Wearable exoskeleton wearer behavior recognition method and system based on residual network

Abstract

The present application belongs to the technical field of wearable external limbs, and provides a wearable external limb wearer behavior recognition method and system based on a residual network, in which, in the residual network model, first, shallow feature extraction is performed on posture information; then, two residual blocks are used to extract deep features of the posture information; in each training iteration process, part of neurons are randomly inactivated; finally, through a full connection layer and an activation function, a numerical vector is normalized into a probability distribution vector, and a final action category is obtained; in the present application, shallow feature extraction is performed on the posture information, two residual blocks are used to extract deep features of the posture information, and in each training iteration process, part of neurons are randomly inactivated, so that the overfitting problem is prevented, the residual network model including the two residual blocks is designed, and the gradient vanishing and deep network degradation problems are avoided to a certain extent.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wearable exolimb, and particularly relates to a wearable exolimb wearer behavior recognition method and system based on a residual network. BACKGROUND

[0002] In the industrial fields such as the building decoration industry, aircraft production and manufacturing, and nuclear power operation and maintenance, due to the limitations of working space, operation scale, and special environment, large robots and mechanical equipment are difficult to enter and play a role, and can only rely on single-person operation to complete tasks. The wearable exolimb can enhance the single-person operation capability in the above limited environment and complex working conditions through the docking and fusion of the mechanical limb and the wearer's limb and the cooperation and mutual assistance. Identifying the behavior of the wearer helps the exolimb robot to better complete the auxiliary operation. In order to enable the exolimb robot to accurately and quickly identify the wearer's intention, thereby better completing the auxiliary operation function, the industry generally uses an inertial measurement unit (IMU) to collect three-dimensional attitude angle information of the wearer, performs data processing and feature extraction, and uses a deep learning model for classification and recognition.

[0003] The inventor found that in the behavior recognition process based on the IMU, when the neural network of the deep learning model reaches a certain depth, with the increase of the number of neural network layers, the accuracy of the neural network will saturate, and even rapidly decrease. Even if batch normalization is used, the training error will not decrease but increase. Once the amount of input data increases, the neural network will also have a network degradation phenomenon to a certain extent, thereby causing the loss of part of the features of the input information. SUMMARY

[0004] To solve the above problems, the present application provides a wearable exolimb wearer behavior recognition method and system based on a residual network. Through the design of the residual network, the present application solves the problems such as network degradation of the neural network and loss of part of the features of the input information in the behavior recognition process based on the IMU.

[0005] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0006] In a first aspect, the present application provides a wearable exolimb wearer behavior recognition method based on a residual network, comprising:

[0007] Obtaining attitude information of multiple parts of the wearer;

[0008] Identifying the wearer according to the attitude information and a preset residual network model;

[0009] In the residual network model, first, shallow feature extraction is performed on the posture information; then, two residual blocks are used to extract deep features of the posture information; in each training iteration process, part of the neurons are randomly inactivated; finally, through the full connection layer and the activation function, the numerical vector is normalized into a probability distribution vector, and the final action category is obtained.

[0010] Further, the posture information at least includes one or more of the posture angle and the acceleration.

[0011] Further, the spikes and burrs in the curve corresponding to the posture information are removed, and the curve corresponding to the collected posture information is smoothed.

[0012] Further, the posture information is cut into a plurality of windows of the same length without interval; the length of the plurality of windows and the window overlap rate are fixed.

[0013] Further, the feature data of the posture information is mapped to a distribution with a mean of 0 and a standard deviation of 1.

[0014] Further, the residual network model includes a gated recurrent unit, two residual blocks and a random inactivation module connected in series.

[0015] Further, the residual block includes two residual sub-modules connected in series;

[0016] In the residual sub-module, a gated recurrent unit module is used to extract time series features; the input data is mapped to the data line by the cross-layer identity transformation and is propagated forward.

[0017] In a second aspect, the present application further provides a wearable exolimb wearer behavior recognition system based on a residual network, comprising:

[0018] The data acquisition module is configured to acquire posture information of multiple parts of the wearer.

[0019] The behavior recognition module is configured to recognize the wearer according to the posture information and a preset residual network model.

[0020] In the residual network model, first, shallow feature extraction is performed on the posture information; then, two residual blocks are used to extract deep features of the posture information; in each training iteration process, part of the neurons are randomly inactivated; finally, through the full connection layer and the activation function, the numerical vector is normalized into a probability distribution vector, and the final action category is obtained.

[0021] In a third aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the wearable exolimb wearer behavior recognition method based on the residual network of the first aspect.

[0022] In a fourth aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method for recognizing the behavior of a wearer of an external limb based on a residual network according to the first aspect when executing the program.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] 1. In the present application, the pose information is subjected to shallow feature extraction, two residual blocks are used to extract deep features of the pose information, part of the neurons are randomly inactivated in each training iteration process, the problem of overfitting is prevented, the residual network model including two residual blocks is designed, the problems of gradient disappearance and deep network degradation are avoided to a certain extent, and the problems of network degradation of the neural network, loss of part of the features of the input information and the like in the behavior recognition process based on the IMU are solved.

[0025] 2. In the present application, the pose information is subjected to filtering, segmentation, normalization and the like, the influence of the factors such as sensor jitter, external noise interference and the like on the collection of the pose information is avoided, and the behavior recognition accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] The drawings constituting a part of the present embodiment are used to provide further understanding of the present embodiment, the illustrative embodiments of the present embodiment and the description thereof are used to explain the present embodiment, and do not constitute improper limitation on the present embodiment.

[0027] Figure 1 It is a flow chart of the behavior recognition process based on the IMU of the present embodiment 1.

[0028] Figure 2 It is a wearing mode and sensor point position map of the present embodiment 1.

[0029] Figure 3 It is a raw data curve graph collected by the present embodiment 1.

[0030] Figure 4 It is a filtered curve graph of the present embodiment 1.

[0031] Figure 5 It is a schematic diagram of the present embodiment 1 with a window overlap rate of 0.

[0032] Figure 6 It is a schematic diagram of the present embodiment 1 with a window overlap rate of λ.

[0033] Figure 7 It is a joint angle change curve graph after normalization of the present embodiment 1.

[0034] Figure 8 RNN structure schematic diagram for embodiment 1 of the present application;

[0035] Figure 9 RNN Cell calculation process diagram for embodiment 1 of the present application;

[0036] Figure 10 LSTM Cell schematic diagram for embodiment 1 of the present application;

[0037] Figure 11 GRU Cell schematic diagram for embodiment 1 of the present application;

[0038] Figure 12 Network direct connection schematic diagram for embodiment 1 of the present application;

[0039] Figure 13 Network connection with residual for embodiment 1 of the present application;

[0040] Figure 14 Res-GRU Block structure diagram for embodiment 1 of the present application;

[0041] Figure 15 Deep residual GRU network for embodiment 1 of the present application;

[0042] Figure 16 Loss curve change diagram in the training process for embodiment 1 of the present application;

[0043] Figure 17 Accuracy rate curve change diagram in the training process for embodiment 1 of the present application;

[0044] Figure 18 Confusion matrix visualization diagram for embodiment 1 of the present application. DETAILED DESCRIPTION

[0045] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0047] Embodiment 1:

[0048] The present embodiment provides a wearable exo-limb wearer behavior recognition method based on a residual network, comprising:

[0049] Obtaining posture information of multiple parts of the wearer;

[0050] According to the posture information and a preset residual network model, the wearer is recognized.

[0051] In the residual network model, first, the pose information is shallow feature extracted; then, two residual blocks are used to extract deep features of the pose information; in each training iteration process, part of the neurons are randomly inactivated; finally, through the full connection layer and the activation function, the numerical vector is normalized to the probability distribution vector, and the final action category is obtained. In each training iteration process, part of the neurons are randomly inactivated, which prevents overfitting problem, the residual network model including two residual blocks avoids the gradient disappearance and deep network degradation problem to some extent, and solves the problems of network degradation of neural network and loss of part of features of input information in the IMU-based behavior recognition process.

[0052] The wearable exolimb wearer behavior recognition method in the embodiment mainly includes:

[0053] S1, pose information data acquisition and processing:

[0054] S1.1, determining the data acquisition mode:

[0055] The signal acquisition of the inertial sensor is the most basic link in the research of the IMU-based human behavior recognition and is also the basis of the entire research. This step directly affects the quality of the human behavior data acquired by the sensor, and further affects the final classification effect. Optionally, there are two methods to acquire human behavior data information, one is to directly use the existing public data set, and the other is to build a corresponding data acquisition platform to acquire the behavior data set. Considering that if the public data set is directly used, it may not meet some specific behaviors during the wearer operation, therefore, the process of building a data acquisition platform to acquire the behavior data set is introduced in the embodiment, one embodiment is to use the Xsens motion capture system to acquire the wearer behavior information and establish the corresponding data set. The acquired pose information at least includes one or more of the pose angle and the acceleration.

[0056] Optionally, when collecting the pose information, sensors can be placed in 17 key parts of the whole body of the collection object (such as the head, shoulder, knee, etc.) to collect data, such as the sensor point diagram shown in FIG. 1. Figure 2

[0057] The captured human pose feature parameters and their roles in the exolimb robot operation strategy semi-physical simulation system are shown in Table 1:

[0058] Table 1: Table of captured pose feature parameters and functions

[0059]

[0060] ​

[0061] Through the software in the motion capture system, the characteristic parameters and action pictures listed in Table 1 can be viewed and recorded in real time, and data streams can be supported to be imported or exported into common 3D software packages to complete various collaborations. In addition, previously recorded files can also be played, analyzed and edited offline.

[0062] S1.2, posture information acquisition and data set making

[0063] In the exoskeleton robot operation strategy semi-physical simulation system, the wearer and the exoskeleton robot mainly complete operation tasks such as auxiliary carrying, support protection, support unloading, etc. In these task scenarios, mainly involved are operations such as standing, squatting, single-hand and double-hand lifting, etc. The embodiment will design experiments to collect these behavior information.

[0064] Determine a plurality of collection objects between 22-30 years old, 1.72-1.85 m in height and 62-75 kg in weight, each collection object wears an Xsens MVN suit and an MTw at the main joint point during collection, the sensor collection frequency can be 60 Hz, and a total of 6 kinds of human behavior data including standing, walking, squatting, left hand lifting, right hand lifting and double hand lifting are collected. The collection time of each behavior data of each collection object is not less than 4 minutes, and after collection, each behavior is added with a category label and saved for subsequent use.

[0065] S1.3, data preprocessing:

[0066] When using IMU to collect human action information, when wearing, due to factors such as sensor jitter, external noise interference, etc., the original signal collected by the sensor inevitably contains some useless noise information in addition to the desired human posture information. Therefore, after obtaining the data, before using the model classification, the data needs to be preprocessed to remove the noise components in the data, and at the same time, the data needs to be aligned, etc. to facilitate model classification.

[0067] S1.3.1, data filtering:

[0068] When collecting human posture signal data, due to the influence of sensor jitter, the collected serialized data curve may appear phenomena such as spikes or burrs, as shown in FIG. 1. Figure 3 Therefore, the embodiment uses an exponential weighted moving average (EWMA) filtering algorithm to clean up the data.

[0069] EWMA assigns a larger weight to data closer to the current time, and a smaller overall weight to data farther from the current time, and its formula is:

[0070]

[0071] From the above formula, it can be deduced that:

[0072] EMA t = β · x t + (1 - β) · EMA t-1

[0073] = β · x t + (1 - β) · (β · x t-1 + (1 - β) · EMA t-2 )

[0074] = β · x t + β · x t-1 - β 2 · x t-1 + (1 - β) 2 · EMA t-2

[0075] =...

[0076] wherein x0is the data value at t0time; β is a weighting coefficient; x t is the data value at t time.

[0077] It can be seen that the EWMA is a recursive function, and the recursive property causes the weight to decay exponentially. The data curve after using the EWMA filtering is much smoother than the curve in Figure 3 , as shown in Figure 4 .

[0078] S1.3.2, data segmentation:

[0079] Compared with the high sampling frequency (60Hz) of the sensor, human behavior activities are performed in a relatively long time period (in seconds or minutes); in addition, a single sample at a specific time (such as the elbow joint angular velocity of 0.6ms -1 ) cannot provide sufficient information to describe the performed behavior, and the time required for each action is also different. If the collected data is directly sent to the recognition model, it is not conducive to judge a single complete action.

[0080] Therefore, in order to facilitate classification, the embodiment uses a sliding window to segment the original data signal, and cuts the original signal into the same length without interval between windows, which can well guarantee the continuity of the data and is suitable for real-time recognition. The principle is shown in Figure 5 .

[0081] When using a sliding window to segment a dataset, two key factors are window length and window overlap rate. While a shorter window can improve feature extraction performance, it may not provide enough information to describe the complete activity; conversely, a longer window may allow for multiple activities to occur within a single time window. In this embodiment, the window length is set to 1 second during preprocessing.

[0082] When segmenting data, using a fixed window overlap ratio can more accurately handle time windows with switching actions. In this embodiment, the window overlap ratio is set to 0.5 when using code to process the data.

[0083] The dataset obtained by setting the above parameters is shown in Table 2:

[0084] Table 2 Sample Dataset Table

[0085]

[0086] S1.3.3, Data Normalization:

[0087] To eliminate the influence of different units between data features, this embodiment normalizes the feature data to make different indicators comparable.

[0088] Zero-mean normalization (Z-Score Normalization) maps raw data to a distribution with a mean of 0 and a standard deviation of 1. Specifically, assuming the mean of the original features is μ and the standard deviation is σ, the normalization formula is defined as:

[0089]

[0090] Where z represents the normalized feature data. Figure 4 The curve shown is normalized, and the result is as follows: Figure 7 As shown, before normalization, the numerical range of joint angles is large. By normalization, the overall data range can be reduced while preserving the original features, and the training speed of the model can be accelerated in the later stages.

[0091] S1.3.4, One-hot encoding:

[0092] One-hot encoding, also known as single-hot coding, is a one-bit valid encoding method. It uses an N-bit state register to encode N states, with each state having its own independent register bit, and only one bit is active at any given time. The six behavior state label codes after using one-hot encoding are shown in Table 3:

[0093] Table 3 shows the label status encoding for each row after one-hot encoding.

[0094]

[0095]

[0096] After one-hot encoding, the data becomes sparse, which solves the problem that classifiers have difficulty handling attribute data and can, to some extent, expand the features.

[0097] S2. Deep GRU network based on residual learning:

[0098] With the improvement of computing power and the development of deep learning technology, deep neural networks have been widely used in the field of time series classification (TSC) and have achieved excellent results. This embodiment will introduce different recurrent neural network (RNN) units and the residual learning concept, and construct a deep residual GRU network.

[0099] S2.1 Common RNN Units:

[0100] S2.1.1. RNN is a type of artificial neural network where the connections between nodes can create a loop, allowing the output of some nodes to serve as the input of subsequent nodes of the same type. Therefore, it can exhibit dynamic behavior over time. Its structural principle is as follows: Figure 8 As shown. RNNs originate from feedforward neural networks and can use their internal states to process input sequences of variable length.

[0101] For each RNN cell, the calculation formula is as follows:

[0102] h t =tanh(W ih x t +b ih +W hh h t-1 +b hh )

[0103] Where, x t h is the external input vector at time t; t-1 h t The hidden layer state matrices at times t-1 and t are respectively; W ih W hh These are the input weight matrix and the hidden layer weight matrix, respectively, while b ih and b hh The deviation vectors corresponding to the two are calculated as follows: Figure 9 As shown:

[0104] RNNs were first applied to language models because of their ability to remember long-term dependencies. However, as the time lag increases, the gradient of the network can vanish as the RNN unfolds into a very deep feed-forward neural network. To solve this problem, RNN structures with forget gates, such as LSTM and GRU, are proposed to enable the storage unit to forget part of the information at a certain time to determine the optimal time lag and thus slow down the gradient vanishing.

[0105] S2.1.2, Long Short Term Memory (LSTM) includes a forget gate, an input gate, and an output gate, and its structure is as shown in Figure 10 .

[0106] The input gate obtains new input points from the outside and processes incoming data; the forget gate decides when to forget the output result, thereby selecting the optimal time lag for the input sequence, avoiding long-term dependencies in the context state; the output gate obtains all the calculation results and generates the output for the LSTM unit. For each element in the input sequence, each layer is calculated according to the following function:

[0107]

[0108] where i t , f t and O t are the input, forget, and output gate states at time t, respectively, all of which use the sigmoid function σ as the activation function, and its expression is:

[0109]

[0110] where C t-1 and C t are the memory cells at time t-1 and t, respectively; is the candidate memory cell at time t, which can be fixed in the range [-1, 1] by using the tanh activation function.

