Sitting posture intelligent recognition method, system and device

Abstract

The application discloses a sitting posture intelligent recognition method, system and device, relates to the technical field of neural network architecture, and comprises the following steps: acquiring real-time pressure distribution data of the hip and thigh area of a user through a flexible pressure sensor array embedded with coordinates; transmitting the real-time pressure distribution data to a pre-constructed CNN-LSTM hybrid neural network model for feature extraction and classification, and outputting the sitting posture category at the current moment; the CNN-LSTM hybrid neural model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory (LSTM) network; and based on the sitting posture category, a preset health management strategy is executed. The application realizes high-precision dynamic recognition of the sitting posture by using the CNN-LSTM hybrid neural network under the premise of guaranteeing the privacy and safety of the user and non-perception use, and provides the effect of active health management intervention.

Description

Technical Field

[0001] This application relates to the technical field of neural network architecture, and in particular to a sitting posture intelligent recognition method, system and device. Background Technology

[0002] With changes in modern office and lifestyle patterns, spinal health problems caused by prolonged sitting are receiving increasing attention. Achieving accurate and comfortable posture monitoring has become a crucial requirement in the field of smart health. Currently, posture monitoring technology based on pressure sensor arrays is widely used due to its advantages of seamless experience and privacy protection. Existing technologies typically collect hip pressure distribution data by deploying sensor arrays within the seat cushion and combining them with traditional machine learning algorithms such as Support Vector Machines (SVM) and Random Forests. Posture classification relies on manually designed features such as pressure center offset and pressure area.

[0003] However, this technical solution based on traditional algorithms and manual features has significant drawbacks when dealing with complex and ever-changing real-world application scenarios. Due to the limited ability of manual feature extraction, it is difficult to fully explore the high-dimensional spatial information and temporal dynamic patterns contained in stress data. As a result, this type of method cannot effectively capture subtle changes and dynamic transitions in user posture. Furthermore, the recognition accuracy drops significantly when facing users of different body types and seat surfaces of different materials, and it severely lacks cross-scenario generalization ability. Therefore, it cannot meet the actual needs of high-precision real-time recognition of complex sitting postures to support accurate health management.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a sitting posture intelligent recognition method, system and device, which aims to solve the technical problem that the existing technology is difficult to capture the spatiotemporal dynamic characteristics of pressure data, resulting in low recognition accuracy and failure to meet the requirements of high-precision recognition.

[0006] To achieve the above objectives, this application proposes a sitting posture intelligent recognition method, the method comprising: Real-time pressure distribution data of the user's buttocks and thighs is obtained through a flexible pressure sensor array with embedded coordinates. The real-time pressure distribution data is transmitted to a pre-constructed CNN-LSTM hybrid neural network model for feature extraction and classification, outputting the current sitting posture category; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM). Based on the sitting posture category, a preset health management strategy is executed.

[0007] In one embodiment, acquiring real-time pressure distribution data of the user's buttocks and thighs includes adaptive coordinate correction, specifically including... Detect the initial pressure distribution when the user first sits down and identify the pressure center points on the left and right sides of the buttocks; Calculate the offset between the pressure center point and the geometric center of the flexible pressure sensor array; Generate a coordinate transformation matrix based on the offset; The acquired real-time pressure distribution data is dynamically corrected using the coordinate transformation matrix to obtain real-time pressure distribution data.

[0008] In one embodiment, the construction steps of the CNN-LSTM hybrid neural network model include: Acquire pressure distribution sample data under different seat materials and / or different body types to construct an initial training dataset; The initial training dataset is preprocessed to generate standardized training samples that include the time-series dimension; Construct an initial network architecture that includes convolutional layers, pooling layers, LSTM layers, fully connected layers, and a softmax classification layer; The standardized training samples are input into the initial network architecture for forward propagation calculation to obtain the preset sitting posture category; Calculate the loss function value between the preset sitting posture category and the true label of the sample, and use the backpropagation algorithm to iteratively update the network weight parameters according to the loss function value until the model converges, thus obtaining the pre-constructed CNN-LSTM hybrid neural network model.

[0009] In one embodiment, the step of acquiring real-time pressure distribution data of the user's buttocks and thighs includes: The physical coordinates of the flexible sensor array are used to construct a two-dimensional pressure matrix from the real-time pressure distribution data; The two-dimensional pressure matrix is ​​normalized, and the two-dimensional pressure matrix of continuous time frames is extracted using the sliding window technique to construct a three-dimensional tensor sample containing the temporal dimension.

[0010] In one embodiment, the processing procedure of the convolutional neural network (CNN) includes spatial feature extraction, specifically including: The three-dimensional tensor sample is received, and the local spatial features of the pressure distribution are extracted through multi-layer convolution operations; Downsampling using pooling layers preserves local stress extrema and suppresses background noise; The output includes spatial feature maps of each frame containing the coordinates of the pressure center and the shape information of the high-pressure area, forming a spatial feature sequence.

[0011] In one embodiment, the processing procedure of the Long Short-Term Memory (LSTM) network includes temporal dynamic feature extraction, specifically including: The spatial feature sequence is input into the LSTM network step by step. By using the forget gate, input gate, and output gate mechanisms within the LSTM network, the dynamic variation patterns and long-term dependencies between the pressure distributions of consecutive frames are captured, and a temporal context feature vector containing attitude transition information is output. The temporal context feature vector is input into a fully connected layer for feature fusion mapping, and then processed by a Softmax classifier to output the user's sitting posture category at the current time.

[0012] In one embodiment, the method further includes lightweight deployment of a CNN-LSTM hybrid neural network model, specifically including: Obtain the trained CNN-LSTM hybrid neural network model; The CNN-LSTM hybrid neural network model is structurally pruned to remove redundant convolutional kernel channels whose absolute weights are below a preset threshold, and a simplified sparse network model is output. Perform integer quantization on the sparse network model to convert the model weight parameters and intermediate activation values ​​from floating-point format to fixed-point integer format, and output the quantized model file; The quantization model file is compiled to generate an inference engine adapted to the embedded edge computing chip, and the inference engine is deployed to the microcontroller unit of the seat cushion.

[0013] In one embodiment, the step of executing a preset health management strategy includes: Pre-build a library of corrective actions for various poor sitting postures; Receive the sitting posture category and retrieve a target corrective action that matches the current poor sitting posture from the corrective action library; Obtain the user's historical body data, and adaptively adjust the difficulty level of the target corrective action based on the historical body data; Generate guidance information that includes key points of the adjusted actions.

[0014] Furthermore, to achieve the above objectives, this application also proposes a sitting posture intelligent recognition system, which includes: The data acquisition module is used to acquire real-time pressure distribution data of the user's buttocks and thighs through a flexible pressure sensor array with embedded coordinates; The posture recognition module is used to transmit the real-time pressure distribution data to a pre-built CNN-LSTM hybrid neural network model for feature extraction and classification, and output the posture category at the current moment; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM). The health management module is used to execute preset health management strategies based on the sitting posture category.

[0015] In addition, to achieve the above objectives, this application also proposes a sitting posture intelligent recognition device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the sitting posture intelligent recognition method as described above.

[0016] This application proposes a sitting posture intelligent recognition method, system, and device. The method includes: acquiring real-time pressure distribution data of a user's buttocks and thighs using a flexible pressure sensor array with embedded coordinates; transmitting the real-time pressure distribution data to a pre-constructed CNN-LSTM hybrid neural network model for feature extraction and classification, and outputting the sitting posture category at the current moment; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM); and executing a preset health management strategy based on the sitting posture category. This application, by employing a flexible pressure sensor array with embedded coordinates and combining it with a hybrid neural network model composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM) for feature extraction and classification, solves the problem that existing technologies rely on manual feature extraction, making it difficult to effectively capture the high-dimensional spatial information and temporal dynamic patterns of pressure data. Compared with existing technologies, it achieves automatic deep mining of sitting posture features, significantly improving the accuracy and generalization ability of sitting posture recognition, thereby ensuring the precise execution of health management strategies. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the intelligent posture recognition method of this application. Figure 2This is a flowchart illustrating Embodiment 2 of the intelligent posture recognition method of this application. Figure 3 This is a flowchart illustrating Embodiment 3 of the intelligent posture recognition method of this application; Figure 4 This is a flowchart illustrating Embodiment 4 of the intelligent posture recognition method of this application; Figure 5 This is a schematic diagram of the module structure of the intelligent posture recognition system according to an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the sitting posture intelligent recognition method in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Because existing technologies struggle to capture the spatiotemporal dynamics of stress data, they result in low recognition accuracy, failing to meet the demands for high-precision recognition.

[0024] This application provides a solution that uses a flexible pressure sensor array with embedded coordinates and combines it with a hybrid neural network model consisting of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM) for feature extraction and classification. This solves the problem that existing technologies rely on manual features and cannot effectively capture the high-dimensional spatial information and temporal dynamic patterns of pressure data. Compared with existing technologies, this solution achieves automatic deep mining of sitting posture features, significantly improving the accuracy and generalization ability of sitting posture recognition, thereby ensuring the precise execution of health management strategies.

[0025] Based on this, embodiments of this application provide a sitting posture intelligent recognition method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the intelligent posture recognition method of this application.

[0026] In this embodiment, the intelligent posture recognition method includes steps S10 to S30: Step S10: Real-time pressure distribution data of the user's buttocks and thighs is acquired through a flexible pressure sensor array with embedded coordinates.

[0027] It should be noted that, in this embodiment, the flexible pressure sensor array with embedded coordinates refers to a sensing network composed of a flexible material substrate. This array consists of multiple independently distributed pressure sensing units, and each sensing unit has clear position coordinate information in its physical structure. It can conform to the curvature of the human body and sense changes in the pressure of the contact surface. The buttocks and thigh areas are the main body parts that come into contact with the seat when the user is seated. The pressure distribution in this area directly reflects the user's sitting posture. The real-time pressure distribution data refers to the set of values ​​collected by the sensor array at the current moment and in a continuous time series, reflecting the pressure intensity at each coordinate point.

[0028] This embodiment utilizes the deformation recovery properties of flexible materials and an arrayed spatial layout to convert physical pressure into electrical signals and record the corresponding spatial coordinates, thereby forming raw data containing spatial information. The aim is to capture the mechanical information of the user's body contacting the seat surface imperceptibly through a high-density flexible sensor array.

[0029] In one possible implementation, the flexible pressure sensor array with embedded coordinates can be made using sensing principles based on piezoresistive, capacitive, or piezoelectric effects. Its flexible substrate material can be selected from materials with good biocompatibility and elasticity, such as polydimethylsiloxane (PDMS), polyimide (PI), or conductive fabric. The sensor units can be arranged in a rectangular grid or a honeycomb-like close arrangement to improve the spatial resolution of data acquisition.

[0030] In one possible implementation, M1616L 16 can be used. 16-resolution flexible thin-film pressure sensor (thickness <0.3mm, single-point range 1-50kg). This sensor array is evenly distributed inside the seat cushion, directly collecting real-time pressure distribution data of the user's buttocks and thighs.

