Human body activity distinguishing device and method

Abstract

According to the human body activity distinguishing device and method, through an intelligent insole system integrating an inertial sensor and a plantar pressure sensor and in combination with a multi-scale time window analysis method, gait data of a user are accurately collected and processed, daily activity modes such as walking, running and going upstairs and downstairs and changes of the modes are recognized, and the human body activity distinguishing effect is achieved. And real-time health feedback and rehabilitation support are provided. The intelligent insole system comprises an upper computer system, sensor integrated insole hardware and a module communicating with a user terminal and a cloud server. According to the system, multi-modal data is collected through an inertial sensor and a plantar pressure sensor, data processing is achieved in combination with data fusion and a multi-scale sliding time window method, finally, a random forest classifier is used for classifying activity modes, and the functions of health monitoring, posture improvement suggestion, falling risk assessment and the like are achieved. The exercise performance of the user is improved, and personalized health management is provided.

Description

Technical Field

[0001] The present invention relates to the fields of sensor technology and pattern recognition, and particularly relates to a human activity discrimination device and method. Background Art

[0002] Human activity discrimination aims to identify various actions of the human body in time and space, including gait, posture, and movement, etc. This technology is of great significance in health management and disease prevention. By accurately monitoring and analyzing an individual's daily activities, it can help evaluate changes in motor ability, develop personalized rehabilitation programs, and optimize intervention measures. During the rehabilitation process of patients with movement disorders such as stroke and Parkinson's disease, activity discrimination can help doctors optimize the treatment effect and avoid excessive or insufficient exercise. At the same time, by real-time monitoring the activity patterns of the elderly based on human activity discrimination, abnormalities can be detected early and preventive measures can be taken to reduce accidents such as falls and fractures, and improve the quality of life.

[0003] Existing human activity discrimination methods mainly include monitoring systems based on large devices and wired connections, such as high-precision cameras, force platforms, and motion capture systems. These systems provide high-precision data acquisition and real-time synchronization, and are suitable for specific experimental environments. However, these methods are complex to deploy, lack flexibility, and are difficult to apply to daily life. Although video monitoring methods can identify activities, they are easily affected by environmental factors and it is difficult to cover all activity areas, especially in scenarios with a large number of people or complex spaces. Activity discrimination methods based on wearable devices (such as smart bracelets) provide all-day monitoring, but since the devices can only provide single motion data, it is difficult to accurately identify complex activities, and they rely on wearing habits. Improper wearing or device damage will affect the accuracy. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies in the prior art in human activity discrimination and devices, and to propose a human activity discrimination method and device.

[0005] The present invention relates to a human activity discrimination device and method, which is particularly significant in the fields of health promotion, rehabilitation training, etc. The intelligent insole system of the present invention consists of two main parts: an upper computer system, which is responsible for interacting with the user and performing data analysis; and an intelligent insole hardware system integrating sensors and wireless communication modules, which can accurately collect the user's gait information. By continuously monitoring the user's activity patterns, the system can not only provide feedback on the daily health status, but also provide data support for rehabilitation training, helping rehabilitation doctors evaluate the patient's recovery situation and then optimize the treatment plan. In order to improve the accuracy of activity pattern recognition, the present invention designs an activity pattern recognition algorithm based on a multi-scale time window method. This algorithm can overcome the common limitations in the prior art, such as the recognition method based on single-sensor features cannot distinguish similar actions, such as going up and down stairs and walking on flat ground, and has insufficient sensitivity to complex gait patterns. Through multi-scale analysis, the present invention can more accurately distinguish different activity patterns, improve the application effect of the system in complex environments, and thus has a wider application prospect in the fields of health monitoring and rehabilitation training.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A human activity discrimination device includes a plantar pressure sensor, an inertial sensor, a main control module, a user terminal, and a cloud server. The main control module, the plantar pressure sensor, and the inertial sensor constitute an intelligent insole system. The intelligent insole system is connected to the user terminal, and the user terminal is connected to the cloud server through wireless communication. The cloud server is used to receive data and perform processing and analysis to identify the user's activity pattern. Specifically, it includes the following content:

[0008] The plantar pressure sensor is used to collect foot pressure data in real time. The main control module preprocesses the plantar pressure data and then sends it to the user terminal;

[0009] The inertial sensor includes an accelerometer and a gyroscope, and is used to collect the user's gait motion data;

[0010] The main control module sends the data collected by the sensor to the user terminal through a Bluetooth module;

[0011] The user terminal receives the data and uploads it to the cloud server. The cloud server analyzes the data and provides a health report.