[0111] S2.1.3, Gated Recurrent Unit (GRU), which can make each recurrent unit adaptively capture dependencies of different time scales. Similar to the LSTM unit, the GRU also has a gating unit that can regulate the flow of information within the unit, but does not have a separate storage unit.

[0112] As shown in Figure 11 , the GRU Cell contains a reset gate r t and an update gate z twhere the reset gate helps to capture short-term dependencies in the time series, while the update gate helps to capture long-term dependencies in the time series, and the calculation methods are shown in formulas 3-X:

[0113]

[0114] At the same time, GRU introduces candidate hidden state Auxiliary calculation of hidden state h t :

[0115]

[0116] where b ir , b hr , b hz , b in and b hn represent the bias vectors of different weight matrices.

[0117] Compared with LSTM, GRU reduces the number of gating signals and reduces the amount of training parameters, and can realize faster convergence speed under the condition of ensuring network accuracy.

[0118] S2.2, construct a deep residual GRU network:

[0119] When the network reaches a certain depth, with the increase of the number of neural network layers, the accuracy of the network will be saturated, and even a rapid decline phenomenon will appear, even if batch normalization is used, the phenomenon of training error rising instead of falling will appear. In order to solve this problem, the embodiment proposes a residual learning architecture to simplify the training of the network and solve the problem of difficult training of deep network.

[0120] Compared with network direct connection, in Figure 12 , the input can be transmitted faster through the data line across the layers; record the mapping to be fitted as H(x), x represents the input of the first layer, as shown in Figure 13 , then the residual learning makes the output mapping result of the stacked network of several layers f(x) = H(x)-x, so the original mapping is f(x)+x, which can be realized through the identity transformation mapping with "shortcut connections". If residual connection is used for each stacked layer, the following can be obtained:

[0121] y = f(x, {W i}) + x

[0122] wherein, x and y are the initial input and final output vectors of the network respectively; f(x, {W i}) represents the mapping output before adding residual connection.

[0123] Residual networks make information flow more easily between layers, including providing feature reuse when propagating forward and alleviating vanishing gradient signals when propagating backward, which makes it easier for the network to learn identity transformations at certain layers. Performing identity transformations at certain layers is a constructive solution that makes it possible for deeper networks to train better.

[0124] Deep residual GRU network:

[0125] The common GRU can solve the problem of gradient explosion to a large extent. However, once the amount of input data increases, the GRU will also have network degradation to a certain extent, resulting in the loss of part of the features of the input information. In the embodiment, the GRU is regarded as a calculation block, and residual information is passed around the block to speed up the convergence speed.

[0126] The GRU module is used in the residual block Res-GRU Block to extract time series features. For a single Res-GRU Block, the structure is as shown in Figure 14 wherein x and y are the input and output of the Block module respectively, and each Res-GRU Block contains two Res-GRU Unit modules. By introducing residual connection, the feature information flows more freely in the Block, allowing the gradient to pass through the addition operation between different layers, bypassing some layers that may have limitations, and making the backward propagation of deeper network layers more smooth.

[0127] Based on the target and residual learning idea of human action recognition, by using Pytorch, similar to building blocks, the modules are combined together in the embodiment to build a deep residual GRU network, and the specific structure is as shown in Figure 15 .

[0128] Specifically, in the deep residual GRU network, first, the GRU Cell is used for shallow feature extraction, and then two Res-GRU Block modules are used to extract deep features of the data. In order to prevent overfitting, a Dropout module is added to randomly inactivate part of the neurons in each training iteration process. Finally, through the full connection layer and the Softmax activation function, a numerical vector is normalized to a probability distribution vector to obtain the final action category. Due to the existence of residual layers and GRU layers, the designed network can avoid the problems of gradient vanishing and deep network degradation to a certain extent.

[0129] S3, experiment:

[0130] In some embodiments, in order to better train and evaluate the model, the data set is divided in this embodiment, 70% as the training set, 30% as the validation / test set, the cross entropy loss function (Cross Entropy Loss) and the Adam optimizer are selected, and the part of the hyperparameter values set in the training are as shown in Table 5:

[0131] Table 5 Training parameter setting table

[0132]

[0133] Figure 16 and Figure 17 The change process of loss and accuracy in the training process is shown. After 120 epoch training, the final loss of the model is 1.13, and the accuracy reaches 91.31%.

[0134] As Figure 18 shown, from the confusion matrix visualization diagram, it can be seen that the model has good classification effect on various behaviors. According to the values in the diagram, the final recall rate of the model is calculated to be 91.62%, the precision is 91.44%, and the F1Score is 91.31.

[0135] Embodiment 2:

[0136] The embodiment provides a wearable outer limb wearer behavior recognition system based on a residual network, comprising:

[0137] The data acquisition module is configured to acquire posture information of multiple parts of the wearer.

[0138] The behavior recognition module is configured to recognize the wearer according to the posture information and a preset residual network model.

[0139] In the residual network model, first, the posture information is extracted for shallow features; then two residual blocks are used to extract deep features of the posture information; in each training iteration process, part of the neurons are randomly inactivated; finally, through the full connection layer and the activation function, the numerical vector is normalized to the probability distribution vector, and the final action category is obtained.

[0140] The working method of the system is the same as the residual network-based wearable outer limb wearer behavior recognition method of embodiment 1, which will not be repeated here.

[0141] Embodiment 3:

[0142] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps of the residual network-based wearable outer limb wearer behavior recognition method of embodiment 1.

[0143] Embodiment 4:

[0144] The embodiment provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the steps of the residual network-based wearable exoskeleton wearer behavior recognition method in the embodiment 1 when executing the program.

[0145] The above merely provides preferred embodiments of the embodiment, but is not intended to limit the embodiment. For those skilled in the field, the embodiment can have various modifications and changes. Any modified, equivalent replaced, improved, etc. within the spirit and principle of the embodiment shall be included in the protection scope of the embodiment.

Claims

1. A method for wearable extremity behavior recognition based on residual networks, characterized in that, include: Obtain posture information of multiple parts of the wearer; Based on the posture information and the preset residual network model, the wearer's posture is identified; In the residual network model, shallow features of pose information are first extracted; then deep features of pose information are extracted using two residual blocks; some neurons are randomly deactivated during each training iteration; finally, the numerical vector is normalized into a probability distribution vector through a fully connected layer and activation function to obtain the final action category. The residual network model includes a gated cyclic unit, two residual blocks, and a random deactivation module connected in series. The residual block comprises two cascaded residual sub-modules; The residual submodule uses a gated loop unit module to extract time series features; the input data is propagated forward through cross-layer identity transformation mapping data lines.

2. The wearable extremity behavior recognition method based on residual networks as described in claim 1, characterized in that, The attitude information includes at least one or more of attitude angles and accelerations.

3. The wearable extremity behavior recognition method based on residual networks as described in claim 1, characterized in that, Remove peaks and spikes from the curves corresponding to the attitude information, and smooth the curves corresponding to the acquired attitude information.

4. The wearable extremity wearer behavior recognition method based on residual networks as described in claim 1, characterized in that, The pose information is divided into multiple windows of equal length without gaps; the length of the multiple windows and the window overlap rate are fixed.

5. The wearable extremity behavior recognition method based on residual networks as described in claim 1, characterized in that, The feature data of the attitude information is mapped to a distribution with a mean of 0 and a standard deviation of 1.

6. A wearable extremity behavior recognition system based on residual networks, implementing the steps of the wearable extremity behavior recognition method based on residual networks as described in any one of claims 1-5, characterized in that, include: The data acquisition module is configured to acquire posture information of multiple parts of the wearer; The behavior recognition module is configured to: recognize the wearer's behavior based on the posture information and a preset residual network model; In the residual network model, shallow features of pose information are first extracted; then, deep features of pose information are extracted using two residual blocks; some neurons are randomly deactivated during each training iteration; finally, the numerical vector is normalized into a probability distribution vector through a fully connected layer and an activation function to obtain the final action category.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the wearable extremity wearer behavior recognition method based on residual networks as described in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the wearable extremity wearer behavior recognition method based on residual networks as described in any one of claims 1-5.

Images