Human motion state recognition device and method based on flow sensor

Abstract

The invention relates to a human motion state recognition device and method based on an airflow sensor, and the device comprises a wearable main body support, a sensor collection device, and a data processing device. The wearable main body bracket comprises a head, neck and chest bracket, a waist bracket, a limb wearable bracket and a sensor bracket; the sensor acquisition device comprises a support and an airflow acquisition module; the airflow acquisition module is used for acquiring airflow information generated when a user moves; the data processing device is installed on the waist support and used for receiving the airflow information and inputting the airflow information into a built-in motion state recognition model to obtain posture recognition data of a user, and the posture recognition data are transmitted to terminal equipment through the wireless transmission module. The human body motion state recognition device can keep high motion state accuracy in a complex environment, and provides a reliable alternative recognition way for a user who has difficulty in moving or is not suitable for depending on traditional images and inertial sensors.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human motion state monitoring, in particular to a human motion state recognition device and method based on an air flow sensor. BACKGROUND

[0002] Human motion state recognition, as one of the core technologies in the field of intelligent sensing, has been widely applied in intelligent medical care, rehabilitation training, human-computer interaction, motion monitoring and other important scenarios with the rapid development of sensing technology and intelligent analysis technology. For example, in the field of intelligent medical care, the real-time recognition of the limb motion state of a patient can assist in disease assessment and rehabilitation effect tracking; in the human-computer interaction scenario, accurate motion state recognition can improve the accuracy of device response and the naturalness of interaction; in motion monitoring, the technology can realize quantitative analysis of motion posture and motion intensity, providing data support for scientific training. At present, the mainstream human motion state recognition method mainly relies on two technical paths: one is an identification scheme based on an image sensor, which collects human image or video data through a camera, a depth camera or other devices, extracts human contour, joint point and other features using computer vision algorithms, and then determines the motion state; the other is an identification scheme based on an inertial sensor, which collects human inertial data such as acceleration, angular velocity and angle through wearable devices (such as smart bracelets, inertial measurement units IMU), and realizes motion state classification combined with feature engineering or simple machine learning models.

[0003] However, the above existing technologies have significant limitations in practical application: 1. Poor environmental adaptability: the scheme based on an image sensor is easily affected by factors such as changes in lighting conditions (such as strong light, weak light environment), obstructions (such as clothing obstruction, obstacle obstruction), leading to failure of image feature extraction, and thus reducing the motion state recognition accuracy; the scheme based on an inertial sensor has the problem of sensor drift, and the data error accumulates after long-term use, which also affects the recognition reliability. 2. Limited application population: most traditional schemes require wearing complex sensing devices or relying on fixed image acquisition devices, which have serious deficiencies in applicability and convenience for special populations with difficulty in movement (such as the elderly, rehabilitation patients), scenes that are not suitable for wearing additional devices (such as some medical rehabilitation scenes), and are difficult to meet the actual use requirements. 3. Limitations of feature extraction: existing schemes based on inertial sensors often rely on manually designed features, which are difficult to fully capture complex nonlinear features related to motion state; while some image recognition schemes use deep learning technology, but are limited by the constraints of image data collection, and the feature robustness in complex environments still needs to be improved. Therefore, the existing human motion state recognition method has room for improvement in adaptability to complex environments, applicability to special groups, and stability of recognition accuracy, and a new motion state recognition technology is urgently needed to break through the dependence bottleneck of traditional solutions and achieve high-precision recognition in complex environments while taking into account the convenience of special groups, providing a more reliable and flexible technical approach for human motion state recognition. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provides a human motion state recognition device based on an airflow sensor, which can maintain high motion state accuracy in complex environments and provide a reliable alternative recognition approach for users who are not able to or are not suitable for relying on traditional images and inertial sensors.

[0005] A second object of the present application is to provide a human motion state recognition method based on an airflow sensor.

[0006] The technical solution of the present application to solve the above technical problems is: A human motion state recognition device based on an airflow sensor, comprising a wearable main support, a sensor acquisition device arranged on the wearable main support, and a data processing device, wherein, The wearable main support comprises a head-neck-chest support, a waist support, a limb wearable support, and a sensor support arranged on the head-neck-chest support, the waist support, and the limb wearable support; The sensor acquisition device is installed on the sensor support, and the sensor acquisition device comprises an airflow acquisition module arranged on the sensor support, wherein the airflow acquisition module is used to acquire airflow information including wind speed information and wind direction information generated during user movement; The data processing device is installed on the waist support, and the data processing device is used to receive airflow information and input it into an internally set motion state recognition model to obtain posture recognition data of the user and transmit it to a terminal device through an internally set wireless transmission module.