[0012] A human activity discrimination method includes the following steps:

[0013] a. Data collection: Through the plantar pressure sensor and the inertial sensor, the user's plantar pressure data and gait data are collected in real time;

[0014] b. Data filtering and denoising: Filter the collected sensor data to remove low-frequency noise and high-frequency interference;

[0015] c. Time window segmentation: Segment the data using a multi-scale sliding time window to extract gait features in different time periods;

[0016] d. Feature extraction: Extract relevant feature values from the segmented data, including root mean square value, standard deviation, maximum and minimum values, zero crossing rate, spectral energy and dominant frequency component, step time, stride time, swing time and support time, double support time and double contact time, contact end time;

[0017] e. Feature fusion: Fusion the data features from the plantar pressure sensor and the inertial sensor to form a unified feature vector;

[0018] f. Classification and recognition: Use the random forest classification algorithm to classify the fused features and identify the current activity pattern.

[0019] The implementation method of the present invention is as follows:

[0020] 1). Wearing the device: In daily life, the user wears the intelligent insole according to personal needs. Place them in the left and right shoes of the user respectively to ensure that the intelligent insole is in close contact with the sole for the purpose of accurately collecting plantar pressure information. The user establishes a communication link with the user terminal (such as a smart phone) through Bluetooth connection.

[0021] 2). Device initialization: When the user wears it for the first time, the user needs to complete the preliminary settings of the device in the mobile application on the user terminal for personalized settings. During initialization, the main control module will automatically calibrate the sensor data to ensure the accuracy of each sensor.

[0022] 3). Data collection: After the device starts working, the intelligent insole will collect the plantar pressure information of the user in real time and process these data through the main control module. The main control module fuses the acceleration and angular velocity data collected by the inertial sensor with the plantar pressure data to generate detailed gait information (including dynamic change information such as walking speed, stride, step frequency, and center of gravity transfer).

[0023] 4). Monitoring of daily activity patterns: The user terminal will use the built-in algorithm to classify the user's daily activities in real time according to the real-time transmitted gait data.

[0024] 5). Data Upload and Analysis: During the activity, the user terminal will regularly upload the collected gait data to the cloud server. The cloud server uses a machine learning model to analyze the historical data, generate a detailed activity pattern report, and comprehensively evaluate it in combination with the user's activity data to infer possible health risks, such as fall risk and sedentary warning.

[0025] 6). Health Management and Feedback: The cloud server generates a health report based on the analysis results of the uploaded data and feeds it back to the user terminal. The report includes content such as exercise volume statistics, postural improvement suggestions, balance assessment, and fall risk analysis, helping users adjust their daily activity habits and improve their health conditions.

[0026] 7). Data Visualization and User Interaction: The user terminal provides an intuitive real-time data visualization function. Users can view gait curves, pressure distribution maps, and activity pattern classification results in the mobile application, so as to comprehensively understand their gait performance and physical health status in daily activities. Users can set and adjust goals based on the visualized data to improve their exercise and health levels. Description of the Drawings

[0027] Figure 1 is a flowchart of the implementation of a human activity discrimination device;

[0028] Figure 2 is a schematic diagram of the plantar pressure sensor with eleven pressure acquisition points in the present invention;

[0029] Figure 3 is a schematic diagram of the hardware design of the human activity discrimination device;

[0030] Figure 4 is a schematic diagram of the data transmission process of the human activity discrimination device;

[0031] Figure 5 is a flowchart of the activity pattern recognition algorithm based on the multi-scale sliding time window method;

[0032] Figure 6 is a schematic diagram of the multi-scale window segmentation method. Detailed Embodiments

[0033] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, the present invention adopts the following technical solutions.

[0034] The present invention provides a human activity discrimination device, which is a wearable device applied to human activity discrimination, including a plantar pressure sensor, an inertial sensor, a main control module, a user terminal, and a cloud server. The main control module, the plantar pressure sensor, and the inertial sensor form an intelligent insole. The intelligent insole is connected to the user terminal, and the user terminal is connected to the cloud server through wireless communication. The cloud server is used to receive data and perform processing and analysis to identify the user's activity pattern. Figure 1 The flowchart of the human activity discrimination device based on a wearable device is as follows:

[0035] Start: System startup preparation;

[0036] Wear the device: The user wears the intelligent insole device;

[0037] Device initialization: The system performs device identification and sensor initialization;

[0038] Data acquisition: The intelligent insole collects plantar pressure and IMU data;

[0039] Daily activity pattern monitoring: The system analyzes the user's activity status in real time;

[0040] Data upload and analysis: The terminal uploads the data to the cloud server for further processing;

[0041] Health management and feedback: Generate a health assessment report based on the analysis results and feedback it to the user;

[0042] Data visualization and user interaction: The user terminal displays the monitoring results in a visual way and provides interactive functions;

[0043] End: Complete a complete data processing process.

[0044] The plantar pressure sensor is connected to the main control module, on which an inertial sensor is integrated and can perform signal processing and communication. The user terminal that can communicate wirelessly with the main control module is connected to a cloud server.

[0045] The intelligent insole is on the user's foot and is used to collect the plantar pressure information of the user and transmit this information to the main control module.

[0046] Further, as Figure 2 shown, the plantar pressure sensor is made of flexible PET material, and the distribution of 11 pressure sensing acquisition points on the intelligent insole is shown. The sensing points are respectively arranged in the heel (2), arch (2), midfoot (2), forefoot (3), and toe (2) regions. All pressure sensors are made of flexible PET material, which can adapt to the changes in the plantar morphology and sense the pressure changes in real time.

[0047] The main control module is used to collect the foot movement information of the user and simultaneously process the collected plantar pressure information;

[0048] Furthermore, the main control module dynamically analyzes the plantar force distribution and foot movement conditions by combining the data of the inertial sensors. The main control module uses a data fusion algorithm to integrate the inertial sensor and plantar pressure data, and provides complete gait information for the user terminal.

[0049] Furthermore, the inertial sensor integrates an accelerometer and a gyroscope, and can simultaneously capture the linear velocity and angular velocity data of the foot during the wearer's activities.

[0050] Figure 3 It is a hardware structure block diagram of a human activity discrimination device. This block diagram shows the circuit connection relationship and data flow path between the hardware components. It mainly includes: the battery module provides energy for the system, and the voltage is regulated to 3.3V through a buck circuit for power supply; the charge and discharge pipeline module is used for charging control and voltage protection; the data acquisition circuit is connected to two sensor modules: the plantar pressure sensor and the inertial sensor, and is responsible for multi-modal data acquisition; the collected data is transmitted to the single-chip microcomputer main control module through the data acquisition circuit for processing; the single-chip microcomputer wirelessly transmits the data to the host computer (user terminal) through BLE (low-power Bluetooth) to realize data storage and analysis.

[0051] The described intelligent insoles are respectively placed in the left and right shoes of the user, and include a left shoe insole, a left shoe main control module, a right shoe insole and a right shoe main control module.

[0052] Furthermore, the user terminal initiates a broadcast through Bluetooth. When the device and the user terminal are within the communication range, the left shoe main control module and the right shoe main control module can communicate with the user terminal through Bluetooth connections respectively.

[0053] Furthermore, after the connection is established, a data transmission link is established between the left shoe main control module and the right shoe main control module and the user terminal through the Bluetooth protocol, and is ready to collect the foot information of the left and right feet.

[0054] Furthermore, as Figure 4 shown, during the data transmission process, the user terminal will regularly send data packets with time stamps to the left shoe main control module and the right shoe main control module to ensure the time stamp consistency between the left and right shoes, so as to ensure the timing alignment of the data.

[0055] Furthermore, during the time stamp calibration process, the user terminal adopts a time synchronization algorithm based on the average error method. The average error formula is as follows:

[0056] ;

[0057] where is the number of synchronization intervals, represents the timestamp sent by the master node (user terminal), represents the timestamp when the client (the main control modules of the left and right shoes) receives the synchronization packet. The specific implementation steps of time synchronization are as follows: The user terminal regularly broadcasts time synchronization data packets through the Bluetooth protocol, and each data packet is attached with the current timestamp of the user terminal. The main control modules of the left and right shoes, as slave nodes, record the local timestamp after receiving the data packet and calculate the difference between the local time and the master node timestamp. Estimate the clock frequency drift through the above formula and further adjust the local clock to ensure time synchronization between the main control modules of the left and right shoes.

[0058] The user terminal is used to receive the gait information of the user uploaded by the main control module and transmit it to the cloud server through the network, and at the same time receive the user's gait data and activity pattern analysis information feedback by the cloud server; The cloud server is used to receive the gait information of the user uploaded by the user terminal and perform analysis, generate the daily activity pattern recognition information of the user, and return the result to the user terminal.

[0059] Furthermore, after receiving the gait feature data, the user terminal can use the built-in algorithm to perform a local preliminary classification of the daily activity pattern, including basic activity types such as walking, standing, sitting, running, and going up and down stairs. At the same time, the user terminal can also upload the complete gait data to the cloud server. The cloud server combines the user's historical gait data and daily activity pattern, mines and analyzes the uploaded historical data through a machine learning model, identifies more complex activity patterns, such as sedentary warning, gait abnormality detection, fall risk assessment, etc., and generates a health report based on the analysis results. The health report includes but is not limited to exercise volume statistics, posture improvement suggestions, balance assessment, and fall risk analysis, etc., and is presented to the user through the mobile application interface of the user terminal.

[0060] Figure 5 is the activity pattern recognition flowchart based on the multi-scale sliding time window method, and the process includes:

[0061] Collect data from the IMU and FSI respectively;

[0062] Perform data filtering and denoising respectively;

[0063] Perform multi-scale sliding window data segmentation using short windows, medium windows, and long windows respectively;

[0064] Perform feature extraction on the segmented data respectively;

[0065] Fuse the features of the IMU and FSI data;

[0066] Classify the fused feature data using the random forest algorithm;

[0067] Implement activity pattern discrimination.

[0068] In the data collection stage, multi-modal information is collected from two different sensors (Inertial Measurement Unit (IMU) and Foot Pressure Sensor (FSI)) of the smart insole. The IMU sensor includes an accelerometer and a gyroscope, which are used to measure acceleration and angular acceleration in three-dimensional space. The accelerometer records the acceleration signal, while the gyroscope records the angular acceleration signal. The sampling frequency of the IMU data is 100 Hz, collecting 100 data points per second, and each data point contains three acceleration components and three angular acceleration components. The IMU data is stored in a time series form, and the shape of the data matrix is N×6, where N is the number of samples and 6 is the number of sensor features. The FSI sensor collects plantar pressure distribution data, which is usually composed of multiple pressure sensor arrays. Each sensor records the contact pressure and can sense the reaction force when the ground contacts the sole. The sampling frequency of the FSI data is also 100 Hz, collecting 100 data points per second, and records the pressure distribution data. The FSI data is stored in a time series form, and the shape of the data matrix is N×M, where N is the number of samples and M is the number of sensors.

[0069] After the data collection is completed, data preprocessing is required. First, data filtering and denoising are performed, and a low-pass filter is used to remove high-frequency noise in the IMU and FSI data.

[0070] Assume is the original sensor signal, and the filtered signal can be expressed as:

[0071] ;

[0072] where is the operation of the low-pass filter, which can be implemented by a digital filter:

[0073] ;

[0074] where are the filter coefficients, is the order of the filter. The cut-off frequency of the filter for IMU data is set to 20 Hz, and the cut-off frequency of the FSI data is set to 5 Hz. After filtering, the data becomes smoother, removing irrelevant high-frequency interference, and the data is clearer and suitable for subsequent processing.

[0075] In the feature extraction stage, the multi-scale sliding time window method is used to extract data features and capture motion patterns of different frequencies. This method captures data features by setting time windows of different sizes , Perform data cutting and extract features from the original data within each time window. Figure 6 It is a schematic diagram of the multi-scale time window segmentation method. It shows how to divide data segments on the time axis with different window lengths (short: 0.1 s, medium: 0.5 s, long: 2 s). The sliding step of each time window is half of the corresponding window length, forming an overlapping window to ensure the continuity and robustness of feature extraction.

[0076] Furthermore, the short time window is 0.1 second, which is used to capture rapidly changing signals and is suitable for extracting features of actions such as running and rapid swinging; the medium time window is 0.5 second, which is applicable to the feature extraction of regular gaits; the long time window is 2 seconds, which is suitable for capturing complex motion patterns or slowly changing features, such as standing still or walking slowly. Each window slides along the time series to extract the features within the window, and the sliding step of the window is half of the window length.

[0077] Furthermore, for the short time window seconds, assuming the sensor sampling frequency is , then the corresponding number of sampling points . Assuming the IMU data within the window is , , …, , the features of each time window can be extracted in the following way:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] ,

[0084] where , , are respectively , , the acceleration data in the , , axis directions, and , , are respectively the angular acceleration data in the

[0085] Among them, the time-domain features and frequency-domain features of the IMU data are calculated as follows:

[0086] The calculation method of the root mean square value RMS is: ;

[0087] Among them, is the acceleration data;

[0088] The calculation method of the standard deviation STD is: ;

[0089] Among them, is the acceleration data, is the signal The mean value, and the calculation formula is: ;

[0090] The calculation methods of the maximum value and the minimum value are: , ;

[0091] Among them, is the acceleration signal;

[0092] The calculation method of the zero crossing rate ZCR is: ;

[0093] Among them, is the acceleration signal, is the indicator function;

[0094] The calculation method of the spectral energy Energy is: ;

[0095] Among them, is the Fourier transform of, is the frequency component.

[0096] Furthermore, the feature extraction of the FSI data includes step time, step duration, swing time and support time, double support time and double contact time, etc. The step time and step duration are used to describe the time interval and duration between steps, the swing time and support time are used to analyze the dynamic characteristics of the gait, and the double support time and double contact time are used to distinguish the gait pattern.

[0097] Furthermore, there are two sensors in each of the heel area and the toe area of the plantar pressure sensor, which are respectively denoted as: heel area sensor: and , toe area sensor and ;

[0098] At any moment Next, calculate the resultant pressure in the heel region: Calculate the resultant pressure in the toe region: ;

[0099] Among them, (where 1, 2, …, );

[0100] To detect the touchdown or liftoff events in each region, corresponding thresholds need to be set: heel threshold , toe threshold , and at the same time, define the contact indication function for each region: , ;

[0101] Using the above regional contact determination, two critical moments can be determined:

[0102] Initial contact event IC. When the heel region changes from the non-contact state 0 to the contact state 1, it is considered that the heel touches the ground, and this moment is recorded as the heel touchdown time, denoted as ; Contact end event CE. When the heel region changes from the contact state 1 to the non-contact state, it is considered that the toe lifts off the ground, and this moment is recorded as the heel touchdown time, denoted as ;

[0103] Furthermore, the FSI data within the window is , , … , and the characteristics of each time window can be extracted in the following way:

[0104] ;

[0105] Among them, calculating the gait characteristics of the FSI data includes the following:

[0106] Stride time is the time interval between two consecutive initial contact events of the same foot: ;

[0107] Among them, is the sampling point sequence number corresponding to the th detected heel touchdown event;

[0108] Step time is the time difference between the initial contact events of the left and right feet respectively: ;

[0109] Among them, represents the heel touchdown sampling point sequence number, represents the heel touchdown sampling point sequence number of the contralateral foot;

[0110] Support time The duration of the foot's contact with the ground during a gait cycle, that is, the time from heel strike to toe-off: ;

[0111] in, Indicates the sequence number of the toe-off sampling point;

[0112] Swing Time is the time the foot is in non-contact state during the gait cycle: ;

[0113] Double support time When collecting left and right foot data, both left and right feet meet the contact event at the same time. At this time, the contact state of the left and right feet is: , it is considered that both feet are supporting at the same time, and in the current gait cycle: ;

[0114] in, is the indicator function, the interval Indicates the range of sampling points in the current gait cycle;

[0115] Double foot contact time is the time interval during which both the left and right feet are in contact at the same time in a gait cycle: ;

[0116] in, and are the initial contact moments of the left and right feet, and The contact end moments of the left and right feet, respectively.

[0117] Furthermore, assuming that the sensor sampling frequency is , then the time window The length is 0.5 seconds, corresponding to the number of sampling points In the time window, for any signal , assuming that the IMU data in the window is , ,…, , the features of each time window can be extracted by:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] ,

[0124] Among them, , , are respectively , , acceleration data in the axial direction, , ,, are respectively , , angular acceleration data in the axial direction.

[0125] Among them, the characteristic calculation formulas are respectively:

[0126] Root mean square value: ;

[0127] Standard deviation: ;

[0128] Maximum and minimum values: , ;

[0129] Zero crossing rate: ;

[0130] Spectrum energy: ;

[0131] The FSI data within the time window is , ,… The characteristics of each time window can be extracted in the following way:

[0132] ;

[0133] Among them, calculating the gait characteristics of FSI data includes the following:

[0134] Stride time = ;

[0135] Step time = ;

[0136] Support time = ;

[0137] Swing time = ;

[0138] Double - support time = ;

[0139] Double - foot contact time = ;

[0140] Wherein, is the acceleration data, is the mean value of the signal ; is the indicator function, is 's Fourier transform, is the frequency component, (where 1, 2, …, ), is the initial contact time, is the contact end time, represents the heel - strike sampling point number, represents the heel - strike sampling point number of the contralateral foot, is the th sampling point number corresponding to the detected heel - strike event, represents the toe - off sampling point number, is the contact state of the left and right feet, and are respectively the initial contact times of the left and right feet, and are respectively the contact end times of the left and right feet;

[0141] Furthermore, assuming the sensor sampling frequency is , then the long - time window has a length of 2 seconds, corresponding to the number of sampling points . In the long - time window, for any signal , assuming the IMU data within the window is , , …, , the features of each time window can be extracted in the following way:

[0142]

[0143]

[0144]

[0145]

[0146]

[0147] ,

[0148] Among them, , , are respectively , , acceleration data in the axial direction, , , are respectively , , angular acceleration data in the axial direction.

[0149] Among them, the characteristic calculation formulas are respectively:

[0150] Root mean square value: ;

[0151] Standard deviation: ;

[0152] Maximum and minimum values: , ;

[0153] Zero crossing rate: ;

[0154] Spectrum energy: ;

[0155] The FSI data within a long time window is , ,… , and the characteristics of each time window can be extracted in the following way:

[0156] ;

[0157] Among them, calculating the gait characteristics of FSI data includes the following:

[0158] Stride time = ;

[0159] Step time = ;

[0160] Support time = ;

[0161] Swing time = ;

[0162] Double support time = ;

[0163] Double - foot contact time = ;

[0164] Wherein, is the acceleration data, is the signal mean value, is the indicator function, is Fourier transform, is the frequency component, (where 1, 2, …, ), is the initial contact time, is the contact end time, represents the serial number of the heel - strike sampling point, represents the serial number of the heel - strike sampling point of the contralateral foot, is the th serial number of the sampling point corresponding to the detected heel - strike event, represents the serial number of the toe - off sampling point, is the contact state of the left and right feet, and are respectively the initial contact times of the left and right feet, and … are respectively the contact end times of the left and right feet. By moving a sliding window on the time series, the feature vectors of each window can be obtained and further concatenated into the final feature vector.

[0165] Furthermore, in the feature fusion stage, first use the min - max normalization method to scale the features of IMU and FSI to a unified range , ensuring that features from different sources are on the same scale, thereby enhancing the stability of the classification model. For each feature (a feature in IMU or FSI), its normalization expression is:

[0166] ;

[0167] Wherein, and are respectively the minimum and maximum values of the feature , is the normalized feature value, ensuring . The normalized IMU and FSI features are respectively and .

[0168] Next, the features extracted by the IMU and FSI are concatenated into a long feature vector. Suppose is the feature vector of the IMU, is the feature vector of the FSI. Then the fused feature vector is:

[0169] ;

[0170] The length of the concatenated feature vector is , where and are the numbers of IMU and FSI features respectively.

[0171] In the activity pattern classification stage, a random forest classifier is used to classify the features. During training, the labeled training data set is divided into two parts: features and labels. The features are the fused feature vectors , and the labels are the activity types: .

[0172] Use the random forest classifier , which is composed of decision trees. The maximum depth of each tree is 10. The construction process of each tree depends on information gain or Gini index for node splitting. The output of each decision tree is the activity class prediction , which is the prediction of the input sample by the tree according to the features . During the classification process, for each input sample , the random forest makes predictions through each tree. Suppose the random forest contains trees, and each tree outputs a prediction result , where . The final activity class prediction result is the voting result of all tree prediction results:

[0173] ;

[0174] Among them, is the indicator function. If the prediction result of the th tree is the activity type , then the function value is 1, otherwise it is 0. represents selecting the class with the most votes among all possible activity classes . Finally, the activity type is classified as the class predicted by the random forest model.

[0175] In the comprehensive evaluation and optimization stage, the system performance is optimized based on the experimental results. The optimization objectives include adjusting the time window size, filter design, and hyperparameters of the random forest, etc. Methods such as grid search and cross-validation are used for optimization.

[0176] Furthermore, the main control module adopts a low-power design, supporting long-term continuous gait monitoring. The built-in lithium battery effectively extends the device's battery life through an optimized power management algorithm, suitable for scenarios of long-term wearing and continuous monitoring of daily activities. At the same time, the main control module performs real-time preprocessing on the data collected by the sensors, reducing data redundancy and improving transmission efficiency.

[0177] Furthermore, the user terminal supports the real-time data visualization function, and users can view gait curves, pressure distribution maps, and activity pattern classification results through a mobile application. These visualized data intuitively display the user's movement trajectory and plantar pressure changes, helping users better understand their gait performance and physical health status.

Claims

1. A human activity discrimination device, characterized in that, It includes a plantar pressure sensor, an inertial sensor, a main control module, a user terminal, and a cloud server. The main control module, the plantar pressure sensor, and the inertial sensor constitute an intelligent insole system. The intelligent insole system is connected to the user terminal, and the user terminal is connected to the cloud server through wireless communication. The cloud server is used to receive data, process and analyze it, and identify the user's activity pattern. Specifically, it includes the following content: The plantar pressure sensor is used to collect foot pressure data in real time. The main control module preprocesses the plantar pressure data and then sends it to the user terminal. The inertial sensor includes an accelerometer and a gyroscope, and is used to collect the user's gait motion data. The main control module sends the data collected by the sensor to the user terminal through a Bluetooth module. The user terminal receives the data and uploads it to the cloud server. The cloud server analyzes the data and provides a health report.

2. The human activity discrimination device according to claim 1, characterized in that: The plantar pressure sensor is composed of multiple sensing elements, which are distributed in different parts of the insole and are used to accurately detect different stages in the user's gait.

3. The human activity discrimination device according to claim 1, wherein: The inertial sensor is integrated into the insole and works in cooperation with the plantar pressure sensor. It transmits data to the user terminal in real time through a Bluetooth module for further processing.

4. The human activity discrimination device according to claim 1, characterized in that: The user terminal is a smart phone, equipped with an application program, which is used to display the data analysis results and synchronize with the cloud server to update the user's health report.

5. A method for discriminating human activities, characterized in that, It includes the following steps: a. Data collection: Through the plantar pressure sensor and the inertial sensor, the user's plantar pressure data and gait data are collected in real time. b. Data filtering and denoising: The collected sensor data is filtered to remove low-frequency noise and high-frequency interference. c. Time window segmentation: The data is segmented using a multi-scale sliding time window to extract gait features in different time periods. d. Feature extraction: Relevant feature values are extracted from the segmented data, including root mean square value, standard deviation, maximum and minimum values, zero crossing rate, spectral energy and dominant frequency component, stride time, step time, swing time and support time, double support time and double foot contact time, contact end time. e. Feature fusion: The data features from the plantar pressure sensor and the inertial sensor are fused to form a unified feature vector. f. Classification and recognition: The random forest classification algorithm is used to classify the fused features to identify the current activity pattern.

6. The method for discriminating human activities according to claim 5, wherein: The data filtering and denoising are processed using a low-pass filter to reduce noise from the environment or the sensor. Among them, the calculation method of the low-pass filter is: ； Among them, is the original sensor signal, is the filtered signal, are the coefficients of the filter, is the order of the filter.

7. A method for discriminating human activities according to claim 5, characterized in that: The multi-scale sliding time window includes a short time window, a medium time window, and a long time window. Feature extraction is performed through time windows of different lengths to distinguish activity pattern features at different scales for subsequent model training. Among them, the short time window is 0.1 second, the medium time window is 0.5 second, the long time window is 2 seconds, and the step size of the sliding window is half of the window length; For the short time window , assuming the sensor sampling frequency is , the corresponding number of sampling points , the IMU data within the window is , , …, , the features of each time window are extracted in the following way: ； Among them, , , are respectively , , the acceleration data in the , , are respectively , , the angular acceleration data in the Among them, calculating the time domain features and frequency domain features of IMU data includes the following: The calculation method of the root mean square value RMS is as follows: ; wherein, is acceleration data; The calculation method of the standard deviation STD is as follows: ; Among them, is the acceleration data, is the signal mean value, and the calculation formula is: ; The calculation methods of the maximum and minimum values are: ， ； Among them, is the acceleration signal; The zero-crossing rate ZCR calculation method is as follows: ; Among them, is the acceleration signal, is the indicating function; The calculation method of spectral energy is as follows: ; wherein, is the Fourier transform of, and is the frequency component.

8. A method for discriminating human activities according to claim 5, characterized in that: For the FSI data within a short time window is , , … , the features of each time window are extracted as follows: ； Among them, calculating the gait features of FSI data includes the following: Stride time is the time interval between two consecutive initial contact events of the same foot: ; Among them, is the serial number of the sampling point corresponding to the th detected heel strike event; Step time is the time difference between the initial contact events of the left and right feet respectively: ; Among them, represents the serial number of the heel strike sampling point, represents the serial number of the heel strike sampling point of the opposite foot; Support time It is the duration of the foot's contact with the ground within a gait cycle, that is, the time from heel strike to toe off: ; Among them, represents the toe-off sampling point serial number; Swing time is the time when the foot is in a non-contact state during the gait cycle: ; Double support time When collecting data of the left and right feet, at the same moment, both the left and right feet satisfy the contact event, and the contact states of the left and right feet are as follows: , it is considered that both feet are supporting simultaneously. During the current gait cycle: ; wherein is an indicator function, and the interval represents the sampling point range within the current gait cycle; Double-foot contact time It is the time interval during which both the left and right feet are in the contact state simultaneously within one gait cycle: ; Among them, and are the initial contact times of the left and right feet respectively, and are the end times of contact of the left and right feet respectively.

9. The human activity discrimination method according to claim 7, characterized in that: The medium time window and the long time window use different time window lengths for feature extraction than the short time window. The specific definitions are as follows: Assume that the sensor sampling frequency is , then the medium time window has a length of 0.5 seconds, corresponding to the number of sampling points ; Within the medium time window, for any signal , assume that the IMU data within the window is , , …, , the features of each time window can be extracted in the following way: ； Among them, , , are respectively , , acceleration data in the , , are respectively , , angular acceleration data in the Among them, the feature calculation formulas are respectively: Root mean square value: ; Standard deviation: ; Maximum and minimum values: ， ； Zero crossing rate: ; Spectrum energy: ; The FSI data within the window is , , … , and the features of each time window can be extracted in the following way: ； Among them, the gait features for calculating FSI data include the following: Stride time = ; Step time = ; Support time = ; Swing time = ; Double-support time = ; Bipedal contact time = ; wherein, is the acceleration data, is the signal mean value, is the indicator function, is Fourier transform, is the frequency component, (wherein 1, 2, …, ), is the initial contact time, is the contact end time, represents the serial number of the heel strike sampling point, represents the serial number of the heel strike sampling point of the contralateral foot, is the serial number of the sampling point corresponding to the nth detected heel strike event, represents the serial number of the toe-off sampling point, is the contact state of the left and right feet, and are respectively the initial contact times of the left and right feet, and are respectively the contact end times of the left and right feet; The sensor sampling frequency is , then the long time window has a length of 2 seconds, corresponding to the number of sampling points . Within the long time window, for any signal , assuming the IMU data within the window is , , …, , the features of each time window are extracted in the following way: ； Among them, , , are respectively , , acceleration data in the , , are respectively , , angular acceleration data in the Among them, the feature calculation formulas are respectively: Root mean square value: ; Standard deviation: ; Maximum value and minimum value: ， ； Zero crossing rate: ; Spectrum energy: ; The FSI data within the window is , , … , and the features of each time window can be extracted in the following manner: ； Among them, the gait features for calculating FSI data include the following: Stride time = ; Step time = ; Support time = ; Swing time = ; Double support time = ; Bipedal contact time = ; wherein, is the acceleration data, is the signal mean value, is the indicator function, is Fourier transform, is the frequency component, (wherein 1, 2, …, ), is the initial contact time, is the end time of contact, represents the heel strike sampling point number, represents the heel strike sampling point number of the contralateral foot, is the sampling point number corresponding to the heel strike event detected for the represents the toe-off sampling point number, is the contact state of the left and right feet, and are respectively the initial contact times of the left and right feet, and are respectively the end times of contact of the left and right feet.

10. A method for discriminating human activities according to claim 5, characterized in that: The feature fusion method is the min-max normalization method. For each feature , which is a feature in the IMU or FSI, and its normalization expression is: ； wherein, and are respectively the minimum value and the maximum value of the feature , is the normalized eigenvalue, ensuring ; After that, the features extracted by the IMU and FSI are concatenated into a long feature vector. is the feature vector of the IMU. is the feature vector of the FSI. Then the fused feature vector is: ； The length of the concatenated feature vector is , where and are the numbers of IMU and FSI features respectively.

11. A method for discriminating human activities according to claim 5, characterized in that: The random forest classification model is composed of multiple decision trees. The maximum depth of each decision tree is 10, and the data is split according to the information gain or Gini index. Finally, classification is completed through a voting mechanism; Among them, the random forest contains n trees, and each tree outputs a prediction result , where . The final prediction result of the activity category is the voting result of the prediction results of all trees: ； Among them, is an indicator function. If the predicted result of the th tree for the class is the activity type , then the function value is 1; otherwise, it is 0. denotes selecting the class with the most votes among all possible activity categories .

12. The human activity discrimination method according to claim 5, characterized in that: The cloud server processes the collected gait data in real time and feeds back the classification results to the user terminal for the user's reference.

Images