A bed human posture recognition method based on multi-level label collaborative perception

Abstract

The application discloses a bed human posture recognition method based on multi-level label collaborative sensing, and belongs to the technical field of radio frequency identification, and comprises the following steps: acquiring RFID phase data to be recognized; segmenting the RFID phase data to be recognized based on a time window to obtain a phase array; performing phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix; classifying the unwrapped phase matrix based on distance to obtain RFID labels, including barrier labels, near neighbor labels and far end labels; selecting a preset number of far end labels as a reference label set; performing phase data alignment operation on all the unwrapped phase arrays to obtain an input phase matrix; and performing classification based on a trained two-stage neural network to obtain a bed human posture recognition result. The application solves the problems that the prior art cannot recognize complex limb postures on a bed and it is difficult to distinguish supine and prone postures when a user raises arms or approaches the head.

Description

TECHNICAL FIELD

[0001] The application relates to a bed human posture recognition method based on multi-level tag collaborative sensing, and belongs to the technical field of radio frequency identification. BACKGROUND

[0002] Long-term monitoring of sleep posture has important value for health assessment and disease diagnosis. The field of intelligent health monitoring has long been committed to realizing accurate sleep posture recognition under non-invasive conditions. Current mainstream technical paths include three types of schemes of video visual detection, pressure sensor array and radio frequency sensing. Video monitoring collects image sequences through a camera, combines human key point detection and posture estimation algorithms to realize sleep posture classification. Although the accuracy is high, it is difficult to be deployed in the bedroom scene for a long time due to the limitations of light conditions, shielding problems and privacy protection requirements. The pressure sensor array analyzes the pressure distribution pattern through a high-density sensor laid on the surface of the mattress, which can directly reflect the body contact area, but has defects such as high cost, complex wiring and poor long-term stability. In recent years, the radio frequency identification (RFID) technology has shown its advantages in sleep posture recognition due to its passive, non-contact and low-power characteristics. It realizes human existence state and posture recognition by deploying a tag matrix to collect receive signal strength indicator (RSSI) and phase change, but traditional methods mostly regard the tag as a single information source, lacking the modeling ability of the spatial level difference of the tag. Existing technologies attempt to realize high-precision classification through binary data processing or distinguish between prone and supine through head tag mode analysis, but both have significant limitations.

[0003] The prior art has three core defects, which directly lead to insufficient complex limb posture recognition ability and difficulty in distinguishing supine and prone positions. Firstly, the traditional method assumes that the human limbs are in a standard position and does not consider the dynamic changes of the arms in real sleep, such as the hands close to the body, naturally flat or raised above the shoulders. When the user's arms are raised or close to the head, the classification performance decreases significantly because the model cannot distinguish the phase disturbance caused by arm obstruction from the subject posture change. Secondly, the existing technology needs to pre-acquire a reference image of an empty bed state, but in actual deployment, the reference may fail due to changes in the antenna and bed position, and there is a lack of stable reference alignment mechanism to deal with RFID signal phase drift and ambiguity. Thirdly, the existing method does not establish a "tag space level" classification and modeling mechanism. The tag signal far from the human body is stable but the human body information is limited, the tag close to the human body contains local detail information affected by scattering, and the completely blocked tag shows contour features. The three need to be analyzed cooperatively to accurately extract key regional features. In addition, the mainstream dataset does not cover the difference in complex limb placement, which limits the generalization ability of the model and makes it unable to adapt to the diverse limb posture changes in real sleep scenarios. These defects together limit the accuracy and robustness of the existing technology in complex limb posture recognition and supine / prone position distinguishing scenarios. SUMMARY

[0004] The purpose of the present application is to provide a bed human posture recognition method based on multi-level tag collaborative perception, which classifies RFID tags based on distance, selects remote tags as a reference tag set to align phase data, uses a trained two-stage neural network for classification, and obtains bed human posture recognition results, to solve the problem that the existing technology cannot recognize complex limb posture on the bed and it is difficult to distinguish supine and prone positions when the user's arms are raised or close to the head.

[0005] To solve the above technical problems, the present application is realized by using the following technical solutions.

[0006] The present application provides a bed human posture recognition method based on multi-level tag collaborative perception, comprising:

[0007] Obtain RFID phase data to be identified;

[0008] Segment the RFID phase data to be identified based on a time window to obtain a phase array;

[0009] Perform phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix;

[0010] Classify RFID tags based on distance on the unwrapped phase matrix, the RFID tags including blocked tags, near neighbor tags and remote tags;

[0011] Select a preset number of remote labels as a reference label set, perform a phase data alignment operation on all unwound phase arrays to obtain an input phase matrix;

[0012] According to the input phase matrix, classification is performed based on the trained two-stage neural network to obtain a bed human posture recognition result;

[0013] The bed human posture recognition result includes one of supine and prone, and one of left flexion, right flexion, left support lateral, right support lateral, left induced lateral, and right induced lateral.

[0014] Further, the network structure of the two-stage neural network includes a main network and a branch network;

[0015] The main network is used to classify according to the input phase matrix, and output an initial bed human posture recognition result, which includes prone and supine combined lying, left flexion, right flexion, left support lateral, right support lateral, left induced lateral, and right induced lateral.

[0016] The branch network is used to determine the input phase matrix corresponding to the waist region according to the spatial distribution range of the blocking label from the input phase matrix corresponding to the prone and supine combined lying according to the initial bed human posture recognition result, and identify and output the supine or prone posture result.

[0017] Further, the main network includes:

[0018] A first dual-channel input layer is used to receive the input phase matrix and a first mask matrix, wherein the first mask matrix is obtained by binarizing the input phase matrix: traversing the input phase matrix, if the element at the current position in the input phase matrix is a non-numeric value, then set the mask value of the corresponding position of the binary mask matrix to 1, otherwise set it to 0, and after traversal, generate a binary mask matrix with the same dimension as the input phase matrix as the first mask matrix.

[0019] A feature concatenation layer is used to concatenate the input phase matrix and the first mask matrix in the channel direction to obtain a concatenated feature;

[0020] A first feature extraction module is used to process the concatenated feature through a plurality of convolution layers and residual blocks to extract spatial features;

[0021] A first attention module is used to enhance the spatial features through a convolution block attention module to obtain enhanced features;

[0022] A first pooling layer is used to perform global average pooling on the enhanced features to extract a global feature vector;

[0023] a feature calculation module, configured to calculate a statistical feature vector based on the first mask matrix;

[0024] a first feature fusion layer, configured to concatenate the global feature vector and the statistical feature vector to obtain a first concatenated vector;

[0025] a first feedforward network, configured to perform nonlinear transformation and feature dimension reduction on the first concatenated vector to obtain a first high-level feature vector;

[0026] a first classification layer, configured to classify the first high-level feature vector to obtain and output an initial on-bed human posture recognition result;

[0027] The statistical feature vector includes: an area proportion of a region where the blocking label is located, a barycentric normalized coordinate of the region where the blocking label is located, a bounding box length-width ratio and a bounding box width-length ratio of the region where the blocking label is located, left and right distribution proportions, up and down distribution proportions, left and right differences, and up and down differences of a number of blocking labels in the region where the blocking label is located, and a phase mean value and a standard deviation of effective labels within a preset distance range of the blocking label.

[0028] Further, the branch network includes:

[0029] a second dual-channel input layer, configured to receive an input phase matrix corresponding to a pitch-merge-lie and extract an input phase matrix corresponding to a waist-torso region, and a second mask matrix, the second mask matrix being obtained by binarization processing of the input phase matrix corresponding to the waist-torso region: traversing the input phase matrix corresponding to the waist-torso region, if an element at a current position in the input phase matrix corresponding to the waist-torso region is a non-number value, setting a mask value of a corresponding position of a binary mask matrix to 1, otherwise setting it to 0, and generating a binary mask matrix with the same dimension as the input phase matrix as the second mask matrix after traversal is completed;

[0030] a second feature extraction module, configured to process the input phase matrix corresponding to the waist-torso region and the second mask matrix through a plurality of convolution layers and residual blocks to extract local sparse features;

[0031] a second attention module, configured to weight importance of each spatial position of the local sparse features through a convolution block attention module to obtain a weighting result;

[0032] a second pooling layer, configured to perform spatial attention pooling on the local sparse features according to the weighting result to obtain a weighted average feature vector;

[0033] a second feature fusion layer, configured to integrate all weighted average feature vectors to obtain integrated features;

[0034] The second feedforward network and the linear classification layer are used for mapping and classifying the integrated features, and outputting a binary classification result of the supine posture or the prone posture.

[0035] Further, the two-stage neural network is trained by using an expanded data set to obtain a body posture recognition result of a person on a bed according to an input phase matrix; wherein the method for obtaining the expanded data set comprises:

[0036] RFID phase data is collected by deploying an RFID tag array on a bed;

[0037] The RFID phase data is segmented based on a time window to obtain a phase array;

[0038] Phase unwrapping is performed on the phase array based on a sliding window to obtain an unwrapped phase matrix;

[0039] The unwrapped phase matrix is classified based on distance to obtain RFID tags, including blocked tags, near tags and far tags;

[0040] A preset number of far tags are selected as a reference tag set, and a phase data alignment operation is performed on all the unwrapped phase arrays to obtain an input phase matrix;

[0041] The input phase matrix is subjected to data enhancement to obtain an expanded data set.

[0042] Further, the RFID phase data is segmented based on a time window to obtain a phase array, comprising:

[0043] The RFID phase data is intercepted based on a preset time window length to obtain initial sample sequences corresponding to multiple time windows;

[0044] For each time window, the first phase observation value of each RFID tag in the current time window in the initial sample sequence is extracted to form an initial sample data reflecting the instantaneous state of the RFID tag array;

[0045] Each RFID tag in the initial sample data is classified according to a preset minimum reading number threshold:

[0046] If the total reading number of a certain RFID tag in the time window is greater than or equal to the preset minimum reading number threshold, the RFID tag is marked as a valid tag;

[0047] If the total reading number of a certain RFID tag in the time window is less than the preset minimum reading number threshold, the RFID tag is marked as a missing tag;

[0048] form a matrix of rows and columns from all valid tags and missing tags as a sample read tag array, wherein, row column represents a number of rows of the sample read tag array, represents a number of columns of the sample read tag array;

[0049] For each missing tag, check the RFID tags in eight directions of a position of the missing tag in the sample read tag array:

[0050] If the RFID tags in the eight directions are all marked as valid tags within a current time window, calculate an average of phase values of the RFID tags in the eight directions, fill in a phase value of the missing tag with the average, and update the missing tag as a valid tag;

[0051] If there is at least one RFID tag not marked as a valid tag among the RFID tags in the eight directions within the current time window, keep the missing tag as a missing tag;

[0052] fill in all the phase arrays marked as valid tags and missing tags as a final phase array;

[0053] The eight directions include directly above, right above, directly right, right below, directly below, left below, directly left, and left above.

[0054] Further, phase unwrapping is performed on the phase array based on a sliding window to obtain an unwrapped phase matrix, including:

[0055] Based on a preset size and sliding step of the sliding window, a sliding window is constructed on the phase array along a time dimension; for the phase array in each sliding window:

[0056] The difference between the phase value at each time and the phase value at the previous time is calculated in sequence;

[0057] If the absolute value of the difference exceeds a phase jump threshold π, it is determined that a phase jump occurs, and phase jump correction is performed by adding or subtracting an integer multiple of 2π from the phase value at the current time, so that the absolute value of the phase difference after phase jump correction is less than π;

[0058] After the phase jump correction of the phase array in all sliding windows is completed, the unwrapped phase matrix is obtained.

[0059] Further, the unwrapped phase matrix is classified based on distance, including:

[0060] The unwrapped phase matrix is converted into a category matrix extracting effective tags and missing tags, wherein 1 represents a blocked tag, 2 represents a near neighbor tag, and 3 represents a far end tag, representing a category matrix;

[0061] initializing the effective tags as far end tags;

[0062] initializing the missing tags as blocked tags;

[0063] traversing each blocked tag:

[0064] if all RFID tags within an eight-neighborhood of the blocked tag are far end tags, updating the blocked tag as a far end tag;

[0065] if at least one RFID tag within the eight-neighborhood of the blocked tag is not a far end tag, maintaining the blocked tag as a blocked tag;

[0066] performing eight-neighborhood connected region marking for each blocked tag;

[0067] if a number of tags of the connected region is less than a preset number of tags of connected region threshold, updating all blocked tags within the connected region as far end tags;

[0068] if the number of tags of the connected region is greater than or equal to the preset number of tags of connected region threshold, maintaining all blocked tags within the connected region as blocked tags;

[0069] searching for non-blocked tags within a preset Manhattan distance range for remaining blocked tags and updating as near neighbor tags.

[0070] Further, a preset number of far end tags are selected as a reference tag set, and a phase data alignment operation is performed on all unwound phase arrays to obtain an input phase matrix, including:

[0071] grouping all unwound phase arrays according to the same reference;

[0072] calculating a reading stability index value, a phase stability index value, and an average minimum distance away from a region where a blocked tag is located for each RFID tag in all unwound phase arrays in a group, and obtaining a comprehensive score based on weighted average;

[0073] a calculation formula of the reading stability index value is represented as:

[0074] ;

[0075] in the formula, represents a reading stability index value of an RFID tag in the ith row and the jth column in the group, is a number of unwound phase arrays in the group, the i-th row and the j-th column in the disentangled phase array of the group the i-th row and the j-th column in the disentangled phase array of the group the number of times the RFID tag in the i-th row and the j-th column is read;

[0076] The formula for calculating the phase stability index value is:

[0077] ;

[0078] In the formula, represents the phase stability index value of the RFID tag in the i-th row and the j-th column in the disentangled phase array of the group, represents a variance calculation function, represents the phase value of the RFID tag in the i-th row and the j-th column;

[0079] The formula for calculating the average minimum distance from the area where the blocking tag is located is:

[0080] ;

[0081] In the formula, represents the average minimum distance from the area where the blocking tag is located to the RFID tag in the i-th row and the j-th column in the disentangled phase array of the group, respectively represent the horizontal coordinate and the vertical coordinate of the blocking tag, represents the i-th disentangled phase array, and the position set of the area where the blocking tag is located.

[0082] Select multiple far-end tags with a comprehensive score higher than a preset comprehensive score threshold as a far-end reference tag set in the group ;

[0083] The formula for calculating the comprehensive score is:

[0084] ;

[0085] In the formula, represents the comprehensive score of the RFID tag in the i-th row and the j-th column in all disentangled phase arrays of the group, respectively represent the stability index value, the phase stability index value, and the weight coefficient of the average minimum distance from the area where the blocking tag is located of the RFID tag in the i-th row and the j-th column, represents a normalization operation;

[0086] Calculate the weighted average phase of the i-th disentangled phase array based on the far-end reference tag set, represented as:

[0087] ; ​

[0088] wherein, denotes the th unwrapped phase array, the weighted average phase calculated by the remote reference tag set denotes the weight, denotes the th unwrapped phase array, the phase value of the RFID tag in the th row and the th column, denotes the minimum value for preventing division by zero;

[0089] After the effective phase points of all the unwrapped phase arrays are weighted average phase, a phase unwrapping operation is performed again to obtain an input phase matrix;

[0090] wherein, the value of the near neighbor tag is set as the phase value of the input phase matrix, the value of the blocked tag is set as 0, and the remote tag is masked.

[0091] Further, the input phase matrix is subjected to data enhancement to obtain an expanded data set, including:

[0092] adding offset noise to the phase of the input phase matrix, and / or,

[0093] adding global Gaussian noise to the phase corresponding to the near neighbor tag, and / or,

[0094] randomly discarding a preset percentage of phase data corresponding to the RFID tag, and / or,

[0095] performing horizontal or vertical translation on the region where the blocked tag is located, and / or,

[0096] scaling the user body size according to the proportion of the number of blocked tags to determine a scaling factor to obtain the expanded data set:

[0097] If the proportion of the number of blocked tags is greater than a preset proportion threshold of the number of blocked tags, it is determined that the user is large in size, and the scaling factor is determined to have a value range less than 1.

[0098] If the proportion of the number of blocked tags is less than a preset proportion threshold of the number of blocked tags, it is determined that the user is small in size, and the scaling factor is determined to have a value range greater than 1.

[0099] Based on the value of the determined scaling factor, coordinate scaling is performed on the region composed of the blocked tags and the near neighbor tags.

[0100] Compared with the prior art, the present application has the following beneficial effects:

[0101] ​​1. The application effectively improves the accuracy of bed posture recognition result judgment by segmenting and unwrapping the RFID phase data based on time window and sliding window respectively, classifying the unwrapped phase matrix according to the distance, selecting the far-end tag for phase data alignment, and finally using the trained two-stage neural network classification.

[0102] 2. The main network of the two-stage neural network uses the input phase matrix and the first mask matrix generated by the blocking tag to perform operations such as splicing, feature extraction, and attention enhancement, and comprehensively analyzes human posture features from multiple dimensions based on statistical feature vectors, accurately identifies the initial bed human posture recognition result, further focuses on the waist region using the branch network, extracts local sparse features using the second mask matrix, and accurately identifies the supine / prone posture through spatial attention mechanism and pooling operation. Compared with traditional methods, the comprehensiveness and accuracy of bed human posture recognition results are greatly improved, and both complex limb postures and easily confused supine / prone postures can be accurately identified.

[0103] 3. The application obtains RFID phase data, and based on time window segmentation, the missing tags are filled to ensure the integrity of the phase array. Then, through sliding window phase unwrapping, the phase jump is accurately corrected, effectively solving the phase wrapping problem and obtaining the accurate phase matrix after unwrapping. In the RFID tag classification process, the blocking tag, near neighbor tag and far-end tag are accurately distinguished by eight-neighborhood connected region marking and Manhattan distance search, and the far-end tag with high comprehensive score is selected as the reference tag set for phase data alignment, ensuring the accuracy and stability of the input phase matrix and improving the reliability of the recognition result.

[0104] 4. The application uses an expanded data set to train a two-stage neural network, adds offset noise and global Gaussian noise to the input phase matrix, randomly discards part of the phase data corresponding to the RFID tag, translates the region where the blocking tag is located, and scales the user's body size according to the proportion of the number of blocking tags, etc. Simulate data changes in various actual scenarios, effectively increase the diversity and richness of the data, so that the model can learn more features in different scenarios, improve the generalization ability of the model. At the same time, combined with the expanded data set, the two-stage neural network is trained, which can make the model better adapt to various complex situations, further improve the accuracy and stability of the bed human posture recognition result judgment, and has higher practical value in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0105] Figure 1is a flowchart of a bed human posture recognition method based on multi-level label collaborative perception provided by an embodiment of the present application.

[0106] Figure 2 is a gray scale reading schematic diagram of an RFID tag array provided by an embodiment of the present application, wherein the black area represents that the RFID tag is not read, and the gray area represents that the RFID tag is read.

[0107] Figure 3 is a schematic diagram of a pretreated RFID tag array provided by an embodiment of the present application.

[0108] Figure 4 is a schematic diagram of an RFID phase array after label classification and phase data alignment provided by an embodiment of the present application.

[0109] Figure 5 is a schematic diagram of the overall structure of a two-stage neural network provided by an embodiment of the present application.

[0110] Figure 6 is a classification result comparison schematic diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0111] The technical solutions of the present application will be described in detail below with the aid of the accompanying drawings and specific embodiments. It should be understood that the specific features in the embodiments and the specific embodiments of the present application are detailed descriptions of the technical solutions of the present application, and are not limitations of the technical solutions of the present application. In the case of no conflict, the technical features in the embodiments and the specific embodiments can be combined with each other.

[0112] The term "and / or", only describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / ", generally represents that the front and rear associated objects are a "or" relationship.

[0113] Embodiment 1

[0114] As shown in Figure 1 , the present embodiment introduces a bed human posture recognition method based on multi-level label collaborative perception, which comprises:

[0115] Step 1: Obtain RFID phase data to be identified.

[0116] The present application provides original data basis for subsequent accurate phase analysis and processing by obtaining RFID phase data to be identified, which ensures that the whole recognition process has sufficient and accurate data source, and helps to improve the accuracy of the final human posture recognition.

[0117] Step two: segmenting the to-be-identified RFID phase data based on a time window to obtain a phase array.

[0118] The present application segments RFID phase data based on a time window, which can effectively convert continuous and complex phase data into an ordered phase array, facilitating subsequent targeted analysis of phase characteristics in different time periods and improving the efficiency and pertinence of data processing and analysis.

[0119] Step three: performing phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix.

[0120] The present application performs phase unwrapping on the phase array based on a sliding window, which can accurately eliminate the winding phenomenon in the phase data and convert the originally winding and complex phase data into a clear and ordered unwrapped phase matrix, providing reliable data support for subsequent accurate tag classification and posture recognition.

[0121] Step four: classifying RFID tags based on the unwrapped phase matrix, wherein the RFID tags include blocked tags, near neighbor tags, and far-end tags.

[0122] The present application classifies RFID tags based on the unwrapped phase matrix, which can clearly distinguish the types of tags in different distance ranges. This classification method helps to more carefully analyze the information reflected by tags in different positions, providing more abundant dimensions and more accurate basis for subsequent accurate recognition of human posture on the bed.

[0123] Step five: selecting a preset number of far-end tags as a reference tag set to perform phase data alignment on all unwrapped phase arrays to obtain an input phase matrix.

[0124] The present application selects a preset number of far-end tags as a reference tag set to align all unwrapped phase arrays, which can effectively eliminate the phase differences between different phase arrays caused by various factors, allowing phase data to be analyzed under a unified reference, improving the consistency and comparability of data, and further improving the accuracy of subsequent posture recognition results.

[0125] Step six: classifying based on the trained two-stage neural network according to the input phase matrix to obtain a human posture recognition result on the bed.

[0126] In the present embodiment, the human posture recognition result on the bed includes one of supine and prone, and one of left flexion, right flexion, left support lateral, right support lateral, left induced lateral, and right induced lateral.

[0127] The application utilizes a trained two-stage neural network to classify according to an input phase matrix, can fully exert the powerful feature extraction and classification capability of the neural network, and accurately identifies a bed human posture recognition result; the application does not need to wear equipment, realizes high-precision human posture recognition, and significantly enhances the generalization capability and environmental adaptability compared with the prior art.

[0128] Embodiment 2

[0129] Based on the same inventive concept as embodiment 1, this embodiment introduces an implementation step of a bed human posture recognition method based on multi-level label collaborative perception, which comprises:

[0130] Step 1: acquiring RFID phase data to be identified.

[0131] Step 2: segmenting the RFID phase data to be identified based on a time window to obtain a phase array.

[0132] Step 3: performing phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix.

[0133] Step 4: classifying the unwrapped phase matrix based on distance, wherein the RFID labels comprise blocking labels, near neighbor labels and far end labels.

[0134] In this embodiment, a gray scale reading schematic diagram of a single RFID label array is shown in Figure 2 , a schematic diagram of the RFID label array after removing isolated points and isolated areas is shown in Figure 3 , a schematic diagram of the RFID phase array after label classification and phase data alignment, i.e. an input phase array, is shown in Figure 4 . By comparison, it can be seen that the preprocessing method proposed in this embodiment can remove noise in the RFID label array data, including abnormal single labels or small label areas that are not read; through the label classification and phase data alignment step, near neighbor label phase data is added near the blocking area, and a higher-dimensional and more discriminative feature representation is constructed.

[0135] Step 5: selecting a preset number of far end labels as a reference label set, performing phase data alignment operation on all unwrapped phase arrays to obtain an input phase matrix.

[0136] Step 6: classifying based on a trained two-stage neural network according to the input phase matrix to obtain a bed human posture recognition result.

[0137] In this embodiment, the bed human posture recognition result comprises one of supine and prone and one of left curved lying, right curved lying, left support lateral lying, right support lateral lying, left induced lateral lying and right induced lateral lying.

[0138] In this embodiment, the network structure of the two-stage neural network includes a main network and branch networks, such as... Figure 5 As shown; the main network is used to classify according to the input phase matrix and output the initial human posture recognition result on the bed. The initial human posture recognition result on the bed includes supine and prone position, left flexed position, right flexed position, left supported side position, right supported side position, left induced side position, and right induced side position. The branch network is used to determine the input phase matrix corresponding to the waist and torso region from the input phase matrix corresponding to the supine and prone position according to the spatial distribution range of the blocking labels, based on the initial human posture recognition result on the bed, and to identify and output the supine or prone posture result.

[0139] In this embodiment, the main network includes: a first dual-channel input layer, used to receive the input phase matrix and a first mask matrix, wherein the first mask matrix is ​​obtained by binarizing the input phase matrix: traversing the input phase matrix, if the element at the current position in the input phase matrix is ​​non-numerical, the mask value at the corresponding position in the binary mask matrix is ​​set to 1, otherwise it is set to 0, and after traversal, a binary mask matrix with the same dimension as the input phase matrix is ​​generated as the first mask matrix; a feature concatenation layer, used to concatenate the input phase matrix and the first mask matrix in the channel direction to obtain concatenated features; and a first feature extraction module, used to process the concatenated features through multiple convolutional layers and residual blocks. The system processes features to extract spatial features; a first attention module enhances the spatial features using a convolutional block attention module to obtain enhanced features; a first pooling layer performs global average pooling on the enhanced features to extract a global feature vector; a feature calculation module calculates a statistical feature vector based on the first mask matrix; a first feature fusion layer concatenates the global feature vector with the statistical feature vector to obtain a first concatenated vector; a first feedforward network performs nonlinear transformation and feature dimensionality reduction on the first concatenated vector to obtain a first high-level feature vector; and a first classification layer classifies the first high-level feature vector to obtain and output the initial bed human pose recognition result. In this embodiment, the main network is trained using the cross-entropy loss function, parameters are updated using the Adam optimization algorithm, and the main network weights are frozen after training.

[0140] In the embodiment, the first feature extraction module includes a convolution network and three layers of residual blocks, the convolution network includes two layers of convolution operations: the first layer is a convolution layer with a convolution kernel size of 3*3, an input channel number of 2, an output channel number of 32, a stride of 1 and padding of 1, followed by a batch normalization operation and a ReLU activation function; the second layer is a convolution layer with a convolution kernel size of 3*3, an input channel number of 32, an output channel number of 64, a stride of 1 and padding of 1, followed by a batch normalization operation and a ReLU activation function, the convolution network extracts complex spatial features through three layers of residual block stacking: the first layer of residual block maps 64 channel features to 128 channel features, the second layer of residual block keeps 128 channel features unchanged, and the third layer of residual block maps 128 channel features to 256 channel features.

[0141] The statistical feature vector includes: an area proportion of the region where the blocking label is located, a barycenter normalized coordinate of the region where the blocking label is located, a boundary box length-width ratio and a boundary box width-length ratio of the region where the blocking label is located, left and right distribution proportions, up and down distribution proportions, left and right difference values and up and down difference values of the number of blocking labels in the region where the blocking label is located, and a phase mean value and a standard deviation of effective labels in a preset distance range of the blocking label.

[0142] In the embodiment, the branch network includes: a second double-channel input layer, configured to receive an input phase matrix corresponding to the supine or prone position and extract an input phase matrix corresponding to a waist and trunk region, and a second mask matrix, which is obtained by binarization processing of the input phase matrix corresponding to the waist and trunk region: traversing the input phase matrix corresponding to the waist and trunk region, if an element at a current position in the input phase matrix corresponding to the waist and trunk region is a non-numeric value, setting a mask value of a corresponding position of the binary mask matrix to 1, otherwise setting the mask value to 0, and generating a binary mask matrix with the same dimension as the input phase matrix as the second mask matrix after the traversal is completed; a second feature extraction module, configured to process the input phase matrix corresponding to the waist and trunk region and the second mask matrix through multiple layers of convolution layers and residual blocks to extract local sparse features; a second attention module, configured to weight the importance of each spatial position of the local sparse features through a convolution block attention module to obtain a weighting result; a second pooling layer, configured to perform spatial attention pooling on the local sparse features according to the weighting result to obtain a weighted average feature vector; a second feature fusion layer, configured to integrate all the weighted average feature vectors to obtain integrated features; and a second feedforward network and a linear classification layer, configured to map and classify the integrated features to output a binary classification result of the supine or prone position. In the embodiment, the branch network is trained by using a binary classification cross-entropy loss function and iteratively trained by using an Adam optimizer, and the weights of the main network remain in a frozen state and do not participate in gradient back propagation during the training process.

[0143] In this embodiment, the second feature extraction module includes a convolutional network and two layers of residual blocks, the convolutional network includes two layers of convolutional operations: the first layer is a convolutional layer with a convolution kernel size of 3x3, an input channel number of 2, an output channel number of 32, a stride of 1, and padding of 1, followed by a batch normalization operation and a ReLU activation function; the second layer is a convolutional layer with a convolution kernel size of 3x3, an input channel number of 32, an output channel number of 64, a stride of 1, and padding of 1, followed by a batch normalization operation and a ReLU activation function, and the convolutional network extracts complex spatial features through two layers of residual blocks: the first layer of residual blocks maps 64-channel features to 128-channel features, and the second layer of residual blocks maps 128-channel features to 256-channel features.

[0144] The two-stage neural network is trained by using an expanded data set to obtain a human posture recognition result on a bed according to an input phase matrix; wherein the method for obtaining the expanded data set comprises:

[0145] Step 6.1: Collect RFID phase data by deploying an RFID tag array on the bed.

[0146] In this embodiment, an RFID tag array containing a total of 459 tags is deployed on a bed that is 2 meters long and 1.5 meters wide, with an array specification of 27 tags per row and 17 tags per column. The tag model uses H47 passive tags produced by Impinj Company. The product electronic code (EPC) of each RFID tag is modified to a unique spatial coordinate number, and the RFID tag array is pasted on the specified position of the plastic film. The plastic film with the pasted RFID tag array is fixed on the mattress.

[0147] In this embodiment, a circularly polarized antenna with a frequency of 920 MHz and an antenna gain of 8dBi is used to collect the backscattering signals of the tag array. The antenna is deployed at a slant position 1 meter away from the bed and 2 meters high to collect RFID phase data. In this embodiment, an R420 reader / writer of Impinj Company is used to receive and process signals from the antenna. The subject lies on the bed and performs specified human postures according to daily habits, including supine, left flexion, right flexion, left support lateral recumbency, right support lateral recumbency, left induced lateral recumbency, right induced lateral recumbency, and prone.

[0148] Step 6.2: Segment the RFID phase data based on a time window to obtain a phase array.

[0149] Step 6.2.1: Based on a preset time window length, the RFID phase data is intercepted to obtain an initial sample sequence corresponding to multiple time windows.

[0150] Step 6.2.2: For each time window, extract the first phase observation of each RFID tag in the initial sample sequence within the current time window to form an initial sample data reflecting the instantaneous state of the RFID tag array.

[0151] Step 6.2.3: Classify each RFID tag in the initial sample data according to a preset minimum read count threshold:

[0152] If the total read count of an RFID tag within the time window is greater than or equal to the preset minimum read count threshold, mark the RFID tag as a valid tag;

[0153] If the total read count of an RFID tag within the time window is less than the preset minimum read count threshold, mark the RFID tag as a missing tag.

[0154] In this embodiment, the preset minimum read count threshold is set to 3 times.

[0155] Step 6.2.4: Form a matrix of rows and columns from all valid tags and missing tags as a sample read tag array. Row Column

[0156] In this embodiment, represents the number of rows of the sample read tag array, represents the number of columns of the sample read tag array.

[0157] Step 6.2.5: For each missing tag, check the RFID tags in the eight directions of the position of the missing tag in the sample read tag array.

[0158] If the RFID tags in the eight directions are all marked as valid tags within the current time window, calculate the average of the phase values of the RFID tags in the eight directions, use the average to fill in the phase value of the missing tag, and update the missing tag to a valid tag;

[0159] If there is at least one RFID tag in the eight directions that is not marked as a valid tag within the current time window, the missing tag remains as a missing tag.

[0160] Step 6.2.6: Take all the phase arrays marked as valid tags and missing tags as the final phase array.

[0161] In this embodiment, the eight directions include directly above, right above, directly right, right below, directly below, left below, directly left, and left above.

[0162] Step 6.2.5: All the phase arrays labeled as valid tags and missing tags are taken as the final phase array.

[0163] Step 6.3: Perform phase unwrapping on the phase array based on the sliding window to obtain an unwrapped phase matrix.

[0164] Step 6.3.1: Based on the preset size and sliding step of the sliding window, construct a sliding window along the time dimension on the phase array.

[0165] Step 6.3.2: For the phase array in each sliding window:

[0166] Calculate the difference between the phase value at each time point of each RFID tag and the phase value at the previous time point in turn;

[0167] If the absolute value of the difference exceeds the phase jump threshold π, it is determined that a phase jump occurs, and phase jump correction is performed by adding or subtracting an integer multiple of 2π to the phase value at the current time point, so that the absolute value of the phase difference after phase jump correction is less than π.

[0168] Step 6.3.3: After the phase jump correction of the phase array in all sliding windows is completed, an unwrapped phase matrix is obtained.

[0169] Step 6.4: Classify the RFID tags based on the unwrapped phase matrix.

[0170] In this embodiment, the RFID tags include blocked tags, near neighbor tags, and far-end tags.

[0171] Step 6.4.1: Convert the unwrapped phase matrix to a category matrix , and extract valid tags and missing tags.

[0172] Wherein, 1 represents a blocked tag, 2 represents a near neighbor tag, and 3 represents a far-end tag, represents the category matrix.

[0173] Step 6.4.2: Initialize the valid tags as far-end tags.

[0174] Step 6.4.3: Initialize the missing tags as blocked tags.

[0175] Step 6.4.4: Traverse each blocked tag:

[0176] If all the RFID tags in the eight-neighborhood of the blocked tag are far-end tags, update the blocked tag as a far-end tag;

[0177] If there is at least one RFID tag in the octet neighborhood of the blocking tag which is not a far-end tag, the blocking tag is kept as a blocking tag.

[0178] Step 6.4.5: Perform octet neighborhood connected region marking for each blocking tag.

[0179] If the number of tags in the connected region is less than the preset number of tags in the connected region threshold, all the blocking tags in the connected region are updated as far-end tags.

[0180] If the number of tags in the connected region is greater than or equal to the preset number of tags in the connected region threshold, all the blocking tags in the connected region are kept as blocking tags.

[0181] Step 6.4.6: Search for non-blocking tags within the preset Manhattan distance range for the remaining blocking tags and update them as near-end tags.

[0182] Step 6.5: Select a preset number of far-end tags as a reference tag set, and perform phase data alignment operation on all the unwrapped phase arrays to obtain an input phase matrix.

[0183] Step 6.5.1: Group all the unwrapped phase arrays according to the same reference.

[0184] In this embodiment, the same reference includes the same subject or the same action.

[0185] Step 6.5.2: Calculate the reading stability index value, phase stability index value and average minimum distance away from the area where the blocking tag is located of each RFID tag in all the unwrapped phase arrays in the group, and obtain a comprehensive score based on weighted average.

[0186] In this embodiment, the calculation formula of the reading stability index value is as follows:

[0187] ;

[0188] In the formula, Rij represents the reading stability index value of the RFID tag in the i-th row and the j-th column in the group, is the number of unwrapped phase arrays in the group, is the number of times the RFID tag in the i-th row and the j-th column in the group is read in all the unwrapped phase arrays in the group. In this embodiment, the calculation formula of the phase stability index value is as follows:

[0189]

[0190] ;​​​

[0191] In the formula, This represents the phase stability index value of the RFID tag in the i-th row and j-th column of the unwrapped phase array within the group. This represents the variance calculation function. This represents the phase value of the RFID tag in the i-th row and j-th column.

[0192] In this embodiment, the formula for calculating the average minimum distance away from the area where the blocking label is located is expressed as follows:

[0193] ;

[0194] In the formula, This represents the average minimum distance from the RFID tag in the i-th row and j-th column of the unwrapped phase array within the group to the area where the blocking tag is located. These represent the x and y coordinates of the blocking label, respectively. Indicates the first The set of locations of the blocking tags in the unwrapped phase array.

[0195] Step 6.5.3: Select multiple remote labels with a comprehensive score higher than the preset comprehensive score threshold as the set of remote reference labels within the group. .

[0196] In this embodiment, the formula for calculating the comprehensive score is expressed as follows:

[0197] ;

[0198] In the formula, This represents the overall score of the RFID tag in the i-th row and j-th column of all unwrapped phase arrays within the group. Let represent the stability index value, phase stability index value, and weighting coefficient of the average minimum distance from the area where the blocking tag is located for the RFID tag in the i-th row and j-th column, respectively. This indicates a normalization operation.

[0199] Step 6.5.4: Calculate the first [value] based on the remote reference tag set. The weighted average phase of the unwrapped phase array is expressed as:

[0200] ;

[0201] In the formula, Indicates the first A detangled phase array, through a set of remote reference tags The calculated weighted average phase. Indicates weight, , Indicates the first the phase value of the RFID tag in the first row and the first column of the unwrapped phase array, the phase value of the RFID tag in the first row and the first column of the unwrapped phase array, the phase value of the RFID tag in the first row and the first column of the unwrapped phase array, representing a minimum value for preventing zero division.

[0202] Step 6.5.5: After weighting and averaging the effective phase points of all the unwrapped phase arrays, a phase unwrapping operation is performed again to obtain an input phase matrix.

[0203] wherein the value of the near neighbor tag is set as the phase value of the input phase matrix, the value of the blocked tag is set as 0, and the far-end tag is masked; the far-end tag is only used for phase reference alignment and weighted average phase calculation, and does not participate in classification as a network input feature.

[0204] Step 6.6: Data augmentation is performed on the input phase matrix to obtain an expanded data set.

[0205] Step 6.6.1: Offset noise is added to the phases of the input phase matrix.

[0206] Step 6.6.2: Global Gaussian noise is added to the phases corresponding to the near neighbor tags.

[0207] Step 6.6.3: A preset percentage of phase data corresponding to RFID tags is randomly discarded.

[0208] Step 6.6.4: The area where the blocked tag is located is horizontally or vertically translated.

[0209] Step 6.6.5: User body size scaling is performed according to the proportion of the number of blocked tags and a decision scaling factor to obtain an expanded data set:

[0210] If the proportion of the number of blocked tags is greater than a preset proportion threshold of the number of blocked tags, it is determined that the user is of a large body size, and the decision scaling factor is determined to have a value range less than 1.

[0211] If the proportion of the number of blocked tags is less than a preset proportion threshold of the number of blocked tags, it is determined that the user is of a small body size, and the decision scaling factor is determined to have a value range greater than 1.

[0212] Step 6.6.6: Based on the value of the decision scaling factor, coordinate scaling is performed on the area composed of the blocked tags and the near neighbor tags.

[0213] In this embodiment, data augmentation is performed on the input phase matrix based on one or a combination of steps 6.6.1-6.6.6 to obtain an expanded data set.

[0214] The embodiment verifies the accuracy of the obtained bed human posture recognition result including one of supine and prone and one of left flexion, right flexion, left support side lying, right support side lying, left induced side lying and right induced side lying through specific experimental results.

[0215] It should be noted that the flexion posture is fetal-like curling and knee bending, facing left is left flexion, and facing right is right flexion; the support side lying is straight legs stacked, and the pelvis is not rotated; the induced side lying is the upper leg stretched forward, and the pelvis is rotated. The human postures selected in the embodiment have certain medical clinical significance, for example, long-term supine can easily increase snoring and sleep apnea, flexion can reduce the risk of snoring and gastroesophageal reflux, and induced side lying can easily increase lower back pain and sacroiliac joint instability. The data collection of the embodiment fully considers the diversity of human postures in the real sleep scene, and divides the common sleep arm positions into three categories, namely, the arm close to the trunk, the arm over the head and the arm naturally placed. For data collection of each human posture, the embodiment requires the subjects to constantly adjust their hand positions and body positions to achieve more diversified collection effect. In addition, the arm placement position during sleep affects the shoulder and neck burden and can provide more detailed medical advice. The embodiment collects human posture data of 12 subjects, and adopts leave-one-out cross-validation, that is, the data of one subject is selected as the test set, and the data of all other subjects is selected as the training set, to ensure that the test subject is not obtained in advance by the model in the training stage, and to accurately evaluate the model precision and generalization ability.

[0216] The embodiment mainly uses the accuracy of the classification precision bed human posture recognition result. The overall performance of the embodiment in the cross-user human posture authentication task is stable, and the comprehensive authentication precision of most users is maintained above 90%, and some users reach the optimal level close to 100%. Under the leave-one-out cross-validation setting, the comprehensive average precision of all users reaches 90.52%, indicating that the proposed two-stage neural network can effectively alleviate the feature drift caused by the difference between users and improve the adaptability of cross-domain human posture recognition. From the initial bed human posture recognition result output by the main network, it has achieved high class distinction ability on most user data, showing good human posture pattern fitting ability. The branch network for supine / prone posture further improves the discrimination precision of the local posture, so that reliable results can still be obtained through the second stage of fine classification when the main network prediction is uncertain. It can be seen from Table 2 that the branch network obtains a classification precision of more than 90% on multiple users, proving that the two-stage neural network provided by the embodiment is superior to relying solely on the main network or the branch network, not only improving the classification robustness, but also maintaining stable performance in the case of large differences between users. Figure 6

[0217] ​To further verify the accuracy of the present application in multi-class human posture recognition tasks, the present application independently classifies the performance of each human posture in the bed human posture recognition result, and the result is shown in Table 1.

[0218] Table 1 Classification results of the present application for different human postures

[0219] Human pose name Classification accuracy Supine 90.20% Left side bend 98.00% Right side bend 92.67% Left side support 91.67% Right side support 92.33% Left induced side 90.00% Right induced side 90.33% Prone 86.70%

[0220] The present application has achieved a high classification accuracy of more than 90% in most human posture categories, indicating that the system can fully utilize the spatial response characteristics of the RFID tag array to distinguish different body posture patterns. Among them, the recognition effect of the left curved lying human posture is the most significant, with a classification accuracy of 98%, indicating that this posture has clear and robust pattern characteristics in the backscattering signal. For symmetrical postures, the classification accuracy is maintained at about 92%, indicating that the model of the present application can effectively distinguish the slight differences in tag reflection characteristics between symmetrical human postures, and has good fine-grained discrimination ability. In addition, the classification accuracy of supine and left / right induced lateral lying postures is maintained at about 90%, showing that the model still has stable generalization performance when facing different body types and differences between different users.

[0221] To verify the effectiveness of the two-stage neural network of the present application, the two-stage neural network of the present application is compared with the single-stage scheme using only the main network, and the result is shown in Table 2.

[0222] Table 2 Comparison of classification results of the present application using a two-stage neural network and only using a main network

[0223] User number Two-stage neural network Only main network user1 100.00% 97.83% user2 98.70% 89.57% user3 91.30% 90.43% user4 90.00% 86.52% user5 84.78% 82.61% user6 85.22% 84.35% user7 99.13% 86.96% user8 93.04% 92.61% user9 86.52% 83.04% user10 76.52% 69.57%

[0224] The two-stage neural network used by the present application achieves accuracy improvement on all users, with a single user improvement range of 0.43% to 12.17%, and an overall average improvement of about 4.17%, proving the effectiveness of the present embodiment in improving human posture authentication accuracy and enhancing cross-user robustness, and having deployability under real-time constraints, fully embodying the practical application value of the present embodiment for RFID passive sleep perception systems.

[0225] Embodiment 3

[0226] Based on the same inventive concept as other embodiments, the present embodiment introduces a computer readable storage medium having computer instructions stored thereon, which when executed by a processor, implements the steps of the method of Embodiment 1 or 2.

[0227] Embodiment 4

[0228] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product comprising computer instructions which, when executed by a processor, implement the steps of the method of embodiment 1 or 2.

[0229] In summary of the above embodiments, the present application effectively improves the accuracy of bed human posture recognition result judgment by segmenting and unwrapping the RFID phase data based on time windows and sliding windows respectively, classifying the unwrapped phase matrix according to the distance, selecting the far-end tags for phase data alignment, and finally using the trained two-stage neural network classification, solves the problem that the prior art cannot recognize complex limb postures on the bed and it is difficult to distinguish supine and prone postures when the user's arms are raised or close to the head.

[0230] The main network of the two-stage neural network of the present application uses the input phase matrix and the first mask matrix generated by the blocking tag to perform operations such as splicing, feature extraction, attention enhancement, and based on the statistical feature vector, the human posture features are analyzed from multiple dimensions, the initial bed human posture recognition result is accurately recognized, the branch network is further focused on the waist trunk area, the second mask matrix is used to assist in extracting local sparse features, through the spatial attention mechanism and the pooling operation, the supine / prone posture is accurately recognized, compared with the traditional method, the comprehensiveness and accuracy of the bed human posture recognition result are greatly improved, whether it is a complex limb posture or an easily confused supine / prone posture, it can be accurately recognized.

[0231] When the present application obtains the RFID phase data and segments the phase array based on the time window, the missing tags are filled to ensure the integrity of the phase array. Then, the phase unwrapping is performed through the sliding window, the phase jump is accurately corrected, the phase wrapping problem is effectively solved, and the accurate phase matrix after unwrapping is obtained. In the RFID tag classification process, the blocking tags, near-end tags and far-end tags are accurately distinguished through eight-neighborhood connected region marking and Manhattan distance search, the far-end tags with high comprehensive score are selected as the reference tag set for phase data alignment, the accuracy and stability of the input phase matrix are ensured, and the reliability of the recognition result is improved.

[0232] The application trains the two-stage neural network by using the expanded data set, adds offset noise and global Gaussian noise through input phase matrix, randomly discards part of the phase data corresponding to the RFID tag, translates the area where the blocked tag is located, and scales the user body according to the proportion of the number of blocked tags, and the like, simulates the data changes in various actual scenes, effectively increases the diversity and richness of the data, so that the model can learn more features in different scenes, and improves the generalization ability of the model. At the same time, the two-stage neural network is trained in combination with the expanded data set, so that the model can better adapt to various complex situations, further improve the accuracy and stability of the bed human body posture recognition result judgment, and has higher practical value in actual application.

[0233] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware aspects. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) having computer usable program code embodied thereon.

[0234] The application is described with reference to flowcharts and / or block diagrams that illustrate the methods, apparatus (systems) and computer program products according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in the flowchart

[0235] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in the flowchart

[0236] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are generated to realize the computer-implemented processes, and the instructions executed on the computer or other programmable devices provide a process for implementing the functions specified in the flowchart Figure 1 one flow or multiple flows and / or the functions specified in the block Figure 1 one flow or multiple flows and / or the functions specified in the block

[0237] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative, but not restrictive, and those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.

Claims

1. A method for human posture recognition on bed based on multi-level label collaborative perception, characterized in that, The method comprises the following steps: acquiring RFID phase data to be identified; segmenting the RFID phase data to be identified based on a time window to obtain a phase array; performing phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix; classifying RFID tags based on distance for the unwrapped phase matrix, wherein the RFID tags include blocked tags, near neighbor tags, and far-end tags; selecting a preset number of far-end tags as a reference tag set, and performing phase data alignment on all unwrapped phase arrays to obtain an input phase matrix; performing classification based on a trained two-stage neural network according to the input phase matrix to obtain a bed occupant posture recognition result; the bed occupant posture recognition result includes one of supine and prone, and one of left lateral recumbent, right lateral recumbent, left support lateral recumbent, right support lateral recumbent, left induced lateral recumbent, and right induced lateral recumbent; the network structure of the two-stage neural network includes a main network and a branch network; the main network is used for classification according to the input phase matrix, and outputs an initial bed occupant posture recognition result, wherein the initial bed occupant posture recognition result includes prone-supine combined recumbent, left lateral recumbent, right lateral recumbent, left support lateral recumbent, right support lateral recumbent, left induced lateral recumbent, and right induced lateral recumbent; the branch network is used for identifying and outputting a supine posture or a prone posture result according to the initial bed occupant posture recognition result from the input phase matrix corresponding to the prone-supine combined recumbent, and determining the input phase matrix corresponding to the waist trunk region according to the spatial distribution range of the blocked tag; classifying RFID tags based on distance for the unwrapped phase matrix comprises: Converting the unwrapped phase matrix into a category matrix , extracting valid and missing labels, where 1 represents a block label, 2 represents a near neighbor label, and 3 represents a far end label, denotes a category matrix; initializing effective tags as far-end tags; initializing missing tags as blocked tags; traversing each blocked tag: if all RFID tags in the eight-neighborhood of the blocked tag are far-end tags, updating the blocked tag as a far-end tag; if there is at least one RFID tag in the eight-neighborhood of the blocked tag that is not a far-end tag, keeping the blocked tag as a blocked tag; performing eight-neighborhood connected region marking for each blocked tag; if the number of tags in the connected region is less than a preset threshold of the number of tags in the connected region, updating all blocked tags in the connected region as far-end tags; if the number of tags in the connected region is greater than or equal to the preset threshold of the number of tags in the connected region, keeping all blocked tags in the connected region as blocked tags; searching for non-blocked tags within a preset Manhattan distance range for the remaining blocked tags and updating them as near neighbor tags. 2.The multi-level label collaborative sensing based human posture recognition method on bed according to claim 1, characterized in that, the main network comprises: The first dual-channel input layer is configured to receive the input phase matrix and a first mask matrix, wherein the first mask matrix is obtained by binarizing the input phase matrix; the binarization process includes: traversing the input phase matrix, if an element at a current position in the input phase matrix is a non-numerical value, setting a mask value at a corresponding position in a binary mask matrix to 1, otherwise setting the mask value to 0, and generating a binary mask matrix with the same dimension as the input phase matrix as the first mask matrix after the traversal is completed; the feature concatenation layer is configured to concatenate the input phase matrix and the first mask matrix in a channel direction to obtain concatenated features; the first feature extraction module is configured to process the concatenated features through a plurality of convolution layers and residual blocks to extract spatial features; the first attention module is configured to enhance the spatial features through a convolution block attention module to obtain enhanced features; the first pooling layer is configured to perform global average pooling on the enhanced features to extract a global feature vector; the feature calculation module is configured to calculate a statistical feature vector based on the first mask matrix; the first feature fusion layer is configured to concatenate the global feature vector and the statistical feature vector to obtain a first concatenated vector; the first feedforward network is configured to perform nonlinear transformation and feature dimension reduction on the first concatenated vector to obtain a first high-level feature vector; the first classification layer is configured to classify the first high-level feature vector to obtain and output an initial in-bed human posture recognition result; wherein the statistical feature vector includes: an area proportion of a region where a blocking label is located, a barycenter normalized coordinate of the region where the blocking label is located, a bounding box length-width ratio and a bounding box width-length ratio of the region where the blocking label is located, left and right distribution proportions, up and down distribution proportions, left and right differences, and up and down differences of a number of blocking labels in the region where the blocking label is located, and a phase mean and a standard deviation of effective labels within a preset distance range of the blocking label. 3.The multi-level label collaborative sensing based human posture recognition method on bed according to claim 2, characterized in that, The branch network comprises: a second dual-channel input layer configured to receive an input phase matrix corresponding to the supine-lying combination and extract an input phase matrix corresponding to a waist torso region, and a second mask matrix obtained by binarization processing of the input phase matrix corresponding to the waist torso region; traversing the input phase matrix corresponding to the waist torso region, if an element at a current position in the input phase matrix corresponding to the waist torso region is a non-number, setting a mask value of a corresponding position of a binary mask matrix to 1, otherwise setting the mask value to 0, and generating a binary mask matrix having the same dimension as the input phase matrix as the second mask matrix after the traversal; a second feature extraction module configured to process the input phase matrix corresponding to the waist torso region and the second mask matrix through a plurality of convolution layers and residual blocks to extract local sparse features; a second attention module configured to weight the importance of each spatial position of the local sparse features through a convolution block attention module to obtain a weighting result; a second pooling layer configured to perform spatial attention pooling on the local sparse features according to the weighting result to obtain a weighted average feature vector; a second feature fusion layer configured to integrate all the weighted average feature vectors to obtain integrated features; and a second feedforward network and a linear classification layer configured to map and classify the integrated features to output a binary classification result of the supine or prone posture. 4.The multi-level label collaborative sensing based human posture recognition method on bed according to claim 3, characterized in that, The two-stage neural network is trained using an expanded data set to obtain a bed posture recognition result according to an input phase matrix; wherein the method for obtaining the expanded data set comprises: RFID phase data is collected by deploying an RFID tag array on a bed; The RFID phase data is segmented based on a time window to obtain a phase array; Phase unwrapping is performed on the phase array based on a sliding window to obtain an unwrapped phase matrix; The unwrapped phase matrix is classified based on distance into RFID tags, including blocked tags, near tags and far tags; A preset number of far tags are selected as a reference tag set, and a phase data alignment operation is performed on all the unwrapped phase arrays to obtain an input phase matrix; The input phase matrix is subjected to data enhancement to obtain an expanded data set.

5. The multi-level label collaborative sensing based human posture recognition method on bed according to claim 4, characterized in that, The RFID phase data is segmented based on a time window to obtain a phase array, comprising: The RFID phase data is intercepted based on a preset time window length to obtain a plurality of initial sample sequences corresponding to the time windows; For each time window, the first phase observation value of each RFID tag in the current time window in the initial sample sequence is extracted to form an initial sample data reflecting the instantaneous state of the RFID tag array; Each RFID tag in the initial sample data is classified according to a preset minimum reading number threshold: If the total number of readings of a certain RFID tag in the time window is greater than or equal to the preset minimum reading number threshold, the RFID tag is marked as a valid tag; If the total number of readings of a certain RFID tag within the time window is less than a preset minimum reading number threshold, the RFID tag is marked as a missing tag; A matrix of rows and columns is formed from all valid tags and missing tags as a sample read mark array, wherein, row column represents a number of rows of the sample read mark array, represents a number of columns of the sample read mark array; For each missing tag, check the RFID tags in eight directions of the position of the missing tag in the sample reading mark array: If the RFID tags in the eight directions are all marked as valid tags within the current time window, calculate the average of the phase values of the RFID tags in the eight directions, use the average to fill in the phase value of the missing tag, and update the missing tag to a valid tag; If there is at least one RFID tag in the eight directions that is not marked as a valid tag within the current time window, the missing tag is kept as a missing tag; The phase arrays marked as valid tags and missing tags are taken as the final phase array; The eight directions include directly above, right above, directly right, right below, directly below, left below, directly left, and left above.

6. The multi-level label collaborative sensing based human posture recognition method on bed according to claim 5, characterized in that, Perform phase unwrapping on the phase array based on a sliding window to obtain an unwrapped phase matrix, including: Based on the size and sliding step of the preset sliding window, a sliding window is constructed on the phase array along the time dimension; for the phase array in each sliding window: Calculate the difference between the phase value at each time and the phase value at the previous time for each RFID tag; if the absolute value of the difference exceeds the phase jump threshold π, it is determined that a phase jump has occurred, and the phase jump is corrected by adding or subtracting an integer multiple of 2π from the current phase value, so that the absolute value of the phase difference after phase jump correction is less than π; after phase jump correction is completed on all phase arrays in the sliding window, an unwrapped phase matrix is obtained. 7.The multi-level label collaborative sensing based human posture recognition method on bed according to claim 1, characterized in that, Select a preset number of far-end tags as a reference tag set, and perform phase data alignment on all unwrapped phase arrays to obtain an input phase matrix, including: Group all unwrapped phase arrays according to the same reference; Calculate the reading stability index value, phase stability index value, and average minimum distance from the area where the blocking tag is located for each RFID tag in all unwrapped phase arrays in the group, and obtain a comprehensive score based on weighted average; The calculation formula of the reading stability index value is: ； In the formula, represents the read stability index value of the RFID tag in the i-th row and j-th column of the group, is the number of the unwrapped phase arrays in the group, is the number of the RFID tag in the i-th row and j-th column of all the unwrapped phase arrays in the group, is the number of the RFID tag in the i-th row and j-th column of all the unwrapped phase arrays in the group, is the number of the RFID tag in the i-th row and j-th column of all the unwrapped phase arrays in the group, The calculation formula of the phase stability index value is: ； In the formula, denotes the phase stability index value of the RFID tag in the i-th row and j-th column of the unwound phase array, denotes a variance calculation function, denotes the phase value of the RFID tag in the i-th row and j-th column. The calculation formula of the average minimum distance from the area where the blocking tag is located is: ； In the formula, represents the average minimum distance from the RFID tag in the i-th row and the j-th column of the disentangled phase array to the area where the blocking tag is located, respectively represent the horizontal coordinate and the vertical coordinate of the blocking tag, represents the area where the blocking tag is located in the i-th disentangled phase array; Selecting multiple remote tags with a comprehensive score higher than a preset comprehensive score threshold as a remote reference tag set in the group ; The calculation formula of the comprehensive score is: ； In the formula, denotes the comprehensive score of the RFID tag in the i-th row and j-th column of all the unwound phase arrays in the group, denote the stability index value, the phase stability index value and the weight coefficient of the average minimum distance away from the area where the blocking tag is located of the RFID tag in the i-th row and j-th column, respectively, denotes the normalization operation; Calculate the first based on the remote reference tag set The weighted average phase of the unwrapped phase array is expressed as: ； wherein represents the array of unwrapped phases, calculated by the remote reference tag set represents the weight, represents the array of unwrapped phases, represents the phase value of the RFID tag in row represents a minimum value for preventing division by zero;​​​​ After weighted average of the valid phase points of all unwrapped phase arrays, perform a phase unwrapping operation again to obtain an input phase matrix; Wherein, the value of the near neighbor tag is set as the phase value of the input phase matrix, the value of the blocking tag is set as 0, and the far-end tag is masked. 8.The multi-level label collaborative sensing based human posture recognition method on bed according to claim 7, characterized in that, Perform data augmentation on the input phase matrix to obtain an expanded data set, including: Add offset noise to the phase of the input phase matrix, and / or, Add global Gaussian noise to the phase corresponding to the near neighbor tag, and / or, Randomly discard a preset percentage of phase data corresponding to the RFID tag, and / or, performing horizontal translation or vertical translation on the area where the blocking label is located, and / or, scaling the user body according to the proportion of the number of blocking labels and determining the scaling factor to obtain an expanded data set: if the proportion of the number of blocking labels is greater than a preset proportion threshold of the number of blocking labels, it is determined that the user is a large body type, and the value range of the scaling factor is less than 1; if the proportion of the number of blocking labels is less than the preset proportion threshold of the number of blocking labels, it is determined that the user is a small body type, and the value range of the scaling factor is greater than 1; performing coordinate scaling on the area composed of the blocking label and the near neighbor label based on the value of the determined scaling factor.

Images