Passive target position and behavior sensing method based on RFID

Abstract

The invention discloses a passive target position and behavior sensing method based on RFID. The method comprises the following steps: firstly, constructing an intelligent sensing space based on RFID, and performing model construction in the intelligent sensing space by using a real-time signal vector between a real-time RSSI (Received Signal Strength Indicator) and a background RSSI; secondly, solving a weight matrix calculation module and a radio image reconstruction module in the model; obtaining the final position coordinate of the target based on a real-time signal vector and an upper and lower layer combined positioning mode; and finally, filtering the tag array in the intelligent sensing space, and extracting an effective tag set according to the position of the target. And establishing a radio image of a plane where the tags are located by using the simplified tag set, and performing behavior identification of upper and lower limb separation. According to the invention, the interaction signal between the tag array and the antenna array is utilized, and the radio tomography and 3D convolutional deep neural network technologies are combined, so that high-precision positioning and activity identification can be completed at the same time.

Description

Technical Field

[0001] The invention belongs to the field of wireless self-organizing networks and mobile radio frequency tag systems, and relates to a passive target position and behavior perception method based on RFID. Background Art

[0002] With the popularization of applications such as smart home, telemedicine and smart retail, people live in an increasingly convenient environment. The basis for the realization of these applications is to provide targeted services to people based on their location and activities. At the same time, the wide coverage and gradually lower prices of sensors have also promoted the development of wireless identification and perception technologies, among which passive perception methods are particularly widely studied. In this scenario, the target does not need to carry any additional equipment, and its precise location and behavior status can be obtained based on the signal obstruction or reflection effect generated by the target on the wireless link.

[0003] In traditional methods, various information collected by devices such as cameras and wearable sensors are generally used to perceive the state of the target. In particular, due to the significant improvement in computer performance and the development of image processing technology, computer vision-based methods have made significant progress. However, problems that plague such solutions, such as dependence on light and privacy restrictions in sensitive places, remain to be solved. Sensor-based wearable methods require the target to carry at least one sensor or signal transceiver to help with location perception and activity recognition. For example, inertial sensors equipped in smartphones can be well applied to pedestrian trajectory estimation and driving mode perception. By fixing RFID tags, Bluetooth or WiFi antennas, etc. on the target's torso, limbs or fingers, coarse-grained activity recognition and fine-grained motion tracking can be achieved. However, the disadvantages of this type of method are relatively obvious. First, the device needs to be carried anytime and anywhere and the battery must be sufficient. Second, it is more inconvenient for the elderly and children.

[0004] Considering the shortcomings of the above methods, there is an increasing number of studies on passive sensing solutions that do not require carrying equipment. In terms of passive positioning, based on parameters such as angle of arrival (AoA), time of flight (TF), and Doppler shift, the positioning accuracy has been improved from the initial meter level to the current decimeter level or even centimeter level. For passive behavior recognition, the perception granularity has been refined from the initial full-body torso activity recognition to gesture recognition, and heartbeat detection and sleep monitoring through wireless signals have also become a hot topic. Although the above two types of passive perception methods have been widely studied, few studies have been able to form a passive perception framework that covers both positioning and behavior recognition tasks. In addition, in the field of activity recognition, most previous technologies use 2D networks to extract full-body features for activity classification, which is not comprehensive enough for signals with high temporal correlation, and they are not flexible enough to add or delete activity categories. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a passive target location and behavior perception method based on RFID. The present invention constructs a real-time passive perception method based on RFID signals that can simultaneously cover the two tasks of positioning and activity recognition. It can achieve accurate and flexible activity recognition while completing high-precision positioning using a 3D convolutional neural network and upper and lower limb separation mechanism.

[0006] The method of the present invention comprises the following steps:

[0007] Constructing intelligent sensing space based on RFID;

[0008] Model building is performed using a real-time signal vector between the real-time RSSI and the background RSSI in the smart sensing space;

[0009] Solve the weight matrix calculation module and radio image reconstruction module in the model;

[0010] The weight matrix calculation includes calculating the ground weight matrix and the weight matrix of the plane where the label array is located;

[0011] The radio image reconstruction is based on the principle of radio tomography technology RTI, which uses weight matrix and real-time signal vector to solve the radio image;

[0012] The final position coordinates of the target are obtained based on the positioning mode combining the real-time signal vector and the upper and lower layers;

[0013] After obtaining the final position coordinates, the label array in the intelligent perception space is filtered to extract the valid label set according to the location of the target;

[0014] The simplified label set is used to build a radio image of the plane where the label is located, and the behavior recognition of upper and lower limbs is carried out separately.

[0015] Beneficial effects of the present invention: The present invention utilizes the interactive signals between the tag array and the antenna array, combined with radio tomography and 3D convolutional deep neural network technology, to simultaneously achieve high-precision positioning and activity recognition. Specifically:

[0016] The present invention designs a passive sensing method that can have both positioning and activity recognition functions. Different from the existing research that only aims at positioning or activity recognition, the present invention can provide the location of a person and his current action at the same time by using fine radio images.

[0017] The present invention proposes a positioning method combining upper and lower layers and an activity recognition mechanism separating upper and lower limbs, which can support high-precision positioning, accurate and flexible activity recognition, and fast and targeted data processing.

[0018] The experiments using commercial RFID equipment in a real environment show that the present invention can achieve an average positioning error of less than 0.22 m and an average activity recognition accuracy of more than 90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of a passive intelligent sensing space based on RFID;

[0020] Figure 2 A schematic diagram of the weight matrix for the ground plane and the label plane, where: (a) the ground plane, (b) the label plane;

[0021] Figure 3 The positioning radio images in three cases are shown in the figure: (a) upper link, (b) lower link, (c) upper and lower link combined;

[0022] Figure 4 To obtain the effective label set, in the figure: (a) label filtering process, (b) effective label set;

[0023] Figure 5 There are eight activities defined in the experiment;

[0024] Figure 6 Radio images of sitting and standing still, in the figure: (a) sitting still (original image), (b) sitting still (after foreground detection), (c) standing still (original image), (d) standing still (after foreground detection);

[0025] Figure 7 This is a diagram of the deep neural network structure based on 3D convolution;

[0026] Figure 8 This is the flow chart of pruning operation;

[0027] Fig. 9 It is the experimental deployment diagram;

[0028] Fig.10 The positioning results are shown in the figure: (a) the overall average positioning error, (b) the effectiveness of the upper and lower layer combined positioning mode, (c) the average positioning error of 6 activities, and (d) the average positioning error of 8 volunteers;

[0029] Fig.11 Confusion matrix for activity recognition;

[0030] Fig.12 Behavior recognition result diagram, showing: (a) the effectiveness of the upper and lower limb separation model, (b) the average recognition accuracy of 8 volunteers. DETAILED DESCRIPTION

[0031] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0032] Since the passive positioning and behavior recognition method proposed in this application is based on the radio frequency tomography technology (Radio Tomographic Imaging, RTI) proposed by Wilson et al. in 2010, the basic principle of this technology is first analyzed, and then the detailed process of converting the real-time received signal strength of RFID into a radio image based on this technology is given, and finally the positioning and activity recognition mode proposed in the present invention is given.

[0033] 1. Basic principles of radio frequency tomography (RTI)

[0034] Radio Tomographic Imaging (RTI) was proposed in 2010. It is a technology used to image the attenuation caused by physical objects in wireless networks. It uses WiFi signals to passively determine the location of people within the sensing range. The following is an introduction to the basic principles of RTI technology and its derivation process:

[0035] When wireless nodes communicate with each other, radio signals pass through the physical area of ​​the network and are absorbed, reflected, diffracted or scattered by objects in the area. The goal of RTI is to determine an image vector that describes the attenuation of radio power caused by physical objects in the area. Since the location of the pixel is known, it is possible to find where the attenuation occurs in the network, that is, where the occluding object is located.

[0036] Let K be the number of nodes in the network. If all nodes are considered to have normal links regardless of whether they communicate with each other, the total number of bidirectional links is The signal strength y of a specific link i at time t i (t) depends on the transmission power p i ; Shadow loss S caused by signal attenuation due to object occlusion i (t); fading loss F due to active or passive interference of narrowband signals in a multipath environment i (t); Static loss L due to distance, antenna pattern, and equipment inconsistency i ; Measurement noise v i (t).

[0037] Theoretically, the received signal strength can be defined as:

[0038] y i (t) = p i -L i -Si (t)-F i (t)-v i (t) (1)

[0039] Furthermore, the shadow loss S i (t) can be approximated as the sum of the attenuation occurring in each pixel. Since each pixel contributes differently to the link attenuation, a weighted approach is required, which can be expressed as:

[0040]

[0041] Among them, x j (t) is the attenuation of pixel j at time t; w ij is the weight of pixel j under link i.

[0042] Imaging only the changing attenuation greatly simplifies the problem, since all static losses can be cancelled out over time. a to b , signal intensity change Δy i for:

[0043] Δy i =y i (t b )-y i (t a )=S(t b )-S i (t a )+F i (t b )-F i (t a )+V i (t b )-V i (t a ) (3)

[0044] (3) can be rewritten as:

[0045]

[0046] where Δx j =x j (t b )-x j (t a ) is from time t a to b The fading variation of pixel j. Noise is a combination of fading and measurement noise:

[0047] n i =F i (t b )-Fi (t a )+V i (t b )-V i (t a ) (5)

[0048] If all links in the network are considered simultaneously, the received signal strength indicator (RSSI) can be systematically represented by a matrix as follows:

[0049] Δy=WΔx+n (6)

[0050] in,

[0051]

[0052] n=[n1,n2,…,n m ] T ,

[0053] [E] i,j =w ij .

[0054] In summary, Δy is the measurement of the RSSI change of all links of length M, n is the noise vector, and Δx is the fading image to be estimated. W is an M×N weight matrix, where each column represents a pixel and each row represents the weight of each pixel under a specific link.

[0055] When calculating the weight matrix, ellipse-based range determination is an available method. Simply put, if a specific pixel falls outside the ellipse determined by the two wireless nodes, the weight of the pixel is set to 0; if the specific pixel is within the LOS path determined by the ellipsoid, its weight is set to be inversely proportional to the square root of the link distance. Intuitively, longer links provide less pixel attenuation information associated with them, because when the link distance is long, the signal will be reflected and diffracted around obstacles; and when obstacles are present, the RSSI changes that occur in relatively short-distance links will be greater. Studies have shown that the variance of link fading does not change with distance, so considering the inverse of the square root of the link length, the weight can be described as:

[0056]

[0057] Where d is the distance between two wireless nodes; d ij (1) d ij (2) are the distances between pixel j and two nodes under link i; λ is the parameter for adjusting the width of the ellipse.

[0058] 2. Radio Image Extraction

[0059] In order to effectively utilize the occlusion effect of the human body to simultaneously complete the two tasks of position determination and activity perception, the embodiment of the present application constructs an RFID-based intelligent perception space, such as Figure 1 Specifically, a tag array including W RFID tags Pasted on the wall, a two-layer antenna array consisting of P antennas Deployed opposite the tag array, forming L = P * Q links. r,c , a r,c represents the tag,antenna located at the rth row and cth column,q r ,q c is the number of rows and columns of the label array, p r , p c is the number of rows and columns of the antenna array, then the RSSI measurement value of the mth round can be expressed as:

[0060]

[0061] in, is the RSSI value collected by antenna i in the mth round; is the RSSI value of the link corresponding to antenna i and tag j.

[0062] Combined with RSSI measurements when no one is in the space R m The difference between R0 and △R m = {R m -R0}.

[0063] Due to changes in environmental factors or human status, even if the same person performs the same activity, the recognition pattern generated using the RSSI absolute value will become completely different. However, using the difference △R between the real-time RSSI and the background RSSI m Model building can amplify the impact of the activity itself on the signal while greatly reducing the influence of external factors. m It is more robust and more suitable for perceptual model building.

[0064] Summarizing the principle of radio frequency tomography RTI, its solution process can be linearly modeled as:

[0065] △R m =W△X m +n (10)

[0066] Where W is the weight matrix of the pixel plane; △X mis the radio image to be estimated; n is the noise vector. Therefore, RTI contains two main parts: a weight matrix calculation module for W calculation and a weight matrix calculation module for △X m Computational radio image reconstruction module.

[0067] In one embodiment, the weight matrix is ​​calculated as follows:

[0068] In order to meet the different requirements of positioning and activity recognition, this embodiment first divides the ground and the plane where the tag array is located into J G , J T pixels, and then calculate the weight matrix respectively.

[0069] Figure 2 A schematic diagram of the weight matrix establishment for the plane where the ground and label array are located is given. Specifically, an ellipse-based weight distribution method is used for the ground plane, such as Figure 2 As shown in (a). For antenna A i and label T j The link l between them has a weight w l It is based on the relationship between each pixel and the focus at A i and T j The relative position between the ellipses is determined by the equation:

[0070]

[0071]

[0072] in, represents the weight value of pixel k in link l; d ij Yes A i With T j The distance between ki ,d kj Respectively represent the distance from the center of pixel k to A i , T j The distance; G is an adjustable parameter used to control the range of the ellipse. In other words, the above operation sets the weight of pixels outside the ellipse to 0, and sets the weight of pixels inside the ellipse to

[0073] For the plane where the label array is located, the weight of each pixel in a particular link depends only on its distance to the corresponding label. Figure 2 As shown in (b), the range of pixels affected by the signal between a specific antenna and the tag is set to be centered on the tag with a radius of ζ T The circle is calibrated using the binarization method to calibrate the weights of each pixel l corresponding to the link l as follows:

[0074]

[0075] Note that since the weight matrix used for positioning is established on a plane parallel to the ground and at the same height as the antenna, this embodiment only uses the link L between the corresponding single-layer antenna and the single-layer tag. c =p c ×q c .

[0076] Following the principle that the closer the pixel is to the link, the greater its weight and contribution, the two weight matrices W can be calculated through the above process. G , W T As shown below, it is further used to obtain the radio image required for positioning and activity recognition:

[0077] W G ={w l |l=1,2,...,L c},W T ={w l ′|l=1,2,...,L} (15)

[0078] In one embodiment, the radio image is reconstructed as follows:

[0079] Based on the linear model of formula (10), radio image reconstruction is actually to minimize the noise vector n and find the optimal solution of the following equation The problem:

[0080]

[0081] However, (16) is an ill-posed inverse problem and cannot be solved simply by setting it to 0. Therefore, Tikhonov regularization can be used to solve this problem. Specifically, by adding a restriction term, the equation can be expressed as the objective function as follows:

[0082] f(x)=||W△X m -△R m || 2 +α||Ψ△X m || 2 (17)

[0083] Among them, α represents the weight value; Ψ is the Tikhonov matrix. Considering that for a 2D image, two vertical and horizontal difference operators can be introduced, namely Γ x , Γ y , (17) can be rewritten as:

[0084] f(x)=||W△X m -△R m ||2 +α(||Γ x △X m || 2 +||Γ y △X m || 2 ) (18)

[0085] The estimated image can be given by setting equation (18) to 0 as follows:

[0086]

[0087] Among them, α represents the weight value, and I represents a unit vector in the standard orthogonal basis;

[0088] In summary, after constructing the weight matrices of the ground and tag planes according to the arrangement of antennas and tags in space, two radio images can be calculated through operations between real-time signal vectors and weight matrices to identify the location and activity of users.

[0089] In one embodiment, the positioning mode design is performed as follows:

[0090] (1) Positioning mode design combining upper and lower layers

[0091] Considering that the occlusion effects of different parts of the human body, especially the upper and lower limbs, are different during movement, the positioning results calculated using only a single-layer antenna and a single-row tag are not accurate enough. Therefore, this embodiment considers combining the positioning results of the upper and lower layers to obtain the final coordinates. Specifically:

[0092] Given the upper and lower antennas are and The corresponding label rows at the same height can be expressed as and Based on the above two layers of links, a positioning mode combining the upper and lower layers can be realized.

[0093] First, we need to capture △R m The two subsets of and Represent the difference between the real-time signal and the background signal of the upper link and the lower link respectively, and then combine the weight matrix W G , use the following formula to calculate the upper image and the underlying image

[0094]

[0095] In order to combine the results of the upper and lower layers, the pixel values ​​of the two images are added to obtain the superimposed image. Then, the final position (X, Y) of the target can be calculated by selecting the first N largest pixel values ​​and weighting them as follows:

[0096]

[0097] Among them, μ i 、(x i ,y i ) are the value and coordinates of pixel i respectively.

[0098] For example, Figure 3 The radio image is given when the target is actually located at (0.6, 1.2). Obviously, the human occlusion is reflected in the brighter color of the pixels around the specific link, while the farther pixels appear darker. Figure 3 As shown in (a) and (b), if the coordinates of the brightest pixel in the image are regarded as the location of the target, the positioning errors of the upper link and the lower link are 0.158m and 0.112m respectively. However, if the results of the two layers are superimposed in a weighted average manner, the positioning accuracy is improved to 0.05m, as shown in Figure 3 As shown in (c).

[0099] (2) Obtaining a valid tag set

[0100] After obtaining accurate positioning results, the next step is to filter the tag array to extract a valid tag set based on the person's location. This can reduce the noise caused by useless tags and speed up the processing of the present invention.

[0101] Figure 4 The process and results of obtaining the effective label set are given. Figure 4 As shown in (a), assuming that antenna A is on the ground plane i The coordinates of (x i ,0); the coordinates of the person are (x,y); the vertical distance between the tag and the antenna is d. The person is regarded as a cylinder with a radius of r. After superimposing the positioning error ε and projecting it onto the ground plane, the occlusion range of the person can be regarded as a circle with a radius of r+ε.

[0102] First, assume that antenna A i The projection on the plane where the label array is located (hereinafter referred to as PT) is B i ; by A i The two tangent lines of the originating circle intersect PT at point Q. Li and Q Ri The two perpendicular tangent lines intersect at point P. Li and P Ri It is tangent to the circle at point C. Li and C Ri Intersects with PT at. Connect A i PLi and A i P Ri And extend it to intersect PT at point G Li and G Ri , these two points can be used to approximate Q Li and Q Ri , to indicate the antenna A i The corresponding left and right boundaries of valid labels. Given B i , C Li and C Ri The coordinates of (x i ,d), (xr-ε,d) and (x+r+ε,d), then G Li and G Ri The x-coordinate of Li 、x Ri ; can be based on △G Li C Li P Li ~△G Li B i A i and △G Ri C Ri P Ri ~△G Ri B i A i The characteristics of these two sets of similar triangles are solved according to the following formula:

[0103]

[0104] In the pair with antenna A i Another antenna A on the same floor j After completing the above operations, A j The corresponding left and right boundaries of the valid label x Lj 、x Rj The x coordinate of can also be obtained in the same way. i , A j The boundary of the two antennas, the range of the target tag can be calculated as:

[0105] x L =min(x Li ,x Lj ),x R =max(x Ri ,x Rj ) (25)

[0106] where [x L ,x R ] represents the final range of valid labels, from which a subset T of the label array can be obtained D ,like Figure 4 As shown in (b), we consider the label set TD The link with the antenna can effectively distinguish human behavior.

[0107] In one embodiment, the behavior recognition model is designed as follows:

[0108] According to formula (19) and using the simplified label set T D Create a radio image of the plane where the tag is located After that, it can be used for subsequent behavior recognition tasks. Since large-scale gestures are the prerequisite for more fine-grained action recognition, and most abnormal behaviors have obvious gesture changes. Therefore, the embodiment selects eight relatively coarse-grained actions formed by the combination of upper limbs and lower limbs to verify the effectiveness of the proposed behavior recognition scheme, such as Figure 5 shown.

[0109] (1) Determination of human height status based on foreground detection

[0110] Considering that sitting or standing posture is the most basic element of a person's movement, it is necessary to first Height judgment is performed to guide the division of sub-images corresponding to the upper limbs and lower limbs. Prior to this, in order to eliminate dynamic noise (background pixels) while fully highlighting the RSSI changes caused by human motion (foreground pixels), the foreground detection method in the field of image processing was borrowed, which mainly includes a background model building module and a foreground detection module.

[0111] In the background model building module, according to the basic idea of ​​the Gaussian mixture model, each pixel in the image is first represented as a mixture of K Gaussian distributions. Then, the value of the current pixel observed at time t is X t The probability is:

[0112]

[0113] Among them, ω i,t is the estimate of the i-th Gaussian weight in the mixture model at time t; η is a Gaussian probability density function, which can be expressed as:

[0114]

[0115] Among them, μ i,t and are the mean and covariance matrix of the i-th Gaussian distribution in the mixed model at time t. i,t σ i,t After sorting, the first N distributions are selected as the background model. N can be calculated as follows, and the threshold P represents the minimum prior probability of background noise in the trace:

[0116]

[0117] Furthermore, in the foreground detection module, foreground pixels can be detected by calculating the distance between the pixel and N distributions, and marking the pixels beyond the standard deviation as foreground pixels. Figure 6 Taking the two postures of sitting still and standing still as examples, the effect of using the foreground detection method is given. Figure 6 (a) and (c) show examples of the original radio images calculated, with distributions representing two actions: sitting still and standing still. After foreground detection and grayscale processing, Figure 6 The results in (b) and (d) illustrate the effectiveness of this method in removing background noise.

[0118] The image after foreground detection and grayscale processing is represented as from Figure 6 From the comparison between (b) and (d), it can be found that if the person is sitting, the highlighting effect caused by the occlusion is concentrated in the middle and lower part of the figure; when the person is standing, this effect is dispersed on the y-axis. Inspired by this observation, this embodiment first The sum of the pixel values ​​is calculated as V, and then the image is evenly divided into two parts, the upper and lower parts. Then the sum of the pixel values ​​of the two parts is further calculated as V u and V l ,definition is the total pixel V l and V u The height status of the target can be determined by the difference in the proportions of V Whether it exceeds the threshold ε (here 0.2) is determined. In other words, diff V ≥ε means that the person is more likely to be sitting; otherwise, the person is more likely to be standing.

[0119] (2) Design of deep neural network based on 3D convolution

[0120] Figure 5 There are eight actions defined in , each of which is a combination of the following two categories of detailed actions:

[0121] Upper limbs: Still, Bow, Wave, Swing; Lower limbs: Sit, Stand, Walk.

[0122] Considering that the sitting and standing of the lower limbs belong to the same type as the stillness of the upper limbs, both represent the characteristic of the limbs being immobile, therefore, sitting and standing are included in the stillness and classified into one category, and finally five training sets are formed: stillness, bending, shaking, swinging and walking.

[0123] When selecting a deep learning algorithm, after investigating existing studies, it was found that most tend to use 2D convolutional neural networks to complete classification tasks. Most of these models are effective, but their biggest problem is that they will cause the loss of useful time information, which is one of the key factors that determine the accuracy of activity recognition classification. Therefore, this embodiment designs a deep neural network based on 3D convolution to extract spatiotemporal features across frames to identify detailed movements of upper and lower limbs.

[0124] The designed network consists of an input layer, a hidden layer, and an output layer. Its network structure is as follows: Figure 7 As shown, in the input layer, first Divided into two sub-images and Corresponding to the upper limbs and lower limbs respectively. Extract S consecutive frames (here 8) from the same type of sub-images and pack them as the input of the network, expressed as:

[0125]

[0126] Define f as the designed network, and the goal of the problem can be expressed as:

[0127] P u =f(I T_u ),P l =f(I T_l ) (30)

[0128] Among them, P u and P l These are the probability distributions of the five types of actions corresponding to the upper limbs and lower limbs.

[0129] In the hidden layer, in order to better capture effective features, the hidden layer contains the following three modules: The edge feature extraction module first extracts the spatiotemporal features of continuous frames and expands the receptive field, which is mainly composed of 4 convolution blocks and 3 maximum pooling layers. Among them, the convolution block consists of a convolution layer, a batch normalization layer, and a ReLU layer, while the maximum pooling layer can expand the receptive field and reduce the training cost. The semantic feature processing module uses a skip connection mechanism, which can increase the depth of the network while avoiding the problem of gradient disappearance, thereby further extracting semantic information of multiple frames from the spatiotemporal features. The fully connected module finally maps the distributed feature representation to the sample space and provides the output of the hidden layer, thereby greatly reducing the influence of the feature position on the classification.

[0130] In the output layer, the probability distribution of five types of actions can be obtained from the output of the hidden layer, which is the basis for subsequent inference of whole-body activities.

[0131] In particular, this embodiment uses a binary cross entropy (BCE) function as a loss function to guide model training, and its mathematical expression is as follows:

[0132] L=BCE(P,label) (31)

[0133] Among them, P is the probability distribution of the network output; "label" is the label information corresponding to the input sample, that is, the real category; L is the loss function. Because the target variable of the classification problem is discrete, and the target variable of the regression problem is continuous. The cross entropy describes the distance between two probability distributions. The smaller the value of the cross entropy, the closer the two are.

[0134] (3) Pruning operation and recognition result output

[0135] Since only a general model was trained in the recognition process of upper and lower limb movements, a decision tree-based solution was subsequently used to obtain the overall activity category. The specific decision process is as follows Figure 8 As shown. At the first level, the current state of the person is first distinguished as sitting or standing, which is achieved by determining the height of the human body. If it is judged to be standing, the sample is further classified by checking whether the lower limbs are recognized as stationary or moving. Finally, the category with the highest probability is the final result corresponding to the upper or lower limbs, and the two parts are combined to form a judgment of the overall activity. By utilizing the above operations, the confusion of sitting / standing posture and upper / lower limbs can be avoided, and the recognition accuracy is improved while reducing the computational overhead.

[0136] Verification example: The experimental environment is deployed as follows Fig. 9 As shown, the spacing between adjacent tags is set to 0.2m, and the size of the overall tag matrix is ​​2.2m×1.4m. The RFID antenna is placed on the same vertical plane 1.8m away from the tag array, with the heights of the upper and lower antennas being 1.2m and 0.6m respectively, and the two antennas on the same layer being 0.75m away from the left and right boundaries. In order to verify the effectiveness of the embodiments of the present application in activity positioning and identification, 8 volunteers (6 males and 2 females of different heights and weights) were invited to participate in the experiment, and 12 corners of 0.3m×0.3m floor tiles were selected as reference positions.

[0137] Specifically, for the two walking-related activities, each volunteer can walk freely within the specified range for 30 minutes, while for the remaining sitting / standing activities, all volunteers are required to repeat 6 specified activities at 12 reference positions for 1 minute each. The total data collection time for all eight activities is 1056 minutes, including 8×2×30=480 minutes of walking and 8×6×12=576 minutes of sitting / standing activities. After the collected activity data are subdivided into upper and lower limbs, the training and test data of each sub-action are obtained as shown in Tables 1 and 2. In particular, in the activity recognition stage, after the entire image is divided into two parts corresponding to the upper and lower limbs, the same half-body samples need to be combined to form a data set for deep network training. When evaluating the performance of the method, the average positioning error and recognition accuracy are mainly used to give the effects of positioning and activity recognition respectively.

[0138] Table 1 Training data table

[0139] Action Details Swing Wave Still Sit Bow Stand Walk quantity 3688 3751 8312 8085 3930 7974 8265

[0140] Table 2 Test data table

[0141]

[0142]

[0143] Positioning accuracy analysis of the verification example:

[0144] Considering that two walking-related activities do not require precise positioning, Fig.10 The average localization errors for the remaining six activities are given in . Fig.10 (a) shows the overall positioning effect of the method. For all 12 reference positions, the positioning error is less than 0.260m, and the total average positioning error is 0.186m. It can also be found that the positioning error of positions 8 and 9 reaches 0.250m, which is slightly larger than other positions. By comparing the real coordinates of each position, it is found that these two positions are in the middle and close to the antenna area, which will cause the signals of a large number of tags to be blocked by the human body, resulting in positioning distortion. However, from the overall results, the embodiment of the present application has stable and accurate positioning performance.

[0145] In order to verify the necessity of using the positioning mechanism combining the upper and lower layers, this experiment compared the positioning results of three cases (upper layer only, lower layer only, and upper and lower layers combined). Fig.10As can be seen in (b), when only the upper or lower antennas are used, the average positioning error is 0.233m and 0.192m, respectively. The difference in accuracy between the two layers may be due to the relatively large amplitude of the upper torso and limbs compared to the basically stationary lower limbs, resulting in unstable signals used for upper layer positioning. Looking further, when the two layers are superimposed, the average positioning error drops to 0.185m, which shows the effectiveness of the combined mechanism proposed in this application. Fig.10 (c) shows the overall positioning accuracy results for different activities. It can be found that the positioning of the present invention is robust to the diversity of activities and can achieve an average positioning error of less than 0.2m.

[0146] In the experiment, the effect of individual diversity on positioning performance is also studied. Fig.10 As shown in (d), the overall positioning accuracy of the eight volunteers remained at a high level of [0.174m-0.212m]. According to the height and weight records of the volunteers, it was found that the volunteers with relatively high positioning errors had relatively unusual body shapes (e.g., Volunteer 1: 156cm54kg, Volunteer 7: 183cm86kg). In general, a high positioning accuracy can be maintained under the condition of individual diversity.

[0147] Analysis of the behavior recognition effect of the verification example:

[0148] Next, the accuracy of the present invention in activity recognition is evaluated. First, the overall behavior recognition effect is analyzed. Table 3 lists the detailed classification results of upper and lower limbs corresponding to each activity. Table 3 Upper and lower limb discrimination results

[0149]

[0150] From Table 3, we can observe the following preliminary results:

[0151] Nearly perfect classification results were obtained for all categories corresponding to the lower limbs;

[0152] For the categories corresponding to the upper limbs, the performance of the present invention is moderate in “bending” and “swaying”;

[0153] For the two categories of walking-related activities, the classification results for the upper limbs were relatively poor.

[0154] Fig.11A confusion matrix for classifying eight activities is given, and the recognition results can be divided into three groups. The first group contains activities 01, 02, 04 and 06, and the activity recognition accuracy of this group exceeds 90%, and three of them are even close to 100%. In the second group, the recognition accuracy of activities 03 and 05 exceeds 80%. In particular, the main reason for the misclassification is that both include upper limb movements such as bending and shaking. The last group contains activities 07 and 08. Since the signal changes caused by arm swinging are negligible in human walking, it is impossible to distinguish between walking with arms swinging and walking with pockets. But overall, although the recognition accuracy in both cases will be reduced, it is still close to 80%. Therefore, through the deep network and pruning operation designed in this application, the ideal activity recognition accuracy can be achieved.

[0155] In order to verify the effectiveness of the activity recognition mechanism using the upper and lower limbs separately compared with the whole body data for activity recognition, the designed deep network is used to directly train and classify the whole body data of eight activities. The comparison of the two modes is shown in Figure 2. Fig.12 As shown in (a). Intuitively, the separation mode adopted makes the classification result more stable, with an average recognition accuracy of 94.6%. Although the accuracy of the full-body recognition scheme is slightly higher than that of the method of the present application by 4.2%, 7.4%, 3.7% and 1.7% for activities 03, 05, 07 and 08, the accuracy gap between the two modes is obvious for the other three activities 02, 04 and 06. In particular, for activity 04, the recognition accuracy of the full-body mode is only 51.9%, which is 40.2% lower than that of the method of the present application.

[0156] The impact of individual diversity on activity recognition performance is as follows Fig.12 As shown in (b), it can be found that the present invention can achieve an accuracy of about 90% for most volunteers. In particular, for volunteer 3, the recognition accuracy can reach 96.1%, while volunteer 7 has the lowest recognition accuracy of 87.2%. Therefore, it can be said that the activity recognition can be well extended to different users.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A passive target location and behavior perception method based on RFID, characterized in that The method comprises the following steps: Constructing intelligent sensing space based on RFID; Model building is performed using a real-time signal vector between the real-time RSSI and the background RSSI in the smart sensing space; Solve the weight matrix calculation module and radio image reconstruction module in the model; The weight matrix calculation includes calculating the ground weight matrix and the weight matrix of the plane where the label array is located; The radio image reconstruction is based on the principle of radio tomography technology RTI, which uses weight matrix and real-time signal vector to solve the radio image; The final position coordinates of the target are obtained based on the positioning mode combining the real-time signal vector and the upper and lower layers; After obtaining the final position coordinates, the label array in the intelligent perception space is filtered to extract the valid label set according to the location of the target; The simplified label set is used to build a radio image of the plane where the label is located, and the behavior recognition of upper and lower limbs is carried out separately.

2. The passive target location and behavior perception method based on RFID according to claim 1, characterized in that: The construction of the RFID-based intelligent sensing space specifically includes: Step 1-1: a tag array including W RFID tags Deployed on the wall, where t r,c represents the label at row r and column c; q r ,q c is the number of rows and columns of the label array; Step 1-2: The two-layer antenna array including P antennas Deployed opposite the tag array, where a r,c represents the tag,antenna located at the rth row and cth column; p r , p c is the number of rows and columns of the antenna array; Step 1-3: Based on this, L=P*Q links are formed.

3. The passive target location and behavior perception method based on RFID according to claim 2, characterized in that: The model construction using the real-time signal vector between the real-time RSSI and the background RSSI is specifically: Step 2-1: Assume the real-time RSSI measurement value R of the mth round m It is expressed as: in, is the RSSI value collected by antenna i in the mth round; is the RSSI value of the link corresponding to antenna i and tag j; Step 2-2: RSSI measurement value when there is no target in the space Step 2-3: Calculate R m The difference between R0 and R1 is the real-time signal vector △R m .

4. The RFID-based passive target location and behavior perception method according to claim 1, characterized in that: The weight matrix calculation module performs the following operations: Step 3-1: Divide the plane where the ground and the tag array are located into J G , J T pixels, and calculate the weight matrix W G , W T ; Step 3-2: Use the ellipse-based weight distribution method for the ground plane. For antenna A i and label T j The link l between them, with the focus on A i and T j Based on the ellipse, for J G The weights of pixels outside the ellipse are set to 0, and the weights of pixels inside the ellipse are set to Among them, d ij Yes A i With T j The distance between Step 3-3: For the plane where the tag array is located, the weight of each pixel in the link depends only on its distance to the corresponding tag. i and label T j The signal influence pixel range is set to be centered on the label with a radius of ζ T The circle uses the binarization method to calibrate the weights of each pixel corresponding to the link l; Step 3-4: Use the link L between the corresponding single-layer antenna and the single-layer tag c =p c ×q c ; Step 3-5: Following the principle that the closer the pixel is to the link, the greater its weight and contribution, two weight matrices W under all links are calculated through steps 3-2, 3-3 and 3-4. G , W T .

5. The passive target location and behavior sensing method based on RFID according to claim 4, characterized in that: In step 3-3, for J T pixels, if the pixel is j The distance between them is less than or equal to ζ T , then its weight is set to 1, otherwise the weight is set to 0.

6. The passive target location and behavior sensing method based on RFID according to claim 1, characterized in that: The radio image reconstruction module performs the following operations: The weight matrix W at a given ground plane G and the weight matrix W of the label plane T After that, the weight matrix and the real-time signal vector △R m The operation between them results in two radio images Used to identify user's location and activities: Among them, α represents the weight value, and I represents the unit vector in the standard orthogonal basis.

7. The passive target location and behavior sensing method based on RFID according to claim 1, characterized in that: The positioning mode of the upper and lower layers is specifically: Step 5-1: Given the upper and lower antennas are and Corresponding label rows at the same height are represented as and Step 5-2: Capture real-time signal vector △R m The two subsets of and Represents the difference between the real-time signal and the background signal of the upper link and the lower link respectively; Step 5-3, based on the weight matrix W G , combined with and Calculate the upper image and the underlying image Step 5-4: Put the two upper images and the underlying image The pixel values ​​are added to obtain the superimposed image Step 5-5: The final position (x, y) of the target is obtained by weighted calculation based on the coordinates of the first N maximum pixel values.

8. The passive target location and behavior sensing method based on RFID according to claim 1, characterized in that: The above-mentioned behavior recognition of separating upper and lower limbs specifically includes determining the height state of a human body based on foreground detection, designing a deep neural network based on 3D convolution and pruning operations, and outputting recognition results; The human height state determination based on foreground detection is specifically: Step 7-1, using foreground detection and grayscale processing methods in the field of image processing to eliminate dynamic noise while fully highlighting the RSSI changes caused by the person's own movement; Step 7-2: The image after foreground detection and grayscale processing is represented as Will The sum of the pixel values ​​is calculated as V, and then the image is evenly divided into two parts, the upper and lower parts; Step 7-3: Calculate the sum of the pixel values ​​of the upper and lower parts as V u and V l ; define diff V is the total pixel V l and V u The difference in proportions of each, the height status of the target is determined by diff V It is determined whether it exceeds the threshold ε.

9. The RFID-based passive target location and behavior perception method according to claim 8, characterized in that: The 3D convolution-based deep neural network performs the following operations: Step 8-1, include sitting and standing in the same category as static, and finally form five training sets: static, bending, shaking, swinging and walking; Step 8-1: In the input layer, Divided into two sub-images and Corresponding to upper and lower limbs respectively; Step 8-2: Extract S consecutive frames from the sub-graphs of the same type and package them as the input of the network; Step 8-3, in the hidden layer, includes the following three modules: The edge feature extraction module extracts the spatiotemporal features of consecutive frames and expands the receptive field. It mainly consists of 4 convolution blocks and 3 maximum pooling layers. The convolution block consists of a convolutional layer, a batch normalization layer, and a ReLU layer; The max pooling layer expands the receptive field and reduces the training cost; Semantic feature processing module, using skip connection mechanism; A fully connected module maps the feature representation of the distribution to the sample space and provides the output of the hidden layer; Step 8-4: In the output layer, the probability distribution of the five types of actions is obtained from the output of the hidden layer.

10. The RFID-based passive target location and behavior perception method according to claim 9, characterized in that: The pruning operation and recognition result output are specifically: Step 9-1, using a human height determination module to distinguish whether the person is currently sitting or standing; Step 9-2, if it is judged to be standing, the sample is further classified by checking whether the lower limbs are still or moving; Step 9-3: The category with the highest probability is the final result corresponding to the upper limb or lower limb, and these two parts are combined to form a judgment of the overall activity.

Images