CN115660042A - Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment - Google Patents
Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment Download PDFInfo
- Publication number
- CN115660042A CN115660042A CN202211018800.7A CN202211018800A CN115660042A CN 115660042 A CN115660042 A CN 115660042A CN 202211018800 A CN202211018800 A CN 202211018800A CN 115660042 A CN115660042 A CN 115660042A
- Authority
- CN
- China
- Prior art keywords
- matrix
- data
- dimensional
- phase
- rssi
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a three-dimensional sleeping posture detection method based on RFID equipment, which comprises the following steps: step 1: deploying an RFID tag array and collecting data; synchronously collecting point cloud data; step 2: data segmentation and elimination: dividing the RFID data collected in the step (1) by utilizing a plurality of time windows, and counting the fluctuation size of the label array signal of each time window to remove the data of the sleep turning; and step 3: preprocessing data; and 4, step 4: carrying out data enhancement; and 5: and building a convolutional neural network model, extracting data characteristics, training to obtain a trained convolutional neural network model, and inputting the data to be tested into the model to obtain the three-dimensional sleeping posture. The invention overcomes the defect that the existing wireless signal can not realize fine-grained three-dimensional sleeping posture and the instability caused by the easy interference of the wireless signal.
Description
Technical Field
The invention belongs to the technical field of wireless signal three-dimensional posture sensing, and relates to a three-dimensional sleeping posture detection method based on RFID equipment.
Background
In sleep health monitoring, sleep posture has important significance and influence on sleep quality, body health, disease detection, adjuvant therapy and the like. Most of the existing sleeping posture monitoring methods focus on sleeping posture classification, such as supine, prone and side lying. However, a single sleep posture classification lacks the more abundant limb posture information in sleep. In fact, three-dimensional position information of key parts of the body, such as the body, the legs, the arms, and the head, is not only an important factor for achieving comfortable sleep, but also for disease diagnosis. The three-dimensional sleeping posture displays the sleeping posture hologram from multiple angles, so that a more natural and real sleeping posture can be identified, meanwhile, rich information is provided for long-term sleep monitoring and disease diagnosis, and the three-dimensional sleeping posture hologram has very strong practical significance and practical value.
The existing wireless sensing sleeping posture detection can only provide posture classification with a coarse granularity, and cannot provide a fine granularity to obtain human body joint three-dimensional positioning due to the limitation of the resolution of low-frequency wireless equipment. For example, wireless devices such as Wi-Fi, FMCW, RFID, etc. are deployed around a bed for sleep posture monitoring by analyzing changes in wireless signal characteristics, including signal received strength RSS, channel state information CSI, phase, etc., generated by different sleep postures. However, the above methods are all focused on recognizing sleeping posture classification or judging body orientation, and accurate recognition of three-dimensional sleeping posture cannot be performed. Meanwhile, wireless signals are easily interfered by external environment, and stability and robustness are not easy to realize.
In addition, there are also some works that use the principle of reflection of wireless signals to infer the feasibility of human anatomy. For example, devices such as WIFI, RF signals, millimeter waves and the like are deployed around a person, and a three-dimensional posture is extracted by capturing the reflection of the signals under the action of the person. However, such methods rely on human walking and movement to achieve good accuracy. In contrast, a person remains stationary in a sleeping posture for a majority of the time, so these methods are not applicable to static three-dimensional sleeping postures.
Disclosure of Invention
The invention aims to provide a three-dimensional sleeping posture detection method based on RFID equipment, and aims to solve the technical problems that the three-dimensional sleeping posture cannot be accurately identified and cannot be applied to static three-dimensional sleeping posture identification in the prior art.
In order to achieve the purpose, the invention adopts the technical scheme that:
a three-dimensional sleeping posture detection method based on RFID equipment specifically comprises the following steps:
step 1: deploying an RFID tag array and collecting data; synchronously collecting point cloud data;
and 2, step: data segmentation and elimination: dividing the RFID data collected in the step (1) by using a plurality of time windows, and counting the fluctuation size of the tag array signal of each time window to remove the data of sleep turning;
and step 3: data preprocessing, namely performing RF image conversion and preprocessing on the data under each time window obtained in the step (2); dividing the obtained data into a training set and a test set;
and 4, step 4: performing data enhancement on the rsi matrix, the phase matrix and the binarization matrix in the training set obtained in the step 3 to obtain an enhanced training set;
and 5: and (3) building a convolutional neural network model, extracting data characteristics of the data obtained in the step (3) and the step (4), training to obtain a trained convolutional neural network model, and inputting the data to be tested into the model to obtain the three-dimensional sleeping posture.
Further, the operation of step 1 is as follows: deploying a 28-by-21 RFID tag array, paving the RFID tag array on a mattress, and placing the RFID tag array above the mattress by using an antenna, wherein the distance between the antenna and the tag array is 1.6-2.3 m; meanwhile, kinect V2 is used for collecting synchronous point cloud information and marking the three-dimensional sleeping postures.
Further, the step 2 comprises the following sub-steps:
step 2.1: data segmentation: the RFID data collected in step 1 is divided in 3s time windows, and the data in each time window is represented as:
rssi m,n (t)＝{rssi 1 m,n ,rssi 2 m,n ,rssi 3 m,n ,...,rssi i m,n }#(1)
phase m,n (t)＝{phase 1 m,n ,phase 2 m,n ,phase 3 m,n ,...,phase i m,n }#(2)
wherein rssi m,n (t) rssi sequence value, phase, of the tag with position m, n in the t-th time window m,n (t) represents a sequence value of the phase of the tag with the position m and n in the t-th time window, and i represents the number of times the tag with the position m and n is read in the t time windows;
step 2.2: data elimination: and calculating the variance sum of the rssi sequence and the phase sequence of all the non-blocking labels under each time window, eliminating the data of the time window larger than the threshold value, and keeping the data of the time window within the threshold value.
Further, in step 2.2, the threshold of the rssi variance sum is taken to be 5, and the threshold of the phase variance sum is taken to be 8.
Further, the step 3 comprises the following substeps:
step 3.1: respectively calculating the average value of the rssi sequence and the phase sequence of the non-blocking label in each time window obtained in the step (2) by using a formula 3 and a formula 4, and simultaneously assigning the rssi average value of the blocking label as-100 and the phase average value as 8; then combining the label array information to obtain an rssi matrix and a phase matrix:
wherein the content of the first and second substances,
represents the mean of the rssi sequence with the tag position m, n under a certain time window,
the mean value of the phase sequence with the label position of m and n under a certain time window is represented;
step 3.2: and (3) carrying out binarization processing on the rssi matrix obtained in the step (3.1), taking a threshold value as-100, taking the value of an element which is greater than the threshold value in the rssi matrix as zero, and taking 1 if the value is equal to the threshold value, so as to obtain a binarization matrix:
wherein, b m,n Binary data which shows that the label position is m and n under a certain time window;
step 3.3: denoising the phase matrix and the rsi matrix obtained in the step 3.1 and the binarization matrix obtained in the step 3.2 respectively to obtain a denoised rsi matrix, a denoised phase matrix and a denoised binarization matrix;
step 3.4: normalizing the denoised rsi matrix and phase matrix obtained in the step 3.3 by using a formula 6 and a formula 7 respectively to obtain a normalized rsi matrix, a normalized phase matrix and a normalized binarization matrix:
wherein rssi m,n The denoised label position is the rssi value, rssi of m, n min Is the minimum value of the denoised rssi matrix, rssi max Is the maximum value of the denoised rssi matrix, phase m,n The denoised label position is m, n phase value, phase min For de-noised phase matrixMinimum value of (1), phase max The maximum value of the denoised phase matrix is obtained;
step 3.5: using a bilinear difference value method to respectively perform two-time upsampling on the normalized rssi matrix, the phase matrix and the binarization matrix obtained in the step 3.4, and enlarge the size of the matrix to be 2 times of the original size to obtain a processed rssi matrix, a processed phase matrix and a processed binarization matrix; the resulting data is divided into a training set and a test set.
Further, in the step 3.3, denoising is performed by using a gaussian distribution function.
Further, the step 4 comprises the following sub-steps:
step 4.1: performing horizontal mirroring on the rssi matrix, the phase matrix and the binarization matrix obtained in the step 3;
step 4.2: carrying out size transformation on the data obtained in the step 3;
step 4.3: and (3) repeating the processing process of the step (3.5) on the rssi matrix, the phase matrix and the binarization matrix obtained in the step (4.1) and the step (4.2) to obtain the enhanced rssi matrix, the phase matrix and the binarization matrix.
Further, the step 4.3 specifically operates as follows: judging a two-dimensional mapping space of the sleeping posture according to the region where the numerical value 1 in the binarization matrix obtained in the step 3.4 is located; then, the left and right boundary ranges of the sleeping postures are found according to the two-dimensional mapping space of the sleeping postures, data outside the boundaries are cut, and the binary matrix is changed from 28 x 21 to 28 x n, n<21; simultaneously, synchronously cutting the rssi matrix and the phase matrix obtained in the step 3.4 to obtain the rssi matrix and the phase matrix with the size of 28 x n; then, respectively splicing two sides of the obtained rssi matrix, phase matrix and binarization matrix with the size of 28 x n
The matrix of (2), the rsi matrix, the phase matrix and the binarization matrix of 28 × n size are changed to 28 × 21 size.
Further, the step 5 comprises the following sub-steps:
step 5.1: building a convolutional neural network in a Pythrch-based environment: firstly, building a two-dimensional convolution layer, introducing an attention mechanism, and then connecting the two-dimensional convolution layers to combine the two-dimensional convolution layer with the attention mechanism; then, the output of the two-dimensional convolution layer is expanded to three-dimensional and connected to a three-dimensional convolution layer of six layers; finally, connecting the three-dimensional convolution layer to a full-connection layer to construct a convolution neural network model;
step 5.2: setting a loss function in a training neural network, and simultaneously training a network model by using an optimizer Adam of a self-adaptive learning rate;
step 5.3: and (4) training the constructed convolutional neural network model by using the enhanced training set obtained in the step (4) to obtain the trained convolutional neural network model.
Step 5.4: and (5) inputting the trained model obtained in the step 5.3 by using the test set, and outputting a detected three-dimensional sleeping posture result.
Further, the step 5.1 comprises the following operations:
step 5.1.1, constructing a layer of two-dimensional convolution layer, wherein the size of a convolution kernel is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 64, an activation function LeakyReLU is activated, and the data is normalized by using BatchNorm2 d; then connecting the channel attention model and then connecting six two-dimensional convolution layers; the structure of the six two-dimensional convolutional layers: the convolution kernel size of the first convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 128, the function LeakyReLU is activated, the data is normalized by using BatchNorm2d, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the second convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the third convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU uses BatchNorm2d to carry out normalization processing on the data, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the fourth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 512, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 1024, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the sixth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 1024, the function LeakyReLU is activated, the BatchNorm2d is used for carrying out normalization processing on the data, and the maximum pooling layer with the size of 2 and the step length of 2 is used;
step 5.1.2, expanding the output of the two-dimensional convolution layer to three dimensions and connecting the two-dimensional convolution layer to six three-dimensional convolution layers; the six three-dimensional convolutional layers have the following structures: the convolution kernel size of the first three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 1025, the activation function LeakyReLU is used for carrying out normalization processing on data by using BatchNorm2 d; the convolution kernel size of the second three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 512, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the third three-dimensional convolution layer uses deconvolution, the convolution kernel size is 2 multiplied by 2, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU uses BatchNorm2d to carry out normalization processing on data; the convolution kernel size of the fourth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 128, the activation function LeakyReLU performs normalization processing on the data by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 64, the activation function LeakyReLU uses BatchNorm2d to perform normalization processing on the data; the convolution kernel size of the sixth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 14, and the function Sigmoid is activated;
step 5.1.3, connecting the three-dimensional convolution layer to a full-connection layer to construct and obtain a convolution neural network model; the full connection layer is constructed by three layers, wherein the first full connection layer is provided with 4096 nodes, the activation function is ReLU, and Dropout is 0.5; the second fully connected layer has 4096 nodes with an activation function of ReLU and Dropout of 0.5; the third full-connection layer has 42 nodes, and the output of the third full-connection layer obtains the final result.
Further, the loss function constructed in step 5.2 is:
wherein L is m Representing loss values in the training convolutional neural network, N representing joint number, P i Three-dimensional coordinates representing the detected ith joint, G i Representing the three-dimensional coordinates of the ith joint in groudtuth.
Compared with the prior art, the invention has the beneficial effects that:
according to the method, the RSSI information, the phase information and the binarization information of a blocking tag and an unblocking tag in the RFID tag array in the three-dimensional sleeping posture are extracted, and a convolutional neural network model is utilized to extract three-dimensional joint points of a human body from the RSSI information, the phase information and the binarization information, so that a three-dimensional sleeping posture detection result is obtained; meanwhile, a data enhancement method is used, so that the high precision, the stability and the generalization capability of the three-dimensional sleeping posture framework detection are improved. Therefore, the defect that the existing wireless signals cannot realize fine-grained three-dimensional sleeping postures is overcome, and the wireless signals are unstable due to easy interference.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a rsi matrix, phase matrix and binary matrix grayscale image;
FIG. 3 is an image after data enhancement horizontal mirroring;
FIG. 4 is a dimension of a data enhancement shift bed size;
FIG. 5 is a block diagram of a convolutional neural network;
FIG. 6 shows the results of the method of the present invention.
Detailed Description
The invention is described in detail below with reference to the drawings and the detailed description.
As shown in fig. 1, the three-dimensional sleeping posture detection method based on the RFID device of the present invention specifically includes the following steps:
step 1: and deploying an RFID tag array for data collection. According to experiment requirements, deploying an RFID tag array, building a scene of data collection, and collecting RFID data to construct a data set. Meanwhile, kinect V2 is used for synchronously collecting point cloud data for marking the three-dimensional sleeping postures.
Specifically, a 28 × 21 RFID tag array is deployed by using H47 RFID tags, each RFID tag has unique position information and a unique name, and then the RFID tag array is laid on a mattress, a circularly polarized antenna with the frequency of 915MHZ and the gain of 8dBi is placed above the mattress, and the distance between the antenna and the tag array is 1.6 m-2.3 m. Meanwhile, kinect V2 is used for collecting synchronous point cloud information, marking the three-dimensional sleeping postures and training and supervising the convolutional neural network model and comparing detection results of the three-dimensional sleeping postures.
In this embodiment, data collection is performed on 36 subjects, 25000 data sets are collected, 20000 data sets are used as training sets for training the convolutional neural network, and 5000 data sets are used as test sets for testing the convolutional neural network.
The advantages of this step are: the step obtains a data set suitable for three-dimensional sleeping posture identification of the RFID wireless signal, and the requirement of the method on data is met. Meanwhile, in the data collection process, external interference factors can be found, and the robustness of the method is improved from the essence of the people.
Step 2: data segmentation and elimination: and (3) dividing the RFID data collected in the step (1) by using a plurality of time windows, and counting the fluctuation size of the tag array signal of each time window to remove the data of the sleep turning. Comprising the following substeps:
step 2.1: and (4) data segmentation. According to the influence of the sleeping posture on the RFID tag, the RFID tag is divided into a blocking tag and a non-blocking tag. The blocking label is a label which can not receive data under the influence of the sleeping posture, the non-blocking label is a label which can still collect signals through an antenna under the influence of the sleeping posture, and the label collected in the step 1 is a non-blocking label. The RFID data collected in step 1 are divided by 3s time windows, and the data in each time window is represented as:
rssi m,n (t)＝{rssi 1 m,n ,rssi 2 m,n ,rssi 3 m,n ,...,rssi i m,n }#(1)
phase m,n (t)＝{phase 1 m,n ,phase 2 m,n ,phase 3 m,n ,...,phase i m,n }#(2)
wherein rssi m,n (t) rssi sequence value, phase, of the tag with position m, n in the t-th time window m,n (t) represents a sequence value of the phase of the tag at the position m, n in the t-th time window, and i represents the number of times the tag at the position m, n is read in the t time windows.
Step 2.2: and (6) data elimination. And calculating the variance sum of the rssi sequence and the phase sequence of all the non-blocking labels under each time window, eliminating the data of the time window larger than the threshold value, and keeping the data of the time window within the threshold value.
Specifically, we obtain the information (including the tag name, rssi information, and phase information) of all non-blocking tags read at each time window, via step 2.1. The rssi sequence and phase sequence of all tags under each time window can be retrieved by tag name.
For step 2.2, considering that most of the sleep is in a relatively static state, data during turning needs to be removed, and static posture data is reserved and collected. Due to the fact that wireless signals are interfered by the environment, when a person is in a turning state, numerical fluctuation of an rsi sequence and a phase sequence is large, variance is an evaluation index reflecting the numerical fluctuation, under the condition, the variance sum of the rsi sequences and the variance sum of the phase sequences of all labels under a time window are counted, and then data of a corresponding time window with the rssi variance sum or the phase variance sum larger than a threshold value are removed. In this embodiment, the threshold for the rssi variance sum is taken to be 5 and the threshold for the phase variance sum is taken to be 8 based on experimental data.
The advantages of this step are: the data are divided according to the time window, the collected data are further refined, the data processing is convenient, meanwhile, redundant data brought in the turning-over process are eliminated, the condition that the data in the data set are all static sleeping posture data in a sleeping state is guaranteed, and the accuracy of the experiment is guaranteed.
And 3, step 3: and (3) data preprocessing, namely performing RF image conversion and preprocessing on the data under each time window obtained in the step (2). Comprising the following substeps:
step 3.1: and (3) respectively averaging the rssi sequence and the phase sequence of the non-blocking label in each time window obtained in the step (2) by using a formula 3 and a formula 4, and simultaneously assigning the rssi average value of the blocking label to be-100 and the phase average value to be 8. And then combining the label array information to obtain an rsi matrix and a phase matrix:
wherein the content of the first and second substances,
represents the mean of the rssi sequence with tag positions m, n under a certain time window,
represents the mean of the phase sequence with the tag position m, n under a certain time window.
The advantages of this step are: the three-dimensional information of the sleeping posture can be extracted from the label array. According to experimental data, the influence of a human body on the tag can divide the tag into a blocked tag and an unblocked tag, the blocked tag represents that the human body completely blocks the communication between the tag and the antenna, and the signal of the unblocked tag can reflect the three-dimensional space information of the human body. The rssi value of the non-blocking label is between-70 and-80, the phase value of the non-blocking label is between 0 and 2 pi, and the difference between the blocking label and the non-blocking label can be distinguished more obviously by assigning the blocking label out of the rssi value interval and the phase value interval of the blocking label. It also provides more accurate information for us to extract three-dimensional pose information from the rssi matrix and the phase matrix. The value of the label and the position information of the label are combined, so that the three-dimensional posture information and the two-dimensional plane information can be combined more favorably, and the accuracy of the three-dimensional sleeping posture is improved.
Step 3.2: and (2) performing binarization processing on the rssi matrix obtained in the step (3.1), taking a threshold value as-100, taking the value of an element which is greater than the threshold value in the rssi matrix as zero, and taking 1 if the value is equal to the threshold value to obtain a binarization matrix:
wherein, b m,n And (3) binary data which shows that the label position is m, n under a certain time window.
See fig. 2, which is a grayscale image of the rssi matrix, the phase matrix, and the binarization matrix.
The steps have the advantages that: the binary matrix comprises two-dimensional mapping information of a three-dimensional gesture, a blocking label and an unblocking label can be distinguished from numerical values of the rssi matrix, the numerical value of the blocking label is-100, the numerical value of the unblocking label is-80 to-70, so that a threshold value is-100, the non-blocking label which is larger than-100 in the rssi matrix is assigned to be 0 in the binary matrix, and the blocking label which is equal to-100 in the rssi matrix is assigned to be 1 in the binary matrix. The binarization matrix reflects the area range of the sleeping posture on a two-dimensional plane and the two-dimensional mapping information of the limbs on the plane.
Step 3.3: using a Gaussian distribution function
And (3) denoising the phase matrix and the rsi matrix obtained in the step (3.1) and the binarization matrix obtained in the step (3.2) respectively to obtain a denoised rsi matrix, a denoised phase matrix and a denoised binarization matrix.
The advantages of this step are: the noise of the rssi matrix, the phase matrix and the binarization matrix can be effectively removed. Due to the communication problem of the tags, some blocking tags which do not belong to the influence of sleeping postures exist, and the influence of the phenomenon can be reduced by using a Gaussian denoising method.
Step 3.4: normalizing the denoised rsi matrix and phase matrix obtained in the step 3.3 by using a formula 6 and a formula 7 respectively to obtain a normalized rsi matrix, a normalized phase matrix and a normalized binarization matrix:
wherein rssi m,n The denoised label position is the rssi value, rssi of m, n min Is the minimum value of the denoised rssi matrix, rssi max Is the maximum value of the denoised rssi matrix, phase m,n The denoised label position is m, n phase value, phase min Phase being the minimum of the denoised phase matrix max Is the maximum value of the denoised phase matrix.
The steps have the advantages that: the difference of the rssi matrix numerical value and the phase matrix data under different time windows is reduced, and the method is favorable for better extracting characteristics from the input convolution neural network.
Step 3.5: and (3) respectively performing two-time upsampling on the normalized rssi matrix, the phase matrix and the binarization matrix obtained in the step (3.4) by using a bilinear difference value method, and enlarging the size of the matrix to be 2 times of the original size to obtain the processed rssi matrix, the phase matrix and the binarization matrix. The resulting data is divided into a training set and a test set.
The advantages of this step are: the size of the matrix is enlarged to 2 times of the original size by a bilinear difference method, and the improvement of the resolution of the matrix is facilitated. Due to the problems of the size and the deployment mode of the label, the obtained rssi matrix, the phase matrix and the binarization matrix are low in resolution, and the input of the rssi matrix, the phase matrix and the binarization matrix into the convolutional neural network is not favorable for extracting features. And 2 times of sampling is performed on the rssi matrix, the phase matrix and the binarization matrix upwards by a bilinear difference value mode, so that the resolution is improved, and the method is favorable for inputting the rssi matrix, the phase matrix and the binarization matrix into a convolutional neural network to extract three-dimensional posture information.
And 4, step 4: and (4) performing data enhancement on the rssi matrix, the phase matrix and the binarization matrix in the training set obtained in the step (3) to obtain an enhanced training set. The data enhancement specifically comprises turning and size change processing so as to enhance a data set and improve the stability and generalization capability of the three-dimensional sleeping posture skeleton detection. Comprising the following substeps:
step 4.1: and (3) performing horizontal mirroring on the rssi matrix, the phase matrix and the binarization matrix obtained in the step (3), and according to a horizontal mirroring formula:
a′ m,n ＝a m,n-k #(8)
wherein a' m,n And obtaining an rssi matrix, a phase matrix and a binarization matrix after horizontal mirroring, wherein the position of the label after mirroring is the value of m and n.
The steps have the advantages that: the data set is expanded, and the time consumption of data collection is saved. Referring to fig. 3, the first row of images is the data obtained in step 3.4 (source data) and the second row of images is the mirror-inverted data. Therefore, synchronous mirror image turning is kept on the rssi matrix, the phase matrix and the binarization matrix, mirror image turning is also carried out on the posture, the diversity of sleeping postures in data set is expanded, and the detection result of the convolutional neural network is favorably improved.
And 4.2: and (4) tag array transformation. Depending on the different sizes of beds, we can deploy tag arrays (28 × n) of different sizes to accommodate beds of different sizes. In the method, the experimental data collected by the user are the data collected under the condition that the size of the label matrix is 28 × 21, and the user simulates the generation of different sizes of the label matrix through the change of the label matrix with the size of 28 × 21, so that the method is suitable for the sizes of beds with different sizes.
In particular, for smaller tag arrays, the size is 28 × n, where n is<21, judging the two-dimensional mapping space of the sleeping posture through the area where the numerical value 1 in the binarization matrix obtained in the step 3.4 is located. Then, the left and right sides of the sleeping posture are searched according to the two-dimensional mapping space of the sleeping postureBoundary range, cut data outside the boundary, change the binarization matrix from 28 × 21 to 28 × n. And synchronously cutting the rssi matrix and the phase matrix obtained in the step 3.4 to obtain the rssi matrix and the phase matrix with the size of 28 x n. Then, respectively splicing two sides of the obtained rssi matrix, phase matrix and binarization matrix with the size of 28 x n
The rssi matrix, the phase matrix and the binarization matrix with the size of 28 × n are changed into the size of 28 × 21, so that the sizes of the input convolutional neural networks are ensured to be consistent.
The advantages of this step are: the data set is expanded, the time consumption of data collection is saved, and the sleeping posture RFID data of different beds are obtained for model training. See fig. 5 for a size transformed image. Through matrix transformation, matrix data with fixed size is converted into matrix data with different sizes, and then the matrix data is converted into matrix data with 28 × 21 sizes. The method is suitable for beds with different sizes, and high-precision three-dimensional sleeping posture detection results are kept. Referring to fig. 3, the first row of images is the data (source data) obtained in step 3.4, and the second row of images is the rssi matrix, phase matrix and binarization matrix data obtained after the label size transformation.
Step 4.3: and (4) repeating the process of the step (3.5) on the rssi matrix, the phase matrix and the binarization matrix obtained in the step (4.1) and the step (4.2) (namely performing double upsampling) to obtain the final enhanced rssi matrix, the phase matrix and the binarization matrix. The sizes of the input convolutional neural networks are ensured to be consistent.
The advantages of this step are: the resolution of the enhanced data is improved, and meanwhile, the input data size of the convolutional neural network is ensured to be consistent.
And 5: and (3) building a convolutional neural network model, performing data characteristic extraction on the data obtained in the steps (3) and (4) to obtain a three-dimensional sleeping posture, training to obtain a trained convolutional neural network model, and inputting the data to be tested into the model to obtain the three-dimensional sleeping posture. Comprising the following substeps:
step 5.1: a convolutional neural network is built in a PyTorch-based environment. Firstly, building a two-dimensional convolution neural network, introducing an attention mechanism model, then building a three-dimensional convolution neural network, connecting the three-dimensional convolution neural network and the two-dimensional convolution neural network together, and finally connecting the three-dimensional convolution neural network and the two-dimensional convolution neural network to a full connection layer. The output of the convolutional neural network is set.
As shown in fig. 5, the following contents are specifically included:
step 5.1.1, a layer of two-dimensional convolution layer is constructed, the size of a convolution kernel is 3 x 3, the filling is 1 x 1, the step length is 1, the number of output channels is 64, an activation function LeakyReLU is activated, and the data is normalized by using BatchNorm2 d. Then, the channel attention model is connected, and then six two-dimensional convolution layers are connected, so that the two-dimensional convolution and the attention mechanism are combined. The structure of the six two-dimensional convolutional layers: the convolution kernel size of the first convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 128, the function LeakyReLU is activated, the BatchNorm2d is used for normalizing the data, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the second convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the third convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU uses BatchNorm2d to carry out normalization processing on the data, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the fourth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 512, the function LeakyReLU is activated, and the data are normalized by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 1024, the function LeakyReLU is activated, and the data are normalized by using BatchNorm2 d; the convolution kernel size of the sixth convolution layer is 3 × 3, the padding is 1 × 1, the step size is 1, the number of output channels is 1024, the activation function LeakyReLU normalizes the data by using BatchNorm2d, and the maximum pooling layer with the size of 2 and the step size of 2 is used.
And 5.1.2, expanding the output of the two-dimensional convolution layer to three dimensions and connecting the two-dimensional convolution layer to six three-dimensional convolution layers. The structures of the six three-dimensional convolution layers are as follows: the convolution kernel size of the first three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 1025, the activation function LeakyReLU is used for carrying out normalization processing on data by using BatchNorm2 d; the convolution kernel size of the second three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 512, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the third three-dimensional convolution layer uses deconvolution, the convolution kernel size is 2 multiplied by 2, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU uses BatchNorm2d to carry out normalization processing on data; the convolution kernel size of the fourth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 128, the activation function LeakyReLU performs normalization processing on the data by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 64, the activation function LeakyReLU uses BatchNorm2d to perform normalization processing on the data; the convolution kernel size of the sixth convolution layer is 3 × 3 × 3, the padding is 1 × 1 × 1, the step size is 1, the number of output channels is 14, and the activation function Sigmoid is activated.
And 5.1.3, connecting the three-dimensional convolution layer to the full-connection layer to construct a convolution neural network model suitable for the method. The full connection layer is constructed by three layers, the first full connection layer has 4096 nodes, the activation function is ReLU, and Dropout is 0.5; the second fully connected layer has 4096 nodes with an activation function of ReLU and Dropout of 0.5; the first full-link layer has 42 nodes, and the output obtains the final result.
The advantages of this step are: and introducing a channel attention mechanism into the two-dimensional convolutional layer, combining the two-dimensional convolutional neural network with the attention mechanism, and extracting two-dimensional features from the rssi matrix, the phase matrix and the binarization matrix. And then converting the two-dimensional feature data into three-dimensional features, and further extracting the three-dimensional skeleton features better through the three-dimensional convolution layer. The model introduces an attention mechanism, and can better extract more useful information from the blocked labels and the unblocked labels in the rssi matrix, the phase matrix and the binarization matrix; then, the two-dimensional information is converted into the three-dimensional information, and a high-precision three-dimensional skeleton can be extracted in a layer progressive mode. The step determines the structure and specific parameters of the convolutional neural network, ensures the feasibility and reliability of the implementation of the method, and realizes the high precision and stability of the three-dimensional sleeping posture.
Step 5.2: a loss function in the training neural network is set while the network model is trained using an optimizer Adam of adaptive learning rate. Wherein the loss function is expressed as:
wherein L is m Representing the values lost in training the convolutional neural network. N represents the joint number, P i Three-dimensional coordinates representing the detected ith joint, G i Representing the three-dimensional coordinates of the ith joint in groudtuth.
The steps have the advantages that: the aim of the method is to optimize a neural network so as to obtain a high-precision three-dimensional sleeping posture detection result. And the loss function and the adam are used for restraining the training process of the neural network, so that the parameters of the neural network can be updated, and the training process of the neural network is optimized.
Step 5.3: and (4) training the constructed convolutional neural network model by using the enhanced training set obtained in the step (4) to obtain the trained convolutional neural network model.
Specifically, the method comprises the following steps: the input of the constructed convolutional neural network is three channels, namely an rssi matrix, a phase matrix and a binarization matrix. And (4) all data in the training set are iterated and traversed for 80 times, and finally a trained convolutional neural network model is obtained and used for detecting a three-dimensional sleeping posture result.
The advantages of this step are: and obtaining a trained convolutional neural network by using the enhanced data set, wherein the trained convolutional neural network is used for detecting the three-dimensional sleeping posture, and a high-precision and stable three-dimensional sleeping posture result can be obtained. Meanwhile, the trained convolutional neural network can be suitable for three-dimensional sleeping posture detection of beds with different sizes.
Step 5.4: and (5) inputting the trained model obtained in the step 5.3 by using the test set, and outputting a detected three-dimensional sleeping posture result.
Specifically, the method comprises the following steps: in order to test the stability and the application range of the trained convolutional neural network for detecting the three-dimensional sleeping posture, the test set comprises the divided test set data in the step 3.5. We also collect the sleeping posture data of different rooms and the sleeping posture data of different beds, and all use the data to test the convolutional neural network trained by the method, so as to evaluate the high precision and stability of the method for detecting the three-dimensional sleeping posture and the applicability of different rooms and different beds.
The advantages of this step are: the usability, high precision, stability and application range of the method are evaluated through the test set. The feasibility, reliability and stability of the method are verified.
The following is the effect of applying the method of the present invention to a data set to demonstrate the accuracy of its detection of three-dimensional sleeping posture.
(1) The table shows the basic information of the data set
TABLE 1
| Number of data set samples | Number of training samples | Number of test samples |
| 25000 | 20000 | 50000 |
We collected 36 people's data as a data set, with 29 people's data used to train the model and 7 people's data used to test the model.
(2) Evaluation criteria:
according to the specific implementation steps, the three-dimensional sleeping posture detection task is completed. We use the average per-joint position error MPJPE as a specific indicator of the three-dimensional skeleton. The formula for MPJPE is as follows:
E MPJPE (i)＝||Pi-Gi|| 2
wherein i represents a joint number, P i Three-dimensional coordinates representing the detected ith joint, G i Representing the three-dimensional coordinates of the ith joint in the groudtruth.
MPJPE is used for evaluating indexes of three-dimensional sleeping postures, the smaller the index is, the higher the precision of the detected three-dimensional sleeping postures is,
(3) And (4) analyzing results:
firstly, the method detects the three-dimensional sleeping posture skeleton through the RFID equipment, and whether the three-dimensional skeleton with fine granularity can be obtained or not is judged. Through the detection of the convolutional neural network, the following results can be obtained:
TABLE 2
| Joint | Error (cm) | Joint | Error (cm) |
| Head with a rotatable shaft | 6.88 | Neck part | 5.05 |
| Left shoulder | 7.05 | Left hip | 6.56 |
| Left elbow | 11.08 | Left knee | 7.63 |
| Left wrist | 16.44 | Left ankle | 10.06 |
| Right shoulder | 7.00 | Right hip | 6.60 |
| Right elbow | 10.48 | Right knee | 7.82 |
| Right wrist | 14.67 | Right ankle | 9.70 |
The final precision of the detection framework of the method is 9.07cm. As shown in Table 2, for the average error of each joint of the three-dimensional sleeping postures detected by the method, the data in the table can show that most joint errors can be kept within 10cm, and the joint errors of the elbow and the wrist are slightly larger. This is because the elbow and wrist degrees of freedom are relatively high compared to other joints, and because the arm is thin and has little effect on the signal, the error is larger than that of other joints. As shown in fig. 6, which is the final result of the method, we can see that the method can generate a three-dimensional sleeping posture with fine force.
Secondly, in order to study the robustness and generalization ability of the method, we applied the method to different sizes of beds and different rooms through a data enhancement method. We therefore compared the results of different rooms and different sizes of beds with and without data enhancement.
The following table shows the resulting accuracy of the sleeping positions of different rooms with and without data enhancement:
TABLE 3
As shown in table 3, it can be seen that, when data enhancement is not used in different rooms, the error is larger and the accuracy is lower, and after data enhancement is used, the error of two rooms is smaller and the accuracy is obviously improved. The data enhancement and improvement method improves the stability and generalization capability of the method, so that the method can be suitable for different rooms and can still keep high-precision detection results.
We collected sleeping posture data in the same room with different sizes of beds to assess the stability and applicability of the method. The size of the tag array can change along with the size of the bed, so the sleeping postures of the beds with different sizes have important influence on the result of the method. The following table shows the resulting accuracy of different bed sizes with and without data enhancement:
TABLE 4
| Error enhancement without data (cm) | Error after data enhancement (cm) | |
| 180cm | 13.05 | 10.07 |
| 150cm | 9.98 | 9.07 |
| 80cm | 29.26 | 9.34 |
Table 4 shows the results of the method of the present invention for three-dimensional sleeping postures under different bed sizes, and also shows the results of the method of the present invention for three-dimensional sleeping postures under different bed sizes after data enhancement. In the table, we can clearly see that the method without data enhancement is that we have the bed precision within 10cm at the size of 150cm, while the three-dimensional sleeping posture results are worse at the sizes of 180cm and 80cm, which indicates that the method can not accurately detect the three-dimensional sleeping postures of the beds with different sizes without data enhancement. After the data enhancement is used, the three-dimensional sleeping posture results of beds with the sizes of 180cm and 80cm are greatly improved and basically stabilized at about 10cm, and the data enhancement leads the method to have stable detection results for the three-dimensional sleeping postures of beds with different sizes.
Claims (11)
1. A three-dimensional sleeping posture detection method based on RFID equipment is characterized by specifically comprising the following steps:
step 1: deploying an RFID tag array and collecting data; synchronously collecting point cloud data;
and 2, step: data segmentation and elimination: dividing the RFID data collected in the step (1) by using a plurality of time windows, and counting the fluctuation size of the tag array signal of each time window to remove the data of sleep turning;
and 3, step 3: data preprocessing, namely performing RF image conversion and preprocessing on the data under each time window obtained in the step (2); dividing the obtained data into a training set and a test set;
and 4, step 4: performing data enhancement on the rsi matrix, the phase matrix and the binarization matrix in the training set obtained in the step 3 to obtain an enhanced training set;
and 5: and (3) building a convolutional neural network model, extracting data characteristics of the data obtained in the step (3) and the step (4), training to obtain a trained convolutional neural network model, and inputting the data to be tested into the model to obtain the three-dimensional sleeping posture.
2. The RFID-device-based three-dimensional sleeping posture detection method according to claim 1, characterized in that the operation of step 1 is as follows: deploying a 28 x 21 RFID tag array, paving the RFID tag array on a mattress, and placing an antenna above the mattress, wherein the distance between the antenna and the tag array is 1.6-2.3 m; meanwhile, kinect V2 is used for collecting synchronous point cloud information and marking the three-dimensional sleeping postures.
3. The RFID-device-based three-dimensional sleeping posture detection method according to claim 1, wherein the step 2 comprises the following sub-steps:
step 2.1: data segmentation: the RFID data collected in step 1 is divided in 3s time windows, and the data in each time window is represented as:
rssi m,n (t)＝{rssi 1 m,n ,rssi 2 m,n ,rssi 3 m,n ,...,rssi i m,n }#(1)
phase m,n (t)＝{phase 1 m,n ,phase 2 m,n ,phase 3 m,n ,...,phase i m,n }#(2)
wherein rssi m,n (t) rssi sequence value, phase, of the tag with m, n in the t-th time window m,n (t) indicates a sequence value of the phase of the tag at the position m, n in the t-th time window, and i indicates the number of times the tag at the position m, n is read in the t time windows;
step 2.2: data elimination: and calculating the variance sum of the rssi sequence and the phase sequence of all the non-blocking labels under each time window, eliminating the data of the time window larger than the threshold value, and keeping the data of the time window within the threshold value.
4. The three-dimensional sleeping posture detection method based on the RFID device according to claim 1, characterized in that in the step 2.2, the rssi variance sum threshold is 5, and the phase variance sum threshold is 8.
5. The RFID-device-based three-dimensional sleeping posture detection method according to claim 1, wherein the step 3 comprises the following sub-steps:
step 3.1: respectively solving the rssi sequence and the phase sequence of the non-blocking label in each time window obtained in the step (2) by using a formula 3 and a formula 4, and simultaneously assigning the rssi average of the blocking label to be-100 and the phase average to be 8; and then combining the label array information to obtain an rsi matrix and a phase matrix:
wherein the content of the first and second substances,
represents the mean of the rssi sequence with tag positions m, n under a certain time window,
means representing a phase sequence with m, n tag positions in a time window;
step 3.2: and (2) performing binarization processing on the rssi matrix obtained in the step (3.1), taking a threshold value as-100, taking the value of an element which is greater than the threshold value in the rssi matrix as zero, and taking 1 if the value is equal to the threshold value to obtain a binarization matrix:
wherein, b m,n Binary data which shows that the label position is m and n under a certain time window;
step 3.3: denoising the phase matrix and the rsi matrix obtained in the step 3.1 and the binarization matrix obtained in the step 3.2 respectively to obtain a denoised rsi matrix, a denoised phase matrix and a denoised binarization matrix;
step 3.4: normalizing the denoised rsi matrix and phase matrix obtained in the step 3.3 by using a formula 6 and a formula 7 respectively to obtain a normalized rsi matrix, a normalized phase matrix and a normalized binarization matrix:
wherein rssi m,n The denoised label position is the rssi value, rssi of m, n min Is the minimum of the denoised rssi matrix, rssi max Is the maximum value of the denoised rssi matrix, phase m,n The denoised label position is m, n phase value, phase min Phase being the minimum of the denoised phase matrix max The maximum value of the denoised phase matrix is obtained;
step 3.5: using a bilinear difference value method to respectively perform two-time upsampling on the normalized rssi matrix, the phase matrix and the binarization matrix obtained in the step 3.4, and enlarge the size of the matrix to be 2 times of the original size to obtain a processed rssi matrix, a processed phase matrix and a processed binarization matrix; the resulting data is divided into a training set and a test set.
6. The three-dimensional sleeping posture detection method based on the RFID device according to claim 1, characterized in that in the step 3.3, the denoising is performed by using Gaussian distribution function.
7. The RFID-device-based three-dimensional sleeping posture detection method according to claim 1, wherein the step 4 comprises the following sub-steps:
step 4.1: performing horizontal mirroring on the rssi matrix, the phase matrix and the binarization matrix obtained in the step 3;
and 4.2: carrying out size transformation on the data obtained in the step 3;
step 4.3: and (3) repeating the processing process in the step (3.5) on the rssi matrix, the phase matrix and the binarization matrix obtained in the step (4.1) and the step (4.2) to obtain the enhanced rssi matrix, the phase matrix and the binarization matrix.
8. The three-dimensional sleeping posture detection method based on the RFID device according to claim 7, wherein the step 4.3 specifically operates as follows: judging a two-dimensional mapping space of the sleeping posture according to the region where the numerical value 1 in the binarization matrix obtained in the step 3.4 is located; then, the left and right boundary ranges of the sleeping positions are searched according to the two-dimensional mapping space of the sleeping positions, data outside the boundaries are cut, and the binary matrix is changed from 28 x 21 to 28 x n, n<21; simultaneously, synchronously cutting the rsi matrix and the phase matrix obtained in the step 3.4 to obtain the rsi matrix and the phase matrix with the size of 28 × n; then, r of 28 x n will be obtainedRespectively splicing two sides of the ssi matrix, the phase matrix and the binarization matrix
The matrix of (2), the rsi matrix, the phase matrix and the binarization matrix of 28 × n size are changed to 28 × 21 size.
9. The RFID-device-based three-dimensional sleeping posture detection method according to claim 1, wherein the step 5 comprises the following sub-steps:
step 5.1: building a convolutional neural network in a PyTorch-based environment: firstly, building a two-dimensional convolution layer, introducing an attention mechanism, and then connecting the two-dimensional convolution layers to combine the two-dimensional convolution layer with the attention mechanism; then, the output of the two-dimensional convolution layer is expanded to three-dimensional and connected to a three-dimensional convolution layer of six layers; finally, connecting the three-dimensional convolution layer to a full-connection layer to construct a convolution neural network model;
step 5.2: setting a loss function in a training neural network, and simultaneously training a network model by using an optimizer Adam of a self-adaptive learning rate;
step 5.3: and (4) training the constructed convolutional neural network model by using the enhanced training set obtained in the step (4) to obtain the trained convolutional neural network model.
Step 5.4: and (5) inputting the trained model obtained in the step 5.3 by using the test set, and outputting a detected three-dimensional sleeping posture result.
10. The RFID-device-based three-dimensional sleeping posture detection method according to claim 9, characterized in that the step 5.1 comprises the following operations:
step 5.1.1, constructing a layer of two-dimensional convolution layer, wherein the size of a convolution kernel is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 64, an activation function LeakyReLU is activated, and the data is normalized by using BatchNorm2 d; then connecting the channel attention model and then connecting six two-dimensional convolution layers; the structure of the six two-dimensional convolutional layers: the convolution kernel size of the first convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 128, the function LeakyReLU is activated, the data is normalized by using BatchNorm2d, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the second convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the third convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 256, the activation function LeakyReLU uses BatchNorm2d to carry out normalization processing on the data, and the maximum pooling layer with the size of 2 and the step length of 2 is used; the convolution kernel size of the fourth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 512, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 1024, the activation function LeakyReLU is used for normalizing the data by using BatchNorm2 d; the convolution kernel size of the sixth convolution layer is 3 × 3, the filling is 1 × 1, the step length is 1, the number of output channels is 1024, the function LeakyReLU is activated, the data is normalized by using BatchNorm2d, and the maximum pooling layer with the size of 2 and the step length of 2 is used;
step 5.1.2, expanding the output of the two-dimensional convolution layer to three dimensions and connecting the two-dimensional convolution layer to six three-dimensional convolution layers; the six three-dimensional convolution layers have the following structures: the convolution kernel size of the first three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 1025, the activation function LeakyReLU is used for carrying out normalization processing on data by using BatchNorm2 d; the convolution kernel size of the second three-dimensional convolution layer is 3 multiplied by 3, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 512, the function LeakyReLU is activated, and the data is normalized by using BatchNorm2 d; the third three-dimensional convolution layer uses deconvolution, the size of a convolution kernel is 2 multiplied by 2, the filling is 1 multiplied by 1, the step length is 1, the number of output channels is 256, an activation function LeakyReLU is used, and the data is normalized by BatchNorm2 d; the convolution kernel size of the fourth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 128, the activation function LeakyReLU performs normalization processing on the data by using BatchNorm2 d; the convolution kernel size of the fifth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 64, the activation function LeakyReLU uses BatchNorm2d to perform normalization processing on the data; the convolution kernel size of the sixth convolution layer is 3 × 3 × 3, the filling is 1 × 1 × 1, the step length is 1, the number of output channels is 14, and the function Sigmoid is activated;
step 5.1.3, connecting the three-dimensional convolution layer to the full-connection layer to construct and obtain a convolution neural network model; the full connection layer is constructed by three layers, wherein the first full connection layer comprises 4096 nodes, the activation function is ReLU, and Dropout is 0.5; the second fully connected layer has 4096 nodes with an activation function of ReLU and Dropout of 0.5; the third full-connection layer has 42 nodes, and the output of the third full-connection layer obtains the final result.
11. The RFID-device-based three-dimensional sleeping posture detection method according to claim 9, wherein the loss function constructed in the step 5.2 is:
wherein L is m Representing loss values in the training convolutional neural network, N representing joint number, P i Three-dimensional coordinates representing the detected ith joint, G i Representing the three-dimensional coordinates of the ith joint in the groudtruth.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211018800.7A CN115660042A (en) | 2022-08-24 | 2022-08-24 | Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211018800.7A CN115660042A (en) | 2022-08-24 | 2022-08-24 | Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115660042A true CN115660042A (en) | 2023-01-31 |
Family
ID=84983517
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211018800.7A Pending CN115660042A (en) | 2022-08-24 | 2022-08-24 | Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115660042A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116030534A (en) * | 2023-02-22 | 2023-04-28 | 中国科学技术大学 | Training method of sleep posture model and sleep posture recognition method |
-
2022
- 2022-08-24 CN CN202211018800.7A patent/CN115660042A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116030534A (en) * | 2023-02-22 | 2023-04-28 | 中国科学技术大学 | Training method of sleep posture model and sleep posture recognition method |
| CN116030534B (en) * | 2023-02-22 | 2023-07-18 | 中国科学技术大学 | Training method of sleep posture model and sleep posture recognition method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP3716020B1 (en) | Systems and methods for three dimensional (3d) reconstruction of human gestures from radar based measurements | |
| US20210352441A1 (en) | Handling concept drift in wi-fi-based localization | |
| CN110517759B (en) | Method for determining image to be marked, method and device for model training | |
| Zhang et al. | Deep neural networks for wireless localization in indoor and outdoor environments | |
| CA2907723C (en) | Scalable real-time location detection based on overlapping neural networks | |
| CN105631436B (en) | Cascade position based on random forest returns the method for face alignment | |
| US20130028517A1 (en) | Apparatus, method, and medium detecting object pose | |
| WO2020240526A1 (en) | Proximity-based model for indoor localization using wireless signals | |
| CN104281835B (en) | Face recognition method based on local sensitive kernel sparse representation | |
| CN108597204A (en) | A kind of intelligent meter data recording system and its implementation | |
| CN105938513A (en) | Apparatus and method for providing reliability for computer aided diagnosis | |
| CN115660042A (en) | Three-dimensional sleeping posture detection method based on RFID (radio frequency identification) equipment | |
| Dai et al. | DMRF-UNet: A two-stage deep learning scheme for GPR data inversion under heterogeneous soil conditions | |
| CN113273998B (en) | Human body sleep information acquisition method and device based on RFID label matrix | |
| Cui et al. | Deep neural network based sparse measurement matrix for image compressed sensing | |
| Naveed et al. | Wavelet based multivariate signal denoising using mahalanobis distance and edf statistics | |
| CN117055004A (en) | Three-dimensional human skeleton estimation method based on millimeter wave radar sparse point cloud | |
| CN113688793A (en) | Training method of face model and face recognition system | |
| CN107622476B (en) | Image Super-resolution processing method based on generative probabilistic model | |
| CN108111973A (en) | A kind of indoor orientation method and device obtained based on real time fingerprint | |
| Deng et al. | Wildar: Wifi signal-based lightweight deep learning model for human activity recognition | |
| JP5450703B2 (en) | Method and apparatus for determining a spatial area in which a target is located | |
| Wu et al. | Device-free human activity recognition with identity-based transfer mechanism | |
| CN113740802B (en) | Signal source positioning method and system for performing matrix completion by using adaptive noise estimation | |
| CN109633531A (en) | Wireless sensor network node positioning system under composite noise condition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |