Human body action classification method, device and equipment based on RFID (Radio Frequency Identification Device) and medium

Abstract

The invention discloses a human motion classification method, device and equipment based on RFID and a medium, and the method comprises the steps: collecting human skeleton phase data, and carrying out the preprocessing of the data, and forming RFID skeleton activity data; constructing a human body action diagram, and connecting hand joint nodes and foot joint nodes in the human body action diagram to form an initial human body skeleton structure diagram; and performing graph convolution processing on the initial human body skeleton structure graph to generate a human body action prediction classification label. By adding an improved graph structure formed by connection between the hands and the feet in the human body structure graph, the topological distance between the nodes is reduced, the nodes can be aggregated to more neighbor node information, the aggregation capability of joint features is improved, and the joint feature aggregation efficiency is improved. And during action recognition of cooperation of both hands and hands and feet, richer action feature information can be captured through aggregation features among non-physically connected joints, and the recognition precision of the model on complex postures is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of human action recognition, and in particular, to a method, device, equipment and medium for classifying human actions based on RFID. Background Art

[0002] The importance and criticality of human action recognition and classification technology in the fields of human-computer interaction, kinematics research, medical diagnosis, and action analysis have been gradually increasing, meeting the seamless connection of smart city construction and intelligent environment interactions such as games and virtual reality, and continuously promoting the continuous development of technology. At the same time, the requirements for human action recognition and classification in various fields are also constantly improving, and the requirement for privacy protection is particularly urgent during posture tracking. Placing sensing and recognition tags on the human body can achieve non-invasive action recognition without direct contact or relying on visual sensors, improving the freedom and comfort of users.

[0003] By placing RFID tags at key parts for action recognition such as human joints, human bone action information can be collected to achieve non-invasive action recognition, which can not only overcome the limitations of traditional vision technology in terms of security and privacy, but also provide more accurate and reliable recognition results. When performing human action recognition based on RFID, RFID data is usually regarded as Euclidean data with a regular structure, ignoring the inherent topological characteristics of human actions, resulting in the difficulty of fully displaying the internal characteristics of human actions, poor self-adaptability, and inaccurate recognition of human actions. Summary of the Invention

[0004] Embodiments of the present invention provide a method, device, equipment and medium for classifying human actions based on RFID to solve the technical problem of difficultly fully capturing the action correlation between human hands and feet and inaccurate recognition of human actions.

[0005] In a first aspect, embodiments of the present invention provide a method for classifying human actions based on RFID, including: S101, collecting human skeleton phase data composed of RFID tags by using a built RFID antenna array, and preprocessing the human skeleton phase data to form RFID skeleton activity data; S102, constructing a human action graph by using the RFID skeleton activity data, connecting the hand joint nodes and foot joint nodes in the human action graph to form an initial human skeleton structure diagram; S103, performing graph convolution processing on the initial human skeleton structure diagram by using a trained graph convolution network, classifying and recognizing human actions, and generating a human action prediction classification label.

[0006] Further, the S103 further includes: When performing graph convolution on the initial human skeleton structure diagram, a self-connection term is introduced into the adjacency matrix through a self-connection identity matrix, and the similarity of the adjacency matrix is calculated to generate a connection enhancement matrix, adding an adaptive term to the adjacency matrix.

[0007] Further, the S102 includes: Establish a human body motion graph based on RFID tags according to the human skeleton structure; Connect the hand joint nodes and foot joint nodes in the human body motion graph; According to the corresponding relationship between the positions of RFID tags on the human body joints and the human skeleton structure in the RFID skeleton activity data, embed the RFID skeleton activity data into the human body motion graph to form an initial human skeleton structure diagram.

[0008] Further, the S101 includes: Collect human skeleton phase data and perform channel hopping buffering processing on the human skeleton phase data; Perform phase unwrapping processing on the human skeleton phase data to form RFID skeleton activity data.

[0009] Further, when training, the trained graph convolutional network is trained using a model-agnostic meta-learning MAML training strategy based on meta-learning.

[0010] Further, the trained graph convolutional network includes: a batch normalization layer, multiple graph convolutional layers, an average pooling layer, and an activation function layer.

[0011] Further, there are six graph convolutional layers, and the number of input channels and output channels of each graph convolutional layer increases sequentially.

[0012] In a second aspect, an embodiment of the present invention provides a human body motion classification device based on RFID, including: A data acquisition module for collecting RFID tag phase data and performing preprocessing; A graph structure establishment module for embedding RFID skeleton activity data into a human body motion graph according to the position correspondence between RFID tags and the human skeleton; A graph convolution module for performing graph convolution processing on the initial human skeleton structure diagram.

[0013] In a third aspect, an embodiment of the present invention provides a device, including: One or more processors; A storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned RFID-based human body motion classification method.

[0014] Fourthly, an embodiment of the present invention provides a storage medium containing computer-executable instructions, and the computer-executable instructions are used to execute the above-mentioned RFID-based human motion classification method when executed by a computer processor.

[0015] A method, device, equipment and medium for classifying human motions based on RFID provided by an embodiment of the present invention. By adding an improved graph structure formed by connecting the hands and feet in the human body structure diagram, the topological distance between nodes is reduced, enabling nodes to aggregate more neighbor node information, enhancing the aggregation ability of joint features. When recognizing motions involving both hands and hands-foot coordination, the aggregation features between non-physically connected joints can capture richer motion feature information, improving the recognition accuracy of the model for complex postures. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 is a flowchart of a method for classifying human motions based on RFID according to Embodiment 1 of the present invention; Figure 2 is a schematic diagram of an improved SNLC strategy graph structure according to Embodiment 1 of the present invention; Figure 3 is a schematic diagram of the model structure of a graph convolutional network according to Embodiment 2 of the present invention; Figure 4 is a flowchart of a method for classifying human motions based on RFID according to Embodiment 3 of the present invention; Figure 5 is a schematic diagram of the human skeleton structure according to Embodiment 3 of the present invention; Figure 6 is a schematic diagram of the structure of a device for classifying human motions based on RFID according to Embodiment 4 of the present invention; Figure 7 is a structural diagram of the equipment according to Embodiment 5 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The present invention will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. Additionally, it should be noted that for the sake of description, only parts related to the present invention are shown in the drawings, rather than all the structures.

[0018] Embodiment 1 Figure 1The figure is a flowchart of a method for classifying human actions based on RFID according to Embodiment 1 of the present invention. In this embodiment, the phase data of RFID tags is collected, and an improved human skeleton structure diagram is constructed. A graph convolutional network is used to identify and classify human actions, which specifically includes the following steps: S101. Use the built RFID antenna array to collect the phase data of the human skeleton composed of RFID tags, and preprocess the phase data of the human skeleton to form RFID skeleton activity data.

[0019] First, an RFID (Radio Frequency Identification) scenario needs to be built for collecting human action behavior data. Use an antenna array composed of RFID readers and writers as transceiver devices, which are responsible for transmitting signals and collecting the phase data returned by RFID tags as receivers to form an RFID communication system. Attach several RFID tags to the human joint parts of the staff. When the staff makes actions in front of the antenna array, the antenna array can identify and analyze the human actions by collecting the signal changes caused by different human actions reflected by the phase data of the RFID tags. Human actions include various single-person actions such as walking forward, bowing, sitting down, standing up, squatting, kicking the left leg, kicking the right leg, jumping, clapping, raising the hand, etc. The response records of all RFID tags during the duration of a single action constitute the original feature data matrix, and phase features are extracted from it to form the phase data of the human skeleton. In order to improve the accuracy of action recognition and reduce the adverse effects such as signal interference, it is necessary to preprocess the phase data of the human skeleton to reduce the influence of interference such as phase distortion, sharp phase change, and noise, so that the data can be used for model calculation and analysis to form RFID skeleton activity data.

[0020] S102. Use the RFID skeleton activity data to construct a human action graph, and connect the hand joint nodes and foot joint nodes in the human action graph to form an initial human skeleton structure diagram.

[0021] Since the RFID tags are placed on the human joint parts, when the staff makes actions, the corresponding human joints will move, driving the positions of the RFID tags to move. The RFID skeleton activity data is composed of the phase change data generated by the displacements of all RFID tags corresponding to the positions of the human joints. Therefore, it is necessary to construct a human action graph according to the positions of the human joints where each RFID tag is located in the RFID skeleton activity data. Since there is associated movement between the two hands and between the hands and feet when the human makes many actions (such as clapping and walking), the two hand nodes and the hand node and the foot node in the human action graph are connected to form an improved graph structure of the SNLC (Skeleton Node Label Construction) strategy, asFigure 2 As shown, node 5 and node 8 in the figure respectively represent the wrist nodes, corresponding to the RFID tags placed on the wrists of the staff. Node 11 and node 14 respectively represent the ankle nodes, corresponding to the RFID tags placed on the ankles of the staff. Node 4 and node 7 respectively represent the elbow nodes, corresponding to the RFID tags placed on the elbow joints of the staff. Node 10 and node 13 respectively represent the knee nodes, corresponding to the RFID tags placed on the knee joints of the staff. Then, according to the positions of each RFID tag on the human joints, the RFID skeleton motion data is embedded into the nodes at the corresponding joint positions in the human motion diagram to form an initial human skeleton structure diagram for subsequent graph convolution processing, and the human motion types are identified through the phase change situation.

[0022] S103, use the trained graph convolutional network to perform graph convolution processing on the initial human skeleton structure diagram, classify and identify human motions, and generate human motion prediction classification labels.

[0023] When performing graph convolution processing on the initial human skeleton structure diagram, since the graph structure is improved by connecting the two hand nodes and the hand and foot nodes, when using multiple graph convolutional layers to aggregate the information of adjacent nodes, new information transmission paths are introduced, reducing the topological distance between nodes, increasing the number of neighborhoods and the receptive field of the graph convolutional network. The receptive field of graph convolution is gradually increased through multiple graph convolutional layers, features from local to global are extracted, human motions are identified and classified from the phase change information of each node, and the final human motion prediction classification labels are generated.

[0024] In this embodiment, by adding an improved graph structure formed by connecting the hands and feet in the human structure diagram, the topological distance between nodes is reduced, enabling nodes to aggregate more neighbor node information, enhancing the aggregation ability of joint features. When recognizing the coordinated motions of both hands and hands and feet, more abundant motion feature information can be captured through the aggregated features between non-physically connected joints, improving the recognition accuracy of the model for complex postures.

[0025] Optionally, the trained graph convolutional network is trained using the model-agnostic meta-learning MAML training strategy based on meta-learning during training.

[0026] Due to the wide range of application fields for human action recognition, it is difficult to obtain comprehensive skeleton type data. When recognizing human actions, one often encounters skeleton types that have not been trained. The meta-learning training strategy can improve the adaptability of untrained skeletons. Specifically, the MAML (Model-Agnostic Meta-Learning) training method is adopted. The initial parameters of the model are optimized through multi-task pre-training. First, the network is initialized to determine the initial model variables, and then it is adapted to the new data domain through training tasks, and the network loss in the new data domain is minimized. That is, the optimization problem of network initialization can be expressed as:

[0027] where, represents the loss function, represents updating the variable D using the data sampled from the new data domain X for k times of gradient descent operations. The loss function L is expressed as:

[0028] where, x is the output matrix of the graph convolutional network, q is the true number of actions in the p th RFID skeleton activity graph among multiple RFID skeleton activity graphs, represents the probability that the p th RFID skeleton activity graph belongs to the q th activity, represents ensuring non-negativity and amplifying the numerical difference through exponentiation, P is the total number of nodes, n is the index of the action type, is the number of human actions. By calculating the probability that each RFID skeleton action graph is correctly classified and mapping it to [0, 1], the mean of the correct classification probabilities is calculated as the loss of the model. The model parameters are optimized by the Adam optimizer. The loss function, through the graph features extracted by the graph convolutional network, can indirectly prompt the model to learn a more generalizable representation, improve the cross-domain generalization performance of the model, and through multi-task joint optimization, avoid overfitting to the trained skeletons and enhance the generalization ability to untrained skeletons. By treating the dataset of untrained skeletons as a task in the meta-learning training strategy, when recognizing human actions, it reduces the dependence on a large amount of labeled data, improves the generalization ability of the model, and can quickly adapt and be applicable to human skeleton action recognition tasks in more fields.

[0029] Embodiment 2 This embodiment is optimized based on the above embodiment. In this embodiment, step S103 further includes: When performing graph convolution on the initial human skeleton structure diagram, a self-connection term is introduced into the adjacency matrix through a self-connection identity matrix, and the similarity of the adjacency matrix is calculated to generate a connection enhancement matrix, adding an adaptive term to the adjacency matrix.

[0030] The improved graph structure formed by connecting the two hand nodes and the hand-foot nodes can be represented as graph G(V,E) , V represents a node, corresponding to a human joint, that is, the distribution of RFID tags on the human body. E represents a spatial edge, reflecting the natural connection of human joints within the framework, such as physical connections (connections between the hand and the elbow) and non-physical connections (coordinated movements between the two hands and between the hands and feet). Introducing a self-connection identity matrix representing self-connection during graph convolution can be expressed as:

[0031] Among them, , A is the adjacency matrix, I represents the self-connection identity matrix, is a trade-off parameter, is the degree matrix, i represents the current node, j represents i the neighbor nodes of F is the input feature, w is the weight vector of multiple stacked output channels. By introducing a self-connection term through the self-connection identity matrix, the information of each vertex itself is aggregated. Therefore, the convolution of vertex in the spatial dimension is defined as:

[0032] Among them, represents the i th joint point, represents the sampling area of the convolution of node ( i.e., the set ), which is the sampling area composed of all points adjacent to node and with a value of 1, is the weight set of all neighbor nodes of node , is a mapping function, representing the adjacency relationship mapping between node and its neighbor nodes , with a value of 1 when adjacent and 0 otherwise. is a normalization factor, aiming to balance the contributions between different neighbor nodes and is used for normalization processing.

[0033] To improve the adaptive ability of the model, by calculating the similarity of the adjacency matrix, a connection strengthening matrix is formed. Using the connection strengthening matrix to form an adaptive term to enhance the adaptive ability of the model, a GSE strategy (Graph Structure Enhancement) is formed. The output of the feature map after convolutional processing of each vertex in the initial human skeleton structure diagram can be expressed as:

[0034] Among them, represents the kernel size of the spatial convolution operation, k represents the current kernel size participating in the operation, represents the normalized adjacency matrix, F represents the input feature, represents the set of weighting functions. Introduce the strengthened connection matrix C representing the similarity of the adjacency matrix, and the output of the formed feature map is expressed as:

[0035] Among them, A represents the adjacency matrix, C represents the connection strengthening matrix, and the connection strengthening matrix is adaptively optimized by calculating the similarity of the adjacency matrix. According to the feature tensor F of the current input sample, through the embedding function and map the original features to a space suitable for calculating similarity, calculate the dynamic similarity between features, and then use softmax for normalization to obtain the connection strengthening matrix C with adaptive ability reflecting the feature correlation strength of the current sample:

[0036] Among them, and represent the embedding functions, with a 1×1 convolutional kernel, used for linear transformation in the channel dimension on each node, and respectively represent the weight parameter matrices of the embedding functions and , represents the transpose of the input feature matrix, is the activation function.

[0037] By introducing self - connection terms, the self - connection between each node is strengthened, and the correlation degree of the connection between the two - hand nodes and the hand - foot nodes is improved. By introducing adaptive terms, the adaptability of the graph convolutional network to human bones is enhanced, and the features of the sample individuals can be mined more accurately for correlation, improving the recognition accuracy of complex actions (such as multi - angle, multi - posture, etc.). GSE improves the feature extraction method. By enhancing the graph structure, when the model is optimized by backpropagation through the loss function, it can indirectly prompt the model to learn more generalizable representations, thus performing better on data in different fields. Through the connection of the two - hand nodes and the hand - foot nodes, introducing self - connection terms and adaptive terms, and the meta - learning training strategy to enhance the cross - domain generalization ability of the model. With the cooperation of the three, a balance of high accuracy, low annotation dependence, and recognition speed can be achieved.

[0038] An optional implementation manner of this embodiment is that the trained graph convolutional network includes: a batch normalization layer, multiple graph convolutional layers, an average pooling layer, and an activation function layer.

[0039] First, the input data is normalized by the batch normalization layer to reduce the scale difference between parameters and improve the model performance. Then, the feature map is processed by multiple graph convolutional layers for graph convolution, and after the graph convolution processing, global average pooling is performed by the average pooling layer to reduce the number of parameters and improve the generalization ability. Finally, the activation function layer uses softmax the activation function for normalization processing and outputs the classification label of human action prediction.

[0040] Specifically, as Figure 3 shown, there are six graph convolutional layers, and the number of input channels and output channels of each graph convolutional layer increases sequentially.

[0041] When performing convolution processing on the initial human skeleton structure diagram through graph convolution layers that gradually increase the number of input channels and output channels, as the number of input channels gradually increases, the convolution receptive field is gradually enhanced. As the number of output channels gradually increases, the feature expression ability is gradually improved, from local features to global features, enhancing the recognition ability of complex actions. Exemplarily, the number of input channels of six graph convolution layers are successively: 4, 64, 128, 128, 256, and 256, and the number of output channels are successively: 64, 64, 128, 128, 256, and 256. The stride - that is, the layer depth of aggregating neighbor nodes - are successively: 1, 2, 1, 2, 1, and 1. Local features are extracted in the first graph convolution layer, and the initial number of channels is expanded; in the second graph convolution layer, local features are further extracted, and downsampling is performed by increasing the stride, reducing the data resolution while retaining the main features, reducing the number of parameters and the amount of computation, and enhancing the understanding of local structures; in the third graph convolution layer, the number of channels is increased to obtain richer skeleton structure information, enhancing the model's expression ability and learning more complex feature relationships; in the fourth graph convolution layer, downsampling is performed again, while deeper features are extracted, compressing the data scale again, reducing redundant information, and enhancing the computational efficiency; in the fifth graph convolution layer, the feature expression ability is enhanced by increasing the number of channels, enabling better understanding of complex action patterns; in the sixth graph convolution layer, high-level semantic information is further extracted, and the same spatial dimension is maintained, further optimizing the high-dimensional feature representation and enhancing the classification ability of the model.

[0042] Embodiment III Figure 4 The following is a flowchart of a method for classifying human actions based on RFID according to Embodiment III of the present invention. This embodiment is optimized based on the above embodiment. In this embodiment, S102 is specifically optimized as: Establish a human action graph based on RFID tags according to the human skeleton structure; Connect the hand joint nodes and the foot joint nodes in the human action graph; According to the corresponding relationship between the positions of RFID tags on the human joints and the human skeleton structure in the RFID skeleton activity data, embed the RFID skeleton activity data into the human action graph to form an initial human skeleton structure diagram.

[0043] Correspondingly, the method for classifying human actions based on RFID provided in this embodiment specifically includes: S301, use the built RFID antenna array to collect human skeleton phase data composed of RFID tags, and preprocess the human skeleton phase data to form RFID skeleton activity data.

[0044] S302, establish a human action graph based on RFID tags according to the human skeleton structure.

[0045] After setting up the acquisition scenario and collecting RFID skeleton activity data, according to the positions of the RFID tags on the human joints, a human motion graph is established, and the nodes in the human motion graph correspond one-to-one with the positions of the RFID tags on the human joints, as Figure 5 (a) shown.

[0046] S303. Connect the hand joint nodes and the foot joint nodes in the human motion graph.

[0047] Since there are coordination relationships between the hands and between the hands and feet when the human body makes some movements, the nodes between the two hands and the nodes between the hands and feet are connected, as Figure 5 (b) shown, so as to capture the coordinated movements between the hands and between the hands and feet during the graph convolution process, which can improve the recognition accuracy of complex movements.

[0048] S304. According to the correspondence between the positions of the RFID tags on the human joints and the human skeleton structure in the RFID skeleton activity data, embed the RFID skeleton activity data into the human motion graph to form an initial human skeleton structure graph.

[0049] According to the positions of the RFID tags at the human joints, embed the phase change data of each RFID tag in the RFID skeleton activity data into the nodes in the human motion graph corresponding to the positions of the RFID tags at the joints to form an initial human skeleton structure graph. In the initial human skeleton structure graph, the data of each node respectively reflects the phase change data of the RFID tag placed at the corresponding human joint. When using the initial human skeleton structure graph for graph convