gait of human lower limbs.

[0006] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0007] Cross-individual gait recognition method based on wearable inertial sensors, including:

[0008] Wearable inertial sensors are used to obtain IMU data from users during daily activities or work. After preprocessing the IMU data, samples are constructed and a dataset is established. The dataset is used to train a gait recognition model, which is then saved and used to recognize real-time IMU data.

[0009] Among them, the gait recognition model is improved based on the InceptionTime network, specifically:

[0010] Improve the input features of the InceptionTime network, use a sliding window to calculate the local statistics of each sample, and concatenate the local statistics and samples in parallel to form a new feature matrix as the input of the InceptionTime network;

[0011] Adjust the gated weight generator of the first Inception module of the InceptionTime network, linearly weight the features output by different convolutional branches of the first Inception module according to the weights, enhance channel features with the help of compression-excitation operation, and use gated linear units to achieve nonlinear feature fusion;

[0012] The second Inception module of the InceptionTime network is adjusted, and the output features of the first Inception module are subjected to secondary nonlinear fusion according to the weights. The fused features enter the subsequent pooling layer and fully connected layer of the InceptionTime network for processing, and finally the classification prediction results are output.

[0013] Furthermore, when constructing the dataset, the pre-processed IMU data is intercepted using a time window with a certain sliding step size. The IMU data of each time window is treated as an independent sample and manually annotated with a corresponding label. The label is the user's action mode / type under the corresponding IMU data, and the action mode / type should be determined based on the foot pressure.

[0014] When labeling, for a sample within a certain time window, if the IMU data in the sample does not fall into the judgment threshold interval of a certain action mode / type, that is, it is transition state data, the label of the sample is divided into the label of the next sample.

[0015] Furthermore, a sliding window is used to calculate the local statistics in each sample, and the local statistics and samples are concatenated in parallel to form a new feature matrix as the input of the InceptionTime network, including:

[0016] For each sample to be processed input to the InceptionTime network, a sliding window is used to calculate the local statistics of the signal in each sample. The local statistics of the signal include the local mean, variance and autocorrelation coefficient;

[0017] A series of signal local statistics are calculated for each sample, and these signal local statistics are concatenated with the sample to form a new input feature matrix as the input feature of the InceptionTime network.

[0018] Furthermore, the sliding window length L is adjusted according to the sampling frequency fs of the IMU data, and the value of L is

[0019] Furthermore, the processing of the gated weight generator of the first Inception module is as follows:

[0020] In the first step, the three convolution branches of the Inception module all output feature tensors F_i of dimension [B, T, C], where B is the batch size, T is the time step, and C is the number of channels. The feature tensors F_i are concatenated according to the channel dimension to form a new feature matrix F_c of dimension [B, T, 3C].

[0021] The second step is to calculate the new feature matrix F_c for the compression-excitation operation of the InceptionTime network. Specifically, the new feature matrix F_c is first transformed into a global quantity S compressed in the time dimension through a pooling operation, and then the global quantity S is stimulated into a channel weight vector W_c through the excitation of the fully connected layer;

[0022] The third step is to obtain the feature tensor F_i corresponding to the time attention branch in the three convolution branches, and calculate the time weight vector W_t of the importance index of each time point through convolution or self-attention mechanism;

[0023] The fourth step is weight fusion: first, the weight vector W_c is split into three weights w1, w2, and w3 according to the three convolution branches; then the time weight vector W_t is multiplied by the three weights w1, w2, and w3 to obtain the spatiotemporal joint weight W_j corresponding to each convolution branch; finally, the spatiotemporal joint weight W_j of each convolution branch is normalized and calculated in the softmax layer of each convolution branch, and finally the final weight matrix W of the three convolution branches is formed;

[0024] In the fifth step, first perform matrix multiplication on the feature tensor F_i and the final weight matrix W to obtain a weighted feature matrix F_w with the same dimension, and then input the weighted feature matrix F_w into the gated linear unit to obtain the final feature matrix F_f as the output feature of the first Inception module.

[0025] Furthermore, the processing of the second Inception module of the InceptionTime network is as follows:

[0026] The first step is to perform multi-scale convolution kernel calculation on the output feature F_f of the first Inception module to obtain the feature matrix F_N = (Fs, Fm, Fl) of different scales; where Fs, Fm and Fl are the feature matrices calculated by the three convolution branches of small, medium and large convolution kernels respectively;

[0027] In the second step, the feature matrix Fs representing the local statistical features in the feature matrix F_N is subjected to weight calculation and weight fusion in the first Inception module, and the final weight matrix obtained is W_D;

[0028] In the third step, the final weight matrix W_D is fused through gated nonlinearity to obtain a new weighted feature matrix as the output of the second Inception module.

[0029] A cross-individual gait recognition system based on a wearable inertial sensor comprises: a wearable inertial sensor, a signal processing module, a storage module and a human-computer interaction module; wherein:

[0030] Wearable inertial sensors are placed on the user's lower limbs and feet to collect IMU data;

[0031] The signal processing module is used to pre-process the received IMU data and divide it into sliding step time windows to obtain the IMU data to be identified. According to the working mode selected by the user, the IMU data is used to construct a data set to train the gait recognition model, or the IMU data is input into the trained gait recognition model to output the action mode / type recognition results.

[0032] The storage module is used to store historical IMU data and motion mode / type recognition results;

[0033] The human-computer interaction module is used for user human-computer interaction, selection of system working mode, display and playback of action mode / type recognition results.

[0034] Furthermore, the working process of the system includes:

[0035] Step 1: Place a single wearable inertial sensor on the user's lower limb foot to collect IMU data while the user is performing daily activities or work, and transmit the IMU data to the signal processing module;

[0036] Step 2: The signal processing module operates according to the working mode selected by the user through the human-computer interaction module, including model training mode and online recognition mode;

[0037] In the training mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into time windows according to the sliding step length to obtain samples. The user enters the labels corresponding to the samples through the human-computer interaction module, thereby constructing a data set and saving it in the storage module. The gait recognition model is trained using the continuously expanding data set.

[0038] In the online recognition mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into a sliding step time window to obtain the IMU data to be recognized; the IMU data is input into the gait recognition model, and the action mode / type recognition results are output through the human-computer interaction module.

[0039] Furthermore, when the wearable inertial sensor is worn, the X, Y, and Z axes of the wearable inertial sensor are aligned with the sagittal plane, coronal plane, and horizontal plane of the foot.

[0040] Furthermore, in the model training mode, the gait recognition model training parameters can be configured through the human-computer interaction module. At the same time, the gait recognition model can be tested by collecting new IMU data or using historical IMU data. When the test results meet the requirements, the gait recognition model can be saved; otherwise, it can be retrained.

[0041] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, the cross-individual gait recognition method based on wearable inertial sensors is implemented.

[0042] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the cross-individual gait recognition method based on wearable inertial sensors is implemented.

[0043] Compared with the prior art, the present invention has the following technical features:

[0044] The present invention addresses the problems of a large number of wearable inertial sensors, data redundancy, and poor cross-individual adaptability in the human motion perception of exoskeleton robots. A single wearable inertial sensor installed at the foot position is used to obtain its data. Through the specific construction of the data set and the construction, training, and storage of a deep learning neural network, online recognition of the time series signals of the wearable inertial sensor is achieved, and the relevant raw data, process data, and result data are displayed and stored, providing a low-cost and highly adaptable method and system for human-computer interaction of exoskeleton. First, the present invention can adapt to the gait characteristics of different users by having dual working modes of model training and online recognition, overcoming the problem of individual differences and expanding its adaptability and scope of use. Second, the use of a localized model training method eliminates dependence on the external and cloud, improving the privacy of use and the convenience of deployment. Finally, the small number of sensors used reduces the data dimension, and the recognition and reasoning process has high real-time performance, which makes the present invention practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a hardware structure diagram of the present invention;

[0046] Figure 2 It is a schematic flow diagram of the method of the present invention. DETAILED DESCRIPTION

[0047] The present invention provides a cross-individual gait recognition method based on wearable inertial sensors, comprising the following steps:

[0048] Wearable inertial sensors are used to obtain IMU data from users during daily activities or work. After preprocessing the IMU data, samples are constructed and a dataset is established. The dataset is used to train a gait recognition model, which is then saved and used to recognize real-time IMU data.

[0049] Among them, the gait recognition model is improved based on the InceptionTime network, specifically:

[0050] Improve the input features of the InceptionTime network, use a sliding window to calculate the local statistics of each sample, and concatenate the local statistics and samples in parallel to form a new feature matrix as the input of the InceptionTime network;

[0051] Adjust the gated weight generator of the first Inception module of the InceptionTime network, linearly weight the features output by different convolutional branches of the first Inception module according to the weights, enhance channel features with the help of compression-excitation operation, and use gated linear units to achieve nonlinear feature fusion;

[0052] The second Inception module of the InceptionTime network is adjusted, and the output features of the first Inception module are subjected to secondary nonlinear fusion according to the weights. The fused features enter the subsequent pooling layer and fully connected layer of the InceptionTime network for processing, and finally the classification prediction results are output.

[0053] The specific design process of the present invention is further described in detail below with reference to the accompanying drawings.

[0054] 1. Construction of dataset.

[0055] After acquiring the user's IMU data (quaternions, acceleration, angular velocity, etc.) through wearable inertial sensors, preprocessing such as filtering and normalization is performed. The preprocessed IMU data is intercepted using a time window with a certain sliding step size. The IMU data in each time window is treated as an independent sample and manually annotated with a corresponding label. The label represents the user's motion mode / type under the corresponding IMU data, and the motion mode / type should be determined based on the plantar pressure commonly used in the field of human kinematics research. However, when labeling, for a sample within a certain time window, if the IMU data in the sample does not fall within the judgment threshold interval of a certain motion mode / type, it is considered to be transitional state data. Traditional methods label the data proportion corresponding to the judgment threshold interval of two motion modes / types. However, this solution does not adopt the traditional method. Instead, the label of the sample is divided into the label of the next sample, that is, the label of the transition state data is considered to be the next motion mode / type. This method can effectively include the data change trend caused by human motion and eliminate the misleading effect of the "latency" of data changes on the model.

[0056] 2. Gait recognition model.

[0057] This paper proposes a gait recognition model, which is improved based on the InceptionTime network, as follows:

[0058] (1) Improvement of InceptionTime network input features.

[0059] In this scheme, for each sample to be processed input to the InceptionTime network, a sliding window is used to calculate the local statistics of the signal in each sample. The sliding window length L is adjusted according to the sampling frequency fs of the IMU data. L is usually taken as The local statistics of the signal include characteristics such as local mean, variance and autocorrelation coefficient, among which the local mean and variance are calculated based on the IMU data in the sliding window, and the autocorrelation coefficient is obtained by autocorrelating the IMU data in the current sliding window with the IMU data in the sliding window at the historical moment.

[0060] A series of signal local statistics are calculated for each sample, and these signal local statistics are concatenated with the sample to form a new input feature matrix as the input feature of the InceptionTime network.

[0061] The input features of the existing InceptionTime network are IMU data, with a dimension of n*m, where n is the dimension of the IMU data and m is the time dimension. The design of this solution can expand the dimension of the input features to (n+1)*m dimensions. This method can improve the saliency of feature expression for periodic time-series IMU data.

[0062] (2) Improvement of the gated weight generator of the first Inception module of the InceptionTime network.

[0063] The features output by the different convolutional branches of the first Inception module are linearly weighted according to their weights, channel features are enhanced using compression-excitation operations, and nonlinear feature fusion is achieved using gated linear units. The different convolutional branches include a temporal attention branch and other dimensional transformation branches; the temporal attention branch is used to transform the time dimension of the input feature matrix, while other dimensional transformation branches (such as the channel attention branch) are used to compress or transform the other dimensions of the input feature matrix.

[0064] This scheme focuses on enhancing the importance of calculating each time point in the temporal attention branch, while the channel attention branch needs to calculate the significance of each channel and improve the feature significance of the data by adjusting the weights.

[0065] Among them, the improvements to the gated weight generator of the first Inception module are as follows:

[0066] In the first step, the three convolution branches of the Inception module all output feature tensors F_i of dimension [B, T, C], where B is the batch size, T is the time step, and C is the number of channels; the feature tensors F_i are concatenated according to the channel dimension to form a new feature matrix F_c of dimension [B, T, 3C].

[0067] The second step is to calculate the new feature matrix F_c for the compression-excitation operation of the InceptionTime network. Specifically, the new feature matrix F_c is first transformed into a global quantity S compressed in the time dimension through a pooling operation, and then the global quantity S is excited into the channel weight vector W_c through the excitation of the fully connected layer.

[0068] The third step is to obtain the feature tensor F_i corresponding to the time attention branch in the three convolution branches, and calculate the time weight vector W_t of the importance index of each time point through convolution or self-attention mechanism.

[0069] The fourth step is to perform weight fusion: first, split the weight vector W_c into three weights w1, w2, and w3 according to the three convolution branches; then multiply the time weight vector W_t with the three weights w1, w2, and w3 to obtain the spatiotemporal joint weight W_j corresponding to each convolution branch; finally, the spatiotemporal joint weight W_j of each convolution branch is normalized and calculated at the softmax layer of each convolution branch, and finally the final weight matrix W of the three convolution branches is formed.

[0070] In the fifth step, first perform matrix multiplication on the feature tensor F_i and the final weight matrix W to obtain a weighted feature matrix F_w with the same dimension, and then input the weighted feature matrix F_w into the gated linear unit to obtain the final feature matrix F_f as the output feature of the first Inception module.

[0071] The feature matrix F_f can enhance the prominent role of signal features in the entire model calculation process and balance the contribution of features in different receptive fields.

[0072] (3) Adjustment of the second Inception module of the InceptionTime network.

[0073] The second Inception module performs a secondary nonlinear fusion of the output features of the first Inception module according to the weights. The fused features enter the subsequent pooling layer and fully connected layer of the InceptionTime network for processing, and finally output the classification prediction results.

[0074] In the quadratic nonlinear fusion, for the three convolution branches of the Inception module, the convolution branch with a small convolution kernel extracts local details of the features, the convolution branch with a medium convolution kernel captures the medium-range features of the features, and the convolution branch with a large convolution kernel obtains the long-term pattern of the features. By combining convolution features of different scales, the generalization ability of time series signals with significant periodic characteristics can be further improved.

[0075] Among them, the small convolution kernel, medium convolution kernel, and large convolution kernel can be customized, for example, they are 3*3, 5*5, and 7*7 convolution kernels respectively.

[0076] The processing of the second Inception module is as follows:

[0077] In the first step, multi-scale convolution kernel calculation is performed on the output feature F_f of the first Inception module to obtain feature matrices of different scales F_N = (Fs, Fm, Fl); among them, Fs, Fm and Fl are feature matrices calculated by three convolution branches of small, medium and large convolution kernels respectively. In this step, the channel data representing local statistical features in the feature matrix F_N will be enhanced through small-scale convolution calculation.

[0078] In the second step, the feature matrix Fs representing the local statistical features in the feature matrix F_N is subjected to weight calculation and weight fusion in the third and fourth steps of the first Inception module, and the final weight matrix obtained is W_D.

[0079] In the third step, the final weight matrix W_D is fused through gated nonlinearity to obtain a new weighted feature matrix as the output of the second Inception module.

[0080] In this step, channels with significant statistical characteristics will be given higher weights, and the weights of insignificant channels will be reduced, so as to enhance the features of high-response channels or suppress low-response channels, thereby improving the contribution of features in the model or suppressing noise interference.

[0081] The output features of the second Inception module enter the subsequent pooling layer and fully connected layer of the InceptionTime network for processing, and finally output the classification prediction result, that is, the prediction result of the gait movement mode / type.

[0082] Based on the above technical solutions, the present invention further provides a cross-individual gait recognition system based on wearable inertial sensors. This system uses data from a single wearable inertial sensor on the foot of the lower limb to identify the movement pattern / type of the lower limb of the human body. With a small number of wearable inertial sensors and good cross-individual adaptability, it can realize the recognition of the movement pattern of the lower limb of different users. The system includes: a wearable inertial sensor, a signal processing module, a storage module, and a human-computer interaction module. The wearable inertial sensor, the storage module, and the human-computer interaction module all have information exchange with the signal processing module, wherein:

[0083] Wearable inertial sensors are placed on the user's lower limbs and feet to collect IMU data;

[0084] The signal processing module is a micro-processing module composed of X86 architecture chips. It pre-processes the received IMU data and divides it into sliding step time windows to obtain the IMU data to be recognized. According to the working mode selected by the user, it uses the IMU data to construct a data set to train the gait recognition model, or inputs the IMU data into the trained gait recognition model to output the movement mode / type recognition results.

[0085] The storage module is used to store historical IMU data and motion mode / type recognition results;

[0086] The human-computer interaction module is used for user human-computer interaction, selection of system working mode, display and playback of action mode / type recognition results.

[0087] The above system works as follows:

[0088] Step 1: Place a single wearable inertial sensor on the user's lower limb foot to collect IMU data while the user is performing daily activities or work, and transmit the IMU data to the signal processing module;

[0089] Step 2: The signal processing module operates according to the working mode selected by the user through the human-computer interaction module, including model training mode and online recognition mode;

[0090] In training mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into time windows according to the sliding step length to obtain samples. The user inputs the labels corresponding to the samples through the human-computer interaction module to construct a data set and save it in the storage module; the gait recognition model is trained using the continuously expanding data set to improve the recognition accuracy of the gait recognition model.

[0091] In the online recognition mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into a sliding step time window to obtain the IMU data to be recognized; the IMU data is input into the gait recognition model, and the action mode / type recognition results are output through the human-computer interaction module.

[0092] Among them, when the wearable inertial sensor is worn, the X, Y, and Z axes of the wearable inertial sensor are consistent with the sagittal plane, coronal plane, and horizontal plane of the foot, and the corresponding order is based on the convenience of installation and use.

[0093] In the model training mode, the gait recognition model training parameters can be configured through the human-computer interaction module. At the same time, the gait recognition model can be tested by collecting new IMU data or using historical IMU data. When the test results meet the requirements, the gait recognition model can be saved; otherwise, it can be retrained.

[0094] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A cross-individual gait recognition method based on wearable inertial sensors, characterized in that: include: Use wearable inertial sensors to obtain IMU data from users during daily activities or work, pre-process the IMU data, construct samples, and establish a dataset; Use the data set to train the gait recognition model, and save the trained gait recognition model for recognition of real-time collected IMU data; Among them, the gait recognition model is improved based on the InceptionTime network, specifically: Improve the input features of the InceptionTime network, use a sliding window to calculate the local statistics of each sample, and concatenate the local statistics and samples in parallel to form a new feature matrix as the input of the InceptionTime network; Adjust the gated weight generator of the first Inception module of the InceptionTime network, linearly weight the features output by different convolutional branches of the first Inception module according to the weights, enhance channel features with the help of compression-excitation operation, and use gated linear units to achieve nonlinear feature fusion; The second Inception module of the InceptionTime network is adjusted, and the output features of the first Inception module are subjected to secondary nonlinear fusion according to the weights. The fused features enter the subsequent pooling layer and fully connected layer of the InceptionTime network for processing, and finally the classification prediction results are output.

2. The cross-individual gait recognition method based on wearable inertial sensors according to claim 1, characterized in that: When constructing the dataset, the pre-processed IMU data is intercepted using a time window with a certain sliding step size. The IMU data of each time window is treated as an independent sample and manually labeled with a corresponding label. The label is the user's action mode / type under the corresponding IMU data, and the action mode / type should be determined based on the foot pressure. When labeling, for a sample within a certain time window, if the IMU data in the sample does not fall into the judgment threshold interval of a certain action mode / type, that is, it is transition state data, the label of the sample is divided into the label of the next sample.

3. The cross-individual gait recognition method based on wearable inertial sensors according to claim 1, characterized in that: The sliding window is used to calculate the local statistics in each sample, and the local statistics and samples are concatenated in parallel to form a new feature matrix as the input of the InceptionTime network, including: For each sample to be processed input to the InceptionTime network, a sliding window is used to calculate the local statistics of the signal in each sample. The local statistics of the signal include the local mean, variance and autocorrelation coefficient; A series of signal local statistics are calculated for each sample, and these signal local statistics are concatenated with the sample to form a new input feature matrix as the input feature of the InceptionTime network.

4. The cross-individual gait recognition method based on wearable inertial sensors according to claim 3 is characterized in that: The sliding window length L is adjusted according to the sampling frequency fs of the IMU data, and the value of L is 2 log2(fs / 10) .

5. The cross-individual gait recognition method based on wearable inertial sensors according to claim 1, characterized in that: The processing of the gated weight generator of the first Inception module is as follows: In the first step, the three convolution branches of the Inception module all output feature tensors F_i of dimension [B, T, C], where B is the batch size, T is the time step, and C is the number of channels. The feature tensors F_i are concatenated according to the channel dimension to form a new feature matrix F_c of dimension [B, T, 3C]. The second step is to calculate the new feature matrix F_c for the compression-excitation operation of the InceptionTime network. Specifically, the new feature matrix F_c is first transformed into a global quantity S compressed in the time dimension through a pooling operation, and then the global quantity S is stimulated into a channel weight vector W_c through the excitation of the fully connected layer; The third step is to obtain the feature tensor F_i corresponding to the time attention branch in the three convolution branches, and calculate the time weight vector W_t of the importance index of each time point through convolution or self-attention mechanism; The fourth step is weight fusion: first, the weight vector W_c is split into three weights w1, w2, and w3 according to the three convolution branches; then the time weight vector W_t is multiplied by the three weights w1, w2, and w3 to obtain the spatiotemporal joint weight W_j corresponding to each convolution branch; finally, the spatiotemporal joint weight W_j of each convolution branch is normalized and calculated in the softmax layer of each convolution branch, and finally the final weight matrix W of the three convolution branches is formed; In the fifth step, first perform matrix multiplication on the feature tensor F_i and the final weight matrix W to obtain a weighted feature matrix F_w with the same dimension, and then input the weighted feature matrix F_w into the gated linear unit to obtain the final feature matrix F_f as the output feature of the first Inception module.

6. The cross-individual gait recognition method based on wearable inertial sensors according to claim 1, characterized in that: The processing process of the second Inception module of the InceptionTime network is as follows: The first step is to perform multi-scale convolution kernel calculation on the output feature F_f of the first Inception module to obtain the feature matrix F_N = (Fs, Fm, Fl) of different scales; where Fs, Fm and Fl are the feature matrices calculated by the three convolution branches of small, medium and large convolution kernels respectively; In the second step, the feature matrix Fs representing the local statistical features in the feature matrix F_N is subjected to weight calculation and weight fusion in the first Inception module, and the final weight matrix obtained is W_D; In the third step, the final weight matrix W_D is fused through gated nonlinearity to obtain a new weighted feature matrix as the output of the second Inception module.

7. A cross-individual gait recognition system based on wearable inertial sensors, characterized in that: include: Wearable inertial sensor, signal processing module, storage module and human-computer interaction module; among which: Wearable inertial sensors are placed on the user's lower limbs and feet to collect IMU data; The signal processing module is used to pre-process the received IMU data and divide it into sliding step time windows to obtain the IMU data to be identified. According to the working mode selected by the user, the IMU data is used to construct a data set to train the gait recognition model, or the IMU data is input into the trained gait recognition model to output the action mode / type recognition results. The storage module is used to store historical IMU data and motion mode / type recognition results; The human-computer interaction module is used for user human-computer interaction, selection of system working mode, display and playback of action mode / type recognition results.

8. The cross-individual gait recognition system based on wearable inertial sensors according to claim 7, characterized in that: The system's working process includes: Step 1: Place a single wearable inertial sensor on the user's lower limb foot to collect IMU data while the user is performing daily activities or work, and transmit the IMU data to the signal processing module; Step 2: The signal processing module operates according to the working mode selected by the user through the human-computer interaction module, including model training mode and online recognition mode; In the training mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into time windows according to the sliding step length to obtain samples. The user enters the labels corresponding to the samples through the human-computer interaction module, thereby constructing a data set and saving it in the storage module. The gait recognition model is trained using the continuously expanding data set. In the online recognition mode, the IMU data acquired in real time by the signal processing module is preprocessed and divided into a sliding step time window to obtain the IMU data to be recognized; the IMU data is input into the gait recognition model, and the action mode / type recognition results are output through the human-computer interaction module.

9. The cross-individual gait recognition system based on wearable inertial sensors according to claim 7, characterized in that: When the wearable inertial sensor is worn, the X, Y, and Z axes of the wearable inertial sensor are aligned with the sagittal plane, coronal plane, and horizontal plane of the foot.

10. The cross-individual gait recognition system based on wearable inertial sensors according to claim 7, characterized in that: In the model training mode, the training parameters of the gait recognition model can be configured through the human-computer interaction module. At the same time, the gait recognition model can be tested by collecting new IMU data or using historical IMU data. When the test results meet the requirements, the gait recognition model can be saved; otherwise, it can be retrained.

Images