[0031] In one specific implementation, when a user sits on a smart cushion equipped with the sensor array, the sensor array collects data at a frequency of 10 to 50 times per second. The system synchronously records the index coordinates of each sensing unit and its corresponding voltage value. After analog-to-digital conversion, a real-time pressure distribution heat map data stream reflecting the pressure on the user's buttocks and thighs is generated.

[0032] Step S20: The real-time pressure distribution data is transmitted to a pre-constructed CNN-LSTM hybrid neural network model for feature extraction and classification, and the current sitting posture category is output; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM).

[0033] It should be noted that in this embodiment, the CNN-LSTM hybrid neural network model is a deep learning architecture that combines convolutional neural networks and long short-term memory networks (LSTM). This architecture integrates the advantages of the two networks through concatenation or fusion to process sequence data with both spatial and temporal characteristics. Feature extraction refers to the process by which the model automatically identifies and extracts key information with discriminative power (such as the trajectory of the pressure center, changes in the shape of the contact surface, etc.) from the original input data. Sitting posture categories refer to a variety of predefined standard sitting posture labels, such as upright sitting posture, forward-leaning sitting posture, backward-leaning sitting posture, left / right leaning, left / right leg pressing, hip forward movement, hunchback, etc. Convolutional neural networks (CNN) are deep neural networks that are specifically designed to process data with similar grid structures (such as images, pressure matrices) and are good at extracting local spatial features. Long short-term memory networks (LSTM) are a special type of recurrent neural network (RNN) that can learn long-term dependencies in sequence data and are good at capturing dynamic changes in the time dimension.

[0034] This embodiment utilizes a convolutional neural network to extract spatial geometric features from the single-frame pressure distribution, and then uses a long short-term memory network to analyze the temporal evolution logic between consecutive frames. Through end-to-end fusion training, it directly establishes a mapping relationship from the original pressure data to the sitting posture category, thereby achieving high-precision real-time recognition. The aim is to use the high-dimensional feature abstraction capability of hybrid deep learning models to solve the problem that traditional algorithms are difficult to handle complex nonlinear mapping relationships.

[0035] In one possible implementation, the end-to-end fusion configuration includes using the final output of the convolutional neural network as the input sequence of the long short-term memory network, or directly applying temporal convolution on the feature map of the convolutional neural network to form a spatiotemporal joint modeling network, and the hybrid model can use the backpropagation algorithm to jointly optimize the parameters of the convolutional layer and the recurrent layer during training.

[0036] In one specific implementation, the system inputs 30 consecutive frames of two-dimensional pressure matrix data into the model. First, it extracts the high-pressure region contour features of each frame through a convolutional neural network. Then, it inputs these feature sequences into a long short-term memory network. The network analyzes the trend of the high-pressure region gradually shifting from "symmetrical distribution" to "left side". Finally, it outputs the sitting posture category "the user is currently crossing their legs on the left side" through the Softmax layer.

[0037] Step S30: Based on the sitting posture category, execute a preset health management strategy.

[0038] It should be noted that, in this embodiment, the health management strategy refers to a series of preset rules or interventions triggered based on the identified sitting posture category, which aim to improve the user's sitting posture habits and prevent health hazards; the system calls the corresponding operation from the strategy library and implements it based on the current identification result.

[0039] This embodiment determines whether the sitting posture conforms to ergonomic standards. When poor sitting posture or prolonged sitting behavior is detected, the corresponding feedback mechanism is activated in a timely manner to guide the user to adjust their posture and reduce the risk of spinal and muscle fatigue. The aim is to transform the identified passive data into proactive health services.

[0040] In one possible implementation, health management strategies include, but are not limited to: providing real-time reminders via voice broadcasts or seat cushion vibrations, pushing corrective movement guidelines in a companion mobile application, recording users' posture health records and generating weekly reports, or controlling the lumbar support and backrest angle of the electric seat for automatic assisted adjustment, etc.

[0041] For example, when the output sitting posture category is "excessive backward leaning" and the duration exceeds 5 minutes, the system will determine that the posture is an unhealthy sitting posture with excessive pressure on the lumbar spine. It will then send a vibration reminder signal to the user's smart bracelet via Bluetooth and pop up a pop-up window on the phone screen with the text "Please pull your waist back and keep the ischial tuberosities vertically stressed" to help the user correct their sitting posture in time.

[0042] This application combines the spatial feature extraction capability of convolutional neural networks with the temporal analysis capability of long short-term memory networks to achieve in-depth mining of pressure distribution data, thereby significantly improving the accuracy and scene adaptability of posture recognition. This enables precise health management strategies to be implemented for different posture states, effectively assisting users in improving their posture habits and preventing health risks caused by prolonged sitting.

[0043] Furthermore, referring to Figure 2 The second embodiment of the intelligent posture recognition method of this application provides a flowchart, based on the above. Figure 2 The illustrated embodiment, "acquiring real-time pressure distribution data of the user's buttocks and thighs," includes adaptive coordinate correction, specifically steps A201-A204: Step A201: Detect the initial pressure distribution when the user first sits down and identify the pressure center points on the left and right sides of the buttocks. Step A202: Calculate the offset between the pressure center point and the geometric center of the flexible pressure sensor array; Step A203: Generate a coordinate transformation matrix based on the offset; Step A204: The acquired real-time pressure distribution data is dynamically corrected using the coordinate transformation matrix to obtain real-time pressure distribution data.