[0007] Preferably, the head-neck-chest support comprises a ring-shaped headband main body and a first sensor support arranged at the front end of the ring-shaped headband main body, wherein the ring-shaped headband main body is an elastic structure with adjustable length, one end of the ring-shaped headband main body is provided with a plug buckle mechanism or a buckle mechanism, and the other end is provided with multiple adjustment holes; the first sensor support is integrally and fixedly connected with the ring-shaped headband main body.

[0008] Preferably, the limb-worn support comprises a ring-shaped belt body and a second sensor support arranged at the front end of the ring-shaped belt body, wherein the ring-shaped belt body is an elastic structure with adjustable length, one end of the ring-shaped belt body is provided with a snap buckle mechanism or a buckle mechanism, and the other end is provided with a plurality of adjustment holes; the second sensor support is fixedly connected with the ring-shaped belt body in an integrated manner.

[0009] Preferably, the waist support comprises a belt body with adjustable length and two groups of third sensor supports mounted on the belt body, wherein the belt body is provided with a snap buckle mechanism or a buckle mechanism; the third sensor supports are fixedly connected with the belt body in an integrated manner.

[0010] Preferably, the data processing device comprises a central control box; the central control box is mounted on the belt body, and a controller is arranged in the central control box; the controller is connected with the airflow collection modules on the head-neck-chest support, the waist support and the limb-worn support through signal transmission lines.

[0011] Preferably, the support comprises a shell and a mounting structure arranged in the shell for mounting the airflow collection modules, wherein the shell is a regular hexagonal structure, the lower part of the shell is connected with the sensor support; the top surface, the bottom surface and the inner sides of the six side surfaces of the shell are each provided with a group of airflow collection modules, and the airflow collection modules are airflow sensors.

[0012] A human motion state recognition method based on an airflow sensor, comprising the following steps: Step 1: wearing a human motion state recognition device, and detecting whether the user wears correctly; when it is detected that the user does not wear the human motion state recognition device correctly, prompting the user to adjust the wearing position and direction until the human motion state recognition device is in the correct state; Step 2: collecting airflow information of the user in the motion process through the airflow collection module, and transmitting the collected airflow information to the data processing device; the data processing device processes the airflow information, inputs the processed airflow information into a motion state recognition model arranged therein, obtains posture recognition data of the user, and transmits the posture recognition data to a terminal device through a wireless transmission module arranged therein.

[0013] Preferably, in step 2, the processing process of the data processing device on the airflow information is as follows: Step 201: performing a preprocessing operation including filtering, normalization and outlier rejection on the airflow information, and performing time sequence alignment on the preprocessed airflow information with different sampling frequencies; Step 202: performing matching analysis on the time-aligned airflow information to determine whether the user is in a motion state; if it is determined that the user is in motion, saving the corresponding airflow information; Step 203: dividing the airflow information according to a predetermined time window, performing feature extraction and fusion on the data in each time window, and obtaining fused features; Step 204: inputting the fused features into a trained motion state recognition model to output a human posture recognition result.

[0014] Preferably, in step 202, the step of determining whether the user is in a motion state is: inputting the preprocessed airflow information into a corresponding neural network classifier to obtain independent motion determination results of the airflow mode and the posture mode, respectively; then performing time alignment and sample statistics on the determination results of the airflow mode and the posture mode in a preset time window according to a unified timestamp; if the proportions of "motion" determination of the airflow mode and the posture mode both exceed the respective set threshold or the unified threshold, a motion state confirmation signal is output; otherwise, a stationary state signal is output.

[0015] Preferably, in step 203, one-dimensional convolutional neural network and bidirectional gated recurrent network are used to extract features from the preprocessed airflow data, respectively, to obtain two types of complementary airflow feature vectors; then, through a weighted splicing mechanism or an attention fusion mechanism, the two types of complementary airflow feature vectors are unified in dimension and information is fused to generate structured fusion features; finally, the fusion features are input into a preset motion state recognition model for subsequent processing.

[0016] Compared with the prior art, the present application has the following beneficial effects: 1. The human motion state recognition method based on airflow sensors of the present application combines multi-channel sensor signals and significantly improves the motion recognition accuracy and robustness through a deep learning algorithm.

[0017] 2. The human motion state recognition method based on airflow sensors of the present application can effectively distinguish between head motion, upright walking, upright running, sitting, lying, turning, bending and other motion states, and can still maintain a very high recognition rate even in an environment with airflow interference or high sensor noise.

[0018] 3. Compared with recognition methods that only rely on vision or inertial data, the human motion state recognition method based on airflow sensors of the present application uses airflow features to compensate for the recognition blind area of inertial sensors in small amplitude motion, thereby achieving more stable and reliable motion state recognition.

[0019] 4、The human body motion state recognition device based on airflow sensor can realize synchronous collection and analysis of airflow signals of multiple parts, and provides reliable technical support for human-computer interaction, motion rehabilitation and other applications. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 It is a structure schematic diagram of a head-neck-chest support in the human body motion state recognition device based on airflow sensor.

[0021] Figure 2 It is a structure schematic diagram of a limb wearable support.

[0022] Figure 3 It is a structure schematic diagram of a waist support.

[0023] Figure 4 It is a flow schematic diagram of the human body motion state recognition method based on airflow sensor.

[0024] In the figure: 1-airflow collection module; 2-head-neck-chest support; 3-limb wearable support; 4-waist support; 5-central control box; 6-signal transmission line. DETAILED DESCRIPTION

[0025] The application will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the application are not limited thereto.

[0026] Referring to Figures 1-3 The human body motion state recognition device based on airflow sensor comprises a wearable main support, a sensor collection device arranged on the wearable main support and a data processing device.

[0027] Referring to Figures 1-3 The wearable main support comprises a head-neck-chest support, a waist support, a limb wearable support and a sensor support arranged on the head-neck-chest support, the waist support and the limb wearable support, wherein, The head-neck-chest support can be installed on the head, the neck and the chest, and comprises a ring-shaped headband main body and a first sensor support arranged at the front end of the ring-shaped headband main body, wherein the ring-shaped headband main body is an elastic structure with adjustable length, one end of the ring-shaped headband main body is provided with a plug buckle or a buckle mechanism, and the other end is provided with a plurality of adjustable holes for adapting to users with different girths; the first sensor support is fixedly connected with the ring-shaped headband main body in an integrated manner. The limb-worn support can be installed on the wrist, elbow, left and right knees or ankle, including a ring-shaped belt body and a second sensor support arranged at the front end of the ring-shaped belt body, wherein the ring-shaped belt body is an elastic structure with adjustable length, one end of the ring-shaped belt body is provided with a snap buckle mechanism or a buckle mechanism, and the other end is provided with a plurality of adjustment holes for adapting to different wearing parts. The waist support includes a belt body with adjustable length and two groups of third sensor supports installed on the belt body, wherein the belt body is provided with a snap buckle or buckle mechanism; and the third sensor supports are fixedly connected with the belt body in an integrated manner.

[0028] Referring to Figures 1-3 The sensor acquisition device is installed on the sensor support, and the sensor acquisition device includes an airflow acquisition module arranged on the sensor support, wherein the airflow acquisition module is used to acquire airflow information including wind speed information and wind direction information generated during user movement.

[0029] Referring to Figures 1-3 The data processing device is installed on the waist support, and the data processing device is used to receive airflow information and input into a motion state recognition model, to obtain posture recognition data of the user, and transmit the posture recognition data to a terminal device through a built-in wireless transmission module, wherein the data processing device includes a central control box, the central control box is installed on the belt body, and a controller is arranged in the central control box; and the controller is connected with the airflow acquisition modules on the head-neck-chest support, waist support and limb-worn support through a signal transmission line.

[0030] In this embodiment, the belt body is connected with the central control box through thirteen signal transmission lines, and a single-chip microcomputer module (i.e. a controller) is arranged in the central control box, which is used to receive airflow signals of each sensor and perform data preprocessing and fusion analysis.

[0031] Referring to Figures 1-3 The sensor support is fixedly connected with the ring-shaped headband body, ring-shaped belt body and belt body or is an integrated structure, wherein the sensor support includes a support column, a housing arranged on the support column and a mounting structure arranged in the housing for mounting an airflow acquisition module, wherein the housing is a regular hexagonal structure, the lower part of the housing is connected with the support column (or an adjusting structure for adjusting the angle and position of the housing is arranged between the housing and the support column); one group of airflow acquisition modules is arranged on the inner side of the top surface, bottom surface and six side surfaces of the housing, and the airflow acquisition modules are airflow sensors. In the embodiment, the support column is a hollow structure, and a cable channel is arranged in the support column, which is used for transmitting power supply and signal lines, so that stable connection between the wind speed and direction sensor array and the data processing device is realized.

[0032] Referring to Figures 1-4 The human motion state recognition method based on the airflow sensor comprises the following steps: Step 1: wearing the human motion state recognition device, and detecting whether the user wears correctly; when it is detected that the user does not wear the human motion state recognition device correctly, prompting the user to adjust the wearing position and direction until the human motion state recognition device is in the correct state; In the embodiment, the head-neck-chest support is worn on the head of the user, and the position of the first sensor support is adjusted through the plug buckle mechanism of the annular headband body in the head-neck-chest support, so that the airflow collection module installed in the first sensor support is in full contact with the respiratory airflow of the user; after wearing the human motion state recognition device of the present application, the quality of wearing needs to be detected automatically, specifically: instructing the user to perform a short-time standard action, which is two deep breaths, one hand lifting and three steps of in-place stepping; the data processing module detects the wind speed peak value and fluctuation amplitude of the airflow sensor at each part of the body in the action stage: if the airflow sensors located at the chest and neck produce periodic low-frequency airflow fluctuations (for example, the airflow fluctuation frequency is lower than the set value) when deep breathing, it is determined that the wearing is good; if the airflow sensors located at the wrist, elbow, knee or ankle do not appear significant wind speed change in the action stage, it is determined that the airflow sensor at the part is loose or the installation direction is improper; if the airflow sensor located at the waist does not detect regular periodic airflow change in the walking action, the user is prompted to adjust the wearing position again.

[0033] Step 2: collecting the airflow information of the user in the motion process through the airflow collection module, and transmitting the collected airflow information to the data processing device; the data processing device performs preprocessing operations including filtering, normalization and outlier rejection on the airflow information, and performs time sequence alignment on the preprocessed airflow information with different sampling frequencies; the airflow information after time sequence alignment is matched and analyzed to determine whether the user is in a motion state; if it is determined that the user is in a motion state, the corresponding airflow information is saved; then the airflow information is divided according to a predetermined time window, the data in each time window is extracted and fused to obtain fusion features; the fusion features are input into the trained motion state recognition model, wherein, The motion state recognition model comprises a one-dimensional convolutional neural network (1D-CNN), a bidirectional gated recurrent unit (Bi-GRU), an attention calculation layer, a long short-term memory unit (LSTM), and a fully connected output layer; wherein the one-dimensional convolutional neural network (1D-CNN) is used to extract local time sequence features of each airflow collection module, the bidirectional gated recurrent unit (Bi-GRU) is used to capture the front-back dependence of the airflow signal, the attention calculation layer is used to calculate the feature weight of the airflow sensor of different parts of the body and perform weighted fusion, the long short-term memory unit (LSTM) is used to model the global time sequence relationship of the fused features, and the fully connected output layer and the Softmax function are used to output the classification results of each motion state; the human motion state recognition result is output through the motion state recognition model, and the recognition result is sent to the terminal device through the wireless transmission module.

[0034] In the above process, the step of determining whether the user is in a motion state is: The airflow is collected in real time through the airflow collection module, and the collected airflow information is filtered, normalized, and feature extracted, wherein the airflow information (or airflow data) collected by the airflow collection modules from different body parts is respectively extracted for specific features: For the end of the limbs: the impact intensity, fluctuation frequency peak value, and wind direction change entropy value of the airflow signal are extracted; For the core of the trunk: the steady flow velocity mean value, periodic fluctuation amplitude, and airflow vortex characteristics of the airflow signal are extracted; The extracted features are input into a multi-branch neural network classifier, each branch in the multi-branch neural network classifier corresponds to a predetermined body region, and the preliminary discrimination of the local motion mode is respectively completed, and finally the motion determination result of the airflow mode and the attitude mode is output, wherein the specific discrimination logic and the corresponding motion type of each branch are as follows: The network branch corresponding to the left wrist and right wrist sensors: discriminates against upper limb movements (such as arm swinging, waving, throwing, etc.); when the wind speed parameter is greater than a certain set threshold and the high change entropy value is detected, the upper limb movement flag is triggered; The network branch corresponding to the left ankle, right ankle, left knee, and right knee sensors: discriminates against lower limb movements (such as stepping, kicking, jumping, etc.); when the characteristic contains a high periodic fluctuation amplitude and a specific frequency peak value (e.g., exceeding the corresponding set threshold) of the wind speed and direction are detected, the lower limb movement flag is triggered; The network branch corresponding to the chest and abdomen sensors: discriminates against trunk movements (such as deep breathing, trunk twisting, etc.); when the characteristic contains a significant increase in steady flow velocity accompanied by a specific vortex pattern, the trunk movement flag is triggered; Network branch corresponding to head and neck sensor: distinguish head movement (such as nodding, shaking, lifting head, etc.); when detecting that the feature contains short-time and one-way airflow impact, trigger the head movement flag.

[0035] The determination results of the airflow mode and the attitude mode are time-aligned and sample statistics are completed within a preset time window according to a unified timestamp; if the "motion" determination proportions of the airflow mode and the attitude mode both exceed the respective set threshold or the unified threshold, a motion state confirmation signal is output; otherwise, a static state signal is output.

[0036] The specific steps of motion judgment of airflow information based on the neural network classifier are as follows: (1) Data preprocessing: real-time acquisition of time series data of wind speed and wind direction, data segmentation is completed through sliding window, and feature parameters in the window are extracted combined with frequency energy analysis; (2) Time series feature modeling: input the preprocessed feature parameters into a recurrent neural network classifier (such as LSTM, GRU), use the built-in memory unit in the network to capture the time series dependence of airflow changes, and output the classification results of each body region branch; (3) Motion state cooperative determination: integrate the classification results of all branches, and finally confirm according to the preset motion cooperative logic rule, wherein the motion cooperative logic rule is: If only the lower limb branch is triggered, and the left and right ankle signals present alternating periodic characteristics, it is determined to be a walking or running state; If the upper limb branch and the lower limb branch are triggered at the same time, and the chest sensor detects a strong vortex feature (for example, exceeding the corresponding set threshold), it is determined to be a high-intensity whole body movement; If only the head and neck branches are triggered, and the trunk sensor signal remains stable, it is determined to be head posture adjustment; If all branch signals are below the determination threshold, it is confirmed that the user is in a static state.

[0037] Through the above motion cooperative logic rule, accurate identification of head movement, upright walking, upright running, sitting movement, lying movement, turning, flexion and other motion states can be realized.

[0038] In addition, the feature fusion and motion state recognition steps of multi-channel airflow information are as follows: A one-dimensional convolutional neural network (1D-CNN) is used to process the time series data of each sensor node (airflow acquisition module) respectively, and the local airflow pattern features of each part are accurately extracted; The local airflow mode characteristics of each part are input into a bidirectional gate recurrent unit (Bi-GRU) network, the preceding and following dependence of the airflow signal of each part in the time dimension is captured, and a high-level time sequence feature vector corresponding to each part is output; The high-level time sequence feature vectors of the 13 parts are spliced to construct a global feature matrix; the global feature matrix is input into an attention calculation layer, the attention calculation layer dynamically learns the importance weight of the airflow information of each part under different motion states (for example, when identifying "running", higher weights are automatically assigned to ankle, knee and chest sensors; when identifying "turning in a sitting position", higher weights are assigned to abdominal and elbow sensors); The learned attention weight is used to weight and fuse the global feature matrix to generate a unified fusion feature vector focusing on the current key motion part; The fusion feature vector is input into a recognition network containing a long short-term memory unit (LSTM), a complete motion sequence across time periods is modeled, and the starting, duration and ending process of the motion is understood; through a fully connected layer and a softmax function, the probability distribution of all predefined motion states such as head motion, upright walking, upright running, sitting position, lying position, turning and flexion is output; the class corresponding to the maximum probability is selected as the final recognition result, which is transmitted to the terminal device through a wireless module.

[0039] The above is the preferred embodiment of the present application, but the embodiments of the present application are not limited by the above, any change, modification, substitution, combination, simplification made without departing from the spirit and principles of the present application should be an equivalent replacement method, which is included in the protection scope of the present application.

Claims

1. A human motion state recognition device based on an air flow sensor, characterized by, The device comprises a wearable body support, a sensor acquisition device arranged on the wearable body support, and a data processing device. The wearable body support comprises a head-neck-chest support, a waist support, a limb wearable support, and sensor supports arranged on the head-neck-chest support, the waist support, and the limb wearable support. The sensor acquisition device is mounted on the sensor supports and comprises airflow acquisition modules arranged on the sensor supports. The data processing device is mounted on the waist support and is configured to receive airflow information and input the information into a built-in motion state recognition model to obtain posture recognition data of a user and transmit the data to a terminal device through a built-in wireless transmission module.

2. The human motion state recognition apparatus based on an air flow sensor according to claim 1, wherein The head-neck-chest support comprises a ring-shaped headband main body and a first sensor support arranged at the front end of the ring-shaped headband main body.

3. The human motion state recognition apparatus based on an air flow sensor according to claim 1, wherein The limb wearable support comprises a ring-shaped band main body and a second sensor support arranged at the front end of the ring-shaped band main body.

4. The human motion state recognition apparatus based on an air flow sensor according to claim 1, wherein The waist support comprises a length-adjustable belt main body and two groups of third sensor supports mounted on the belt main body.

5. The human motion state recognition apparatus based on an air flow sensor according to claim 4, wherein The data processing device comprises a central control box.

6. The human motion state recognition apparatus based on an air flow sensor according to claim 1, wherein The support comprises a shell and a mounting structure arranged in the shell for mounting the airflow acquisition modules.

7. A human motion state recognition method based on an air flow sensor, characterized by, The device comprises the following steps: Step 1: Wear the human motion state recognition device and detect whether the user wears it correctly. When it is detected that the user does not wear the human motion state recognition device correctly, the user is reminded to adjust the wearing position and direction until the human motion state recognition device is in the correct state. Step 2: The airflow information of the user during the exercise is collected by the airflow collection module, and the collected airflow information is transmitted to the data processing device; the data processing device processes the airflow information, and inputs the processed airflow information into the exercise state recognition model built-in, obtains the posture recognition data of the user, and transmits the posture recognition data to the terminal device through the built-in wireless transmission module.

8. The human motion state recognition method based on an air flow sensor according to claim 7, wherein In step 2, the processing process of the data processing device on the airflow information is as follows: Step 201: The airflow information is preprocessed including filtering, normalization and outlier rejection, and the airflow information with different sampling frequencies after preprocessing is executed time alignment; Step 202: The airflow information after time alignment is matched and analyzed to determine whether the user is in a motion state; if it is determined that the user is exercising, the corresponding airflow information is saved; Step 203: The airflow information is divided according to the predetermined time window, and the data in each time window is extracted and fused to obtain the fused features; Step 204: The fused features are input into the trained exercise state recognition model to output the human posture recognition result.

9. The human motion state recognition method based on an air flow sensor according to claim 8, wherein In step 202, the step of determining whether the user is in a motion state is as follows: The preprocessed airflow information is input into the corresponding neural network classifier to obtain independent motion determination results of the airflow mode and the posture mode respectively; then the determination results of the airflow mode and the posture mode are time-aligned and sample-counted in the preset time window according to the unified timestamp; If the "motion" determination proportion of the airflow mode and the posture mode both exceeds the set threshold value or the unified threshold value, a motion state confirmation signal is output; otherwise, a static state signal is output.

10. The human motion state recognition method based on an air flow sensor according to claim 9, wherein In step 203, one-dimensional convolutional neural network and bidirectional gated recurrent network are used to extract features from the preprocessed airflow data to obtain two types of complementary airflow feature vectors; then the two types of complementary airflow feature vectors are dimensionally unified and information fused through a weighted splicing mechanism or an attention fusion mechanism to generate structured fusion features; Finally, the fusion features are input into the preset exercise state recognition model for subsequent processing.

Images