[0044] It should be noted that in this embodiment, adaptive coordinate correction is a data preprocessing mechanism designed to address the problem of relative shifts in the coordinate system of input data caused by inconsistent user seating positions on the flexible pressure sensor array (e.g., to the left, right, or rear), thus affecting the accuracy of model recognition. Specifically, the initial pressure distribution state refers to the static pressure field distribution collected by the sensor array when the user first contacts the cushion and maintains a stable posture; the pressure center points on the left and right sides of the buttocks refer to the coordinate positions of the centers of gravity representing the main force-bearing areas on the left and right sides of the human body, determined through a pressure weighting algorithm or extreme value search; the geometric center of the flexible pressure sensor array refers to the absolute center coordinates of the physical layout of the sensor array, typically defined as the zero point of the array coordinate system; the offset refers to the displacement vector of the user's actual pressure center of gravity relative to the physical center of the sensor array; and the coordinate transformation matrix is ​​a mathematical parameter generated based on the offset, used to map the original sensor coordinates to a standardized coordinate system based on the human body.

[0045] This embodiment quickly locates the deviation between the user's center of gravity and the array center in the early stage of the user's sitting, constructs a geometric transformation relationship, and applies this relationship to correct the data stream in real time, ensuring that the feature data received by the model has spatial consistency; it eliminates the interference caused by the user's sitting posture deviation, so that no matter where the user sits on the cushion, the pressure distribution data input to the recognition model presents uniform spatial characteristics, thereby significantly improving the model's adaptability and robustness to different sitting positions.

[0046] In one possible implementation, the pressure center points on the left and right sides of the buttocks are identified by calculating the pressure torque of a local area, that is, by performing a pressure-weighted average on the left and right halves of the sensor array to determine the coordinates of the center of gravity; when calculating the offset, the displacement components in the horizontal and vertical directions are calculated separately; the coordinate transformation matrix includes a translation transformation matrix, or combines a rotation matrix to adapt to the body angle deflection caused by the user sitting sideways; in addition, the correction process is re-initialized and calculated each time the user's new sitting action is triggered, so as to adapt to the usage habits of different users or the same user at different times.

[0047] Furthermore, the construction steps of the CNN-LSTM hybrid neural network model include A301~A305: Step A301: Obtain pressure distribution sample data collected under different seat materials and / or different body types to construct the initial training dataset.

[0048] It should be noted that in this embodiment, different seat materials include sponge, mesh or hard material, and users of different body types cover groups with different weight distribution and skeletal structures.

[0049] This embodiment introduces rich data variations to avoid overfitting the model to specific scenarios or groups of people, ensuring that the model can maintain high accuracy in recognition when facing different users and usage environments in real applications.

[0050] Step A302: Preprocess the initial training dataset to generate standardized training samples containing the time-series dimension.

[0051] It should be noted that in this embodiment, preprocessing refers to the system's operations of cleaning, denoising, normalizing, and format conversion of the raw pressure data, aiming to eliminate environmental noise and sensor errors during the data acquisition process. Through techniques such as sliding window truncation, the discrete pressure distribution data is constructed into a serialized data structure arranged in chronological order, giving it both spatial and temporal dynamic characteristics.

[0052] This embodiment introduces a temporal dimension to capture dynamic information during the sitting posture change process, thereby significantly improving the model's ability to understand continuous actions and posture transitions.

[0053] Step A303: Construct an initial network architecture that includes convolutional layers, pooling layers, LSTM layers, fully connected layers, and softmax classification layers.

[0054] It should be noted that in this embodiment, the convolutional layer is used to extract local spatial features from the input data, such as the distribution of pressure peak regions; the pooling layer is used to downsample the feature map, reduce the amount of computation, and extract the main features; the LSTM layer is used to process time series data and capture the long-term dependence of pressure distribution over time; the fully connected layer is responsible for integrating and mapping the extracted high-dimensional features; and the softmax classification layer is used to output the probability distribution of each sitting posture category.

[0055] Step A304: Input the standardized training samples into the initial network architecture for forward propagation calculation to obtain the preset sitting posture category.

[0056] It should be noted that in this embodiment, the processed serialized data is processed layer by layer according to the network hierarchy, realizing the data flow from the input layer to the output layer. The preset sitting posture category refers to the predicted classification result output by the model after inference and judgment of the input samples under the current internal parameter weights.

[0057] Step A305: Calculate the loss function value between the preset sitting posture category and the true label of the sample, and use the backpropagation algorithm to iteratively update the network weight parameters according to the loss function value until the model converges, thus obtaining the pre-constructed CNN-LSTM hybrid neural network model.

[0058] It should be noted that in this embodiment, the loss function value is used to measure the degree of difference between the model's prediction and the actual facts; the smaller the loss value, the more accurate the model's prediction. Iterative update is an optimization process in which the system adjusts the connection weights and biases of each neuron in the network layer by layer from the output layer to the input layer based on the calculated loss value gradient, in order to gradually reduce the prediction error. Model convergence refers to the loss function value stabilizing or reaching a preset accuracy threshold.

[0059] This application trains the model by introducing multi-scenario sample data and a complex spatiotemporal feature fusion network architecture, enabling the model to fully learn the pressure distribution characteristics and dynamic changes of different users under different seating conditions, significantly improving the model's generalization ability and recognition accuracy of complex sitting posture patterns.

[0060] Furthermore, the step of acquiring real-time pressure distribution data of the user's buttocks and thighs includes steps A401-A402: Step A401: Construct a two-dimensional pressure matrix from the real-time pressure distribution data based on the physical coordinates of the flexible sensor array.

[0061] It should be noted that in this embodiment, physical coordinates are the specific position information of each sensing unit in the flexible sensor array in a predefined coordinate system, usually represented in the form of (x, y); the two-dimensional pressure matrix is ​​a matrix data structure formed by mapping discrete sensor values ​​according to their physical coordinate positions into a two-dimensional grid structure, wherein the row index and column index of the matrix correspond to the horizontal and vertical coordinates of the physical coordinates, and the element values ​​in the matrix correspond to the pressure values ​​at the corresponding positions.

[0062] In one possible implementation, if some sensing units in the flexible sensor array are damaged or data is missing, the value at the missing location can be estimated using the pressure values ​​of the surrounding effective sensing units through an interpolation algorithm, thereby completing the two-dimensional pressure matrix.

[0063] Step A402: Normalize the two-dimensional pressure matrix and use the sliding window technique to extract the two-dimensional pressure matrix of continuous time frames to construct a three-dimensional tensor sample containing the temporal dimension.

[0064] It should be noted that in this embodiment, the sliding window technique involves sliding the window sequentially on the data stream according to a preset time step to extract continuous data segments; continuous time frames refer to a series of two-dimensional pressure matrices arranged in chronological order, with each frame representing the pressure distribution state at a certain moment; three-dimensional tensor samples refer to multi-dimensional data structures containing three dimensions: height, width, and time, specifically manifested as data blocks formed by stacking multiple continuous two-dimensional pressure matrices in the time dimension.

[0065] This embodiment is used to convert static spatial image data into sequence data containing temporal dynamic information, which not only preserves the spatial characteristics at each moment, but also reflects the evolution of pressure distribution over time.

[0066] In one possible implementation, the normalization process can employ the maximum-minimum normalization method; the length of the sliding window and the sliding step size can be flexibly set according to the real-time requirements of the actual application.

[0067] In another possible implementation, the dynamic window adjustment is achieved by: calculating the inter-frame difference value between the two-dimensional pressure matrices of two adjacent frames; comparing the inter-frame difference value with a preset dynamic change threshold; when the inter-frame difference value is greater than the dynamic change threshold, it is determined to be a period of drastic attitude change, and the time step of the sliding window is shortened to improve the temporal resolution; when the inter-frame difference value is less than the dynamic change threshold, it is determined to be a period of stable attitude, and the time step of the sliding window is extended to improve the stability of the recognition result; wherein, the adjusted sliding window is used to capture continuous time frame data.

[0068] In one specific implementation, the system first divides all values ​​in the 32x32 two-dimensional pressure matrix by the sensor's maximum range of 255 to normalize them to the 0-1 range. Then, a sliding window of length 10 is set to extract data from frame 1 to frame 10 to construct a three-dimensional tensor sample of shape 10x32x32. As time progresses, the window slides to frame 2 to frame 11 to construct the next three-dimensional tensor sample.

[0069] This application transforms discrete sensor data into a standardized input format with clear spatial topology and temporal dynamic information by constructing a two-dimensional matrix and a three-dimensional tensor. This not only eliminates the interference of data magnitude differences on model training, but also deeply preserves the characteristics of pressure distribution evolution over time, effectively improving the recognition accuracy of complex dynamic sitting postures.

[0070] Furthermore, referring to Figure 3 The third embodiment of the intelligent posture recognition method of this application provides a flowchart, based on the above. Figure 3 The illustrated embodiment further refines the step of "the processing of a convolutional neural network (CNN) includes spatial feature extraction," including steps A501 to A503: Step A501: Receive the three-dimensional tensor sample and extract the local spatial features of the pressure distribution through multi-layer convolution operations.

[0071] It should be noted that in this embodiment, the multi-layer convolution operation is a computational process that uses multiple convolution kernels to perform sliding window convolution operations and nonlinear activation processing on the input data to extract deep features layer by layer; local spatial features are attributes that reflect the details of pressure distribution, such as edges, textures and locally clustered spots where pressure gradients change.

[0072] Step A502 involves downsampling through a pooling layer to preserve local pressure extremes and suppress background noise.

[0073] It should be noted that in this embodiment, pooling layer downsampling is a dimensionality reduction process for the feature map, which aims to reduce the amount of data and the number of parameters; the local pressure extreme point is the position where the pressure value is the largest in the local neighborhood, which usually corresponds to the ischial tuberosity or the main weight-bearing area of ​​the thigh.

[0074] Step A503: Output spatial feature maps of each frame containing the coordinates of the pressure center and the shape information of the high-pressure area, forming a spatial feature sequence.

[0075] It should be noted that in this embodiment, the pressure center coordinates refer to the spatial location of the centroid of the entire pressure distribution calculated based on pressure weighting; the high-pressure region shape refers to the geometric contour formed by the region where the pressure value exceeds a specific threshold; and the spatial feature sequence refers to the set of feature maps arranged in chronological order that can characterize the spatial attributes at each moment.

[0076] This embodiment utilizes multi-layer convolution operations to mine local details and deep abstract features in the pressure distribution. Subsequently, it uses pooling layers to downsample and remove redundant information, reducing computational complexity. At the same time, it highlights key pressure extreme points and suppresses background noise. Finally, it outputs spatial feature maps of each frame containing the coordinates of the pressure center and the shape information of the high-pressure area, thereby forming a feature sequence that accurately reflects the changes in the spatial morphology of the sitting posture. This transforms the original high-dimensional three-dimensional tensor sample into a compact and semantically rich spatial feature sequence.

[0077] In one possible implementation, multi-layer convolution operations can employ different kernel sizes and strides, such as combining 1x1 and 3x3 kernels, which can capture local spatial features at different scales while increasing network depth; the pooling layer downsampling can specifically employ a max pooling strategy, which effectively preserves the most significant stress features by selecting the maximum value in a local region.

[0078] Furthermore, referring to Figure 4 The fourth embodiment of the intelligent posture recognition method of this application provides a flowchart, based on the above. Figure 4 The illustrated embodiment further refines the steps of "the processing of the Long Short-Term Memory (LSTM) network, including temporal dynamic feature extraction," including steps A601 to A603: Step A601: Input the spatial feature sequence into the LSTM network step by step.

[0079] It should be noted that in this embodiment, each frame feature map corresponds to the spatial features of pressure distribution at a specific moment. A time step refers to each discrete time point in the time series.

[0080] Step A602: Through the forget gate, input gate and output gate mechanism inside the LSTM network, the dynamic change law and long-term dependency relationship between the pressure distribution of consecutive frames are captured, and the temporal context feature vector containing attitude transition information is output.

[0081] It should be noted that, in this embodiment, the dynamic change pattern refers to the evolution of pressure distribution caused by the user's sitting posture over time, such as the movement trajectory of the center of gravity or the increase or decrease of the force area; the long-term dependency relationship refers to the correlation between the current sitting posture state and the state at a more distant time; the posture transition information describes the dynamic characteristics of the user's transition from one sitting posture to another; the temporal context feature vector refers to a high-dimensional feature vector that integrates information from historical moments and the current moment, representing the spatiotemporal context of the current moment.

[0082] In addition, the forget gate, input gate, and output gate refer to the core units inside the LSTM network used to control the flow of information. The forget gate is used to decide which historical information to discard, the input gate is used to decide which new information to store in the cell state, and the output gate is used to decide what value to output based on the current cell state.

[0083] In one possible implementation, a first weight vector is calculated using a forget gate, and irrelevant redundant features are removed from the cell state of the previous time step based on the correlation between the current spatial feature map and the historical pressure distribution. A second weight vector and candidate cell states are calculated using an input gate to determine the effective pressure change information that needs to be retained in the spatial feature map input at the current time step. The cell state at the current time step is updated based on the first weight vector, the second weight vector, and the candidate cell states to accumulate continuous features of posture evolution in the time dimension. A third weight vector is calculated using an output gate, and the hidden layer state at the current time step is filtered and output based on the updated cell state to form the temporal context feature vector containing posture transition information.

[0084] Step A603: Input the temporal context feature vector into the fully connected layer for feature fusion mapping, and then process it through the Softmax classifier to output the sitting posture category of the user at the current time.

[0085] It should be noted that, in this embodiment, the fully connected layer refers to the layer in the neural network that performs linear transformation and nonlinear activation on the input features to integrate the features; the feature fusion mapping is the transformation process of converting the temporal context feature vector to the classification label space; the Softmax classifier refers to the normalized exponential function that converts the output of the fully connected layer into a probability distribution, which is used to calculate the confidence of each category.

[0086] This embodiment inputs spatial feature sequences sequentially into an LSTM network, then utilizes the LSTM's gating mechanism to transmit and filter information in the temporal dimension, capturing the dynamic changes and long-term dependencies between pressure distributions in consecutive frames. It outputs a temporal context feature vector containing posture transition information. Finally, the feature vector is input into a fully connected layer for deep feature fusion mapping, and then processed by a Softmax classifier to output the user's current posture category. This allows for the extraction of dynamic features with temporal logic from static spatial features, enabling accurate understanding of continuous actions and posture changes.

[0087] In one specific implementation, the system inputs a spatial feature sequence containing 30 frames of feature maps into an LSTM network step by step. The LSTM network ignores the non-steady-state disturbances when the user adjusts their sitting position in the first few seconds through a forget gate, records the trend of the pressure center gradually shifting from the center to the left through an input gate, and outputs a temporal context feature vector reflecting this continuous change process through an output gate at the current moment. Subsequently, a fully connected layer maps this vector into a feature score corresponding to a preset sitting posture category. The Softmax classifier calculates that the "left tilt" category has the highest probability based on the score, and finally outputs the sitting posture category in which the user is currently tilted to the left.

[0088] Furthermore, the method also includes lightweight deployment of the CNN-LSTM hybrid neural network model, specifically including steps A701~A704: Step A701: Obtain the trained CNN-LSTM hybrid neural network model.

[0089] It should be noted that, in this embodiment, the trained CNN-LSTM hybrid neural network model refers to a deep learning model loaded from the training environment that has been trained with a large amount of sample data, whose network weight parameters have converged and whose verification accuracy has reached a preset standard. It includes a complete network topology and high-precision floating-point parameters that accurately identify sitting posture.

[0090] Step A702: Perform structural pruning on the CNN-LSTM hybrid neural network model, remove redundant convolutional kernel channels whose absolute weight values ​​are lower than a preset threshold, and output a simplified sparse network model.

[0091] It should be noted that in this embodiment, structural pruning is a model compression technique that reduces model size by removing redundant connections or neurons in a neural network; the preset threshold is a pre-set numerical limit used to determine whether a convolutional kernel is important; and the sparse network model is a simplified model structure with sparser network connections and a significantly reduced number of parameters after pruning. This step is used to reduce the computational complexity and storage space requirements of the model by eliminating invalid feature channels, making the model more concise.

[0092] Step A703: Perform integer quantization on the sparse network model to convert the model weight parameters and intermediate activation values ​​from floating-point format to fixed-point integer format, and output the quantized model file.

[0093] It should be noted that, in this embodiment, integer quantization is a compression process that reduces the representation precision of data in the model from high-precision floating-point numbers to low-precision integers; intermediate activation values ​​are the output data generated by each layer of the neural network during the operation; and quantized model file refers to the model storage file after data type conversion.

[0094] This step is used to further reduce the model size and reduce hardware resource consumption. It leverages the advantages of higher efficiency and lower power consumption of integer operations on embedded chips to enable the model to run in real time on edge devices, while maintaining the model's accuracy as much as possible through a reasonable quantization strategy.

[0095] Step A704: Compile the quantization model file to generate an inference engine adapted to the embedded edge computing chip, and deploy the inference engine into the microcontroller unit of the seat cushion.

[0096] It should be noted that in this embodiment, the embedded edge computing chip is a processor integrated inside the smart cushion, possessing low power consumption and certain AI computing capabilities; the microcontroller unit refers to the main control chip of the smart cushion. This step is used to successfully port the pruned and quantized optimized algorithm model to resource-constrained terminal hardware, realizing the localized deployment of posture recognition. This ensures that the system can complete real-time posture recognition tasks with low latency and high efficiency even offline, thereby reducing dependence on cloud services and improving system response speed and data privacy security.

[0097] This application significantly reduces the parameter size and computational complexity of the model through structural pruning and integer quantization, solving the problem that complex deep learning models are difficult to deploy on resource-constrained embedded devices. This enables the seat cushion microcontroller unit to run AI inference locally and efficiently, thereby greatly improving the system's response speed, reducing power consumption, and enhancing data privacy protection capabilities.

[0098] Furthermore, the steps for implementing the preset health management strategy include A801~A804: Step A801: Pre-build a library of corrective actions corresponding to various poor sitting postures; Step A802: Receive the sitting posture category and retrieve the target corrective action that matches the current poor sitting posture from the corrective action library; Step A803: Obtain the user's historical body data, and adaptively adjust the difficulty level of the target correction movement based on the historical body data; Step A804: Generate guidance information containing the key points of the adjusted actions.

[0099] It should be noted that, in this embodiment, the corrective action library refers to a database pre-established and stored by the system, which includes physical corrective exercises or stretching movements specifically designed for various types of poor sitting postures (such as hunching, crossing legs, and tilting to the side). Poor sitting postures refer to sitting postures identified by the model that are not ergonomic and may lead to spinal diseases or poor blood circulation if maintained for a long time. Target corrective actions refer to the most targeted corrective solutions that the system accurately matches from the corrective action library based on the currently identified poor sitting posture category. Adaptive adjustment is the process by which the system personalizes the intensity, amplitude, frequency, or duration of standard corrective actions based on the user's physical tolerance.

[0100] This application integrates posture recognition results with users' historical body data to achieve personalized and precise health management strategies. It can not only provide effective correction guidance for specific poor postures, but also dynamically adjust the difficulty of movements according to the user's physical condition, thereby avoiding the risk of physical injury caused by inappropriate exercise intensity and significantly improving the safety, scientific nature and user compliance of health interventions.

[0101] In one possible implementation, the corrective movement library can be updated regularly via a cloud server to incorporate the latest rehabilitation training movements or expert advice; adaptive adjustments to the target corrective movements can be dynamically fine-tuned in conjunction with the user's currently collected physiological indicators (such as heart rate and electromyography signals); in addition, guidance information supports multimodal output, for example, while displaying a diagram of the movement on the screen, the specific vibration frequency of the vibration motor in the cushion can prompt the user on the part of the force exerted or the rhythm.

[0102] In another possible implementation, the health management strategy is deeply linked to the user's sedentary time, forming a periodic intervention mechanism. For example, the system records the user's continuous sitting time, and when the cumulative time exceeds a preset threshold (such as 45 minutes), the health management strategy is forcibly triggered to remind the user to stand up and move around, regardless of whether the current sitting posture is standard.

[0103] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the intelligent posture recognition method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0104] This application also provides a sitting posture intelligent recognition system, please refer to... Figure 5 The intelligent posture recognition system includes: The data acquisition module 10 is used to acquire real-time pressure distribution data of the user's buttocks and thighs through a flexible pressure sensor array with embedded coordinates; The posture recognition module 20 is used to transmit the real-time pressure distribution data to a pre-built CNN-LSTM hybrid neural network model for feature extraction and classification, and output the posture category at the current moment; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network CNN and a long short-term memory network LSTM. The health management module 30 is used to execute a preset health management strategy based on the sitting posture category.

[0105] The intelligent posture recognition system provided in this application, employing the intelligent posture recognition method described in the above embodiments, can solve the technical problem that existing technologies struggle to capture the spatiotemporal dynamic characteristics of pressure data, resulting in low recognition accuracy and failing to meet the requirements for high-precision recognition. Compared with the prior art, the beneficial effects of the intelligent posture recognition system provided in this application are the same as those of the intelligent posture recognition method provided in the above embodiments, and other technical features of the intelligent posture recognition system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0106] This application provides a posture intelligent recognition device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the posture intelligent recognition method in the first embodiment described above.

[0107] The following is for reference. Figure 6The diagram illustrates a structural schematic suitable for implementing the intelligent posture recognition device of this application. The intelligent posture recognition device in this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The illustrated intelligent posture recognition device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0108] like Figure 6 As shown, the posture recognition device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the posture recognition device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the posture recognition device to communicate wirelessly or wiredly with other devices to exchange data. Although posture recognition devices with various systems are shown in the figures, it should be understood that implementing or having all of the systems shown is not required. More or fewer systems may be implemented alternatively.

[0109] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0110] The intelligent posture recognition device provided in this application, employing the intelligent posture recognition method described in the above embodiments, can solve the technical problem that existing technologies struggle to capture the spatiotemporal dynamic characteristics of pressure data, resulting in low recognition accuracy and an inability to meet high-precision recognition requirements. Compared with existing technologies, the beneficial effects of the intelligent posture recognition device provided in this application are the same as those of the intelligent posture recognition method provided in the above embodiments, and other technical features of this intelligent posture recognition device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0111] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent recognition of sitting posture, characterized in that, The method includes: Real-time pressure distribution data of the user's buttocks and thighs is obtained through a flexible pressure sensor array with embedded coordinates. The real-time pressure distribution data is transmitted to a pre-constructed CNN-LSTM hybrid neural network model for feature extraction and classification, outputting the current sitting posture category; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM). Based on the sitting posture category, a preset health management strategy is executed.

2. The intelligent posture recognition method as described in claim 1, characterized in that, The acquisition of real-time pressure distribution data in the user's buttocks and thighs includes adaptive coordinate correction, specifically: Detect the initial pressure distribution when the user first sits down and identify the pressure center points on the left and right sides of the buttocks; Calculate the offset between the pressure center point and the geometric center of the flexible pressure sensor array; Generate a coordinate transformation matrix based on the offset; The acquired real-time pressure distribution data is dynamically corrected using the coordinate transformation matrix to obtain real-time pressure distribution data.

3. The intelligent posture recognition method as described in claim 1, characterized in that, The construction steps of the CNN-LSTM hybrid neural network model include: Acquire pressure distribution sample data under different seat materials and / or different body types to construct an initial training dataset; The initial training dataset is preprocessed to generate standardized training samples that include the time-series dimension; Construct an initial network architecture that includes convolutional layers, pooling layers, LSTM layers, fully connected layers, and a softmax classification layer; The standardized training samples are input into the initial network architecture for forward propagation calculation to obtain the preset sitting posture category; Calculate the loss function value between the preset sitting posture category and the true label of the sample, and use the backpropagation algorithm to iteratively update the network weight parameters according to the loss function value until the model converges, thus obtaining the pre-constructed CNN-LSTM hybrid neural network model.

4. The intelligent posture recognition method as described in claim 1, characterized in that, Following the step of acquiring real-time pressure distribution data for the user's buttocks and thighs, the following is included: The physical coordinates of the flexible sensor array are used to construct a two-dimensional pressure matrix from the real-time pressure distribution data; The two-dimensional pressure matrix is ​​normalized, and the two-dimensional pressure matrix of continuous time frames is extracted using the sliding window technique to construct a three-dimensional tensor sample containing the temporal dimension.

5. The intelligent posture recognition method as described in claim 4, characterized in that, The processing procedure of the Convolutional Neural Network (CNN) includes spatial feature extraction, specifically including: The three-dimensional tensor sample is received, and the local spatial features of the pressure distribution are extracted through multi-layer convolution operations; Downsampling using pooling layers preserves local stress extrema and suppresses background noise; The output includes spatial feature maps of each frame containing the coordinates of the pressure center and the shape information of the high-pressure area, forming a spatial feature sequence.

6. The intelligent posture recognition method as described in claim 5, characterized in that, The processing procedure of the Long Short-Term Memory (LSTM) network includes temporal dynamic feature extraction, specifically including: The spatial feature sequence is input into the LSTM network step by step. By using the forget gate, input gate, and output gate mechanisms within the LSTM network, the dynamic variation patterns and long-term dependencies between the pressure distributions of consecutive frames are captured, and a temporal context feature vector containing attitude transition information is output. The temporal context feature vector is input into a fully connected layer for feature fusion mapping, and then processed by a Softmax classifier to output the user's sitting posture category at the current time.

7. The intelligent posture recognition method as described in claim 1, characterized in that, The method also includes lightweight deployment of CNN-LSTM hybrid neural network models, specifically including: Obtain the trained CNN-LSTM hybrid neural network model; The CNN-LSTM hybrid neural network model is structurally pruned to remove redundant convolutional kernel channels whose absolute weights are below a preset threshold, and a simplified sparse network model is output. Perform integer quantization on the sparse network model to convert the model weight parameters and intermediate activation values ​​from floating-point format to fixed-point integer format, and output the quantized model file; The quantization model file is compiled to generate an inference engine adapted to the embedded edge computing chip, and the inference engine is deployed to the microcontroller unit of the seat cushion.

8. The intelligent posture recognition method as described in claim 1, characterized in that, The steps for implementing the preset health management strategy include: Pre-build a library of corrective actions for various poor sitting postures; Receive the sitting posture category and retrieve a target corrective action that matches the current poor sitting posture from the corrective action library; Obtain the user's historical body data, and adaptively adjust the difficulty level of the target corrective action based on the historical body data; Generate guidance information that includes key points of the adjusted actions.

9. A sitting posture intelligent recognition system, characterized in that, The intelligent posture recognition system includes: The data acquisition module is used to acquire real-time pressure distribution data of the user's buttocks and thighs through a flexible pressure sensor array with embedded coordinates; The posture recognition module is used to transmit the real-time pressure distribution data to a pre-built CNN-LSTM hybrid neural network model for feature extraction and classification, and output the posture category at the current moment; the CNN-LSTM hybrid neural network model is composed of an end-to-end fusion of a convolutional neural network (CNN) and a long short-term memory network (LSTM). The health management module is used to execute preset health management strategies based on the sitting posture category.

10. A sitting posture intelligent recognition device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the sitting posture intelligent recognition method as described in any one of claims 1 to 8.

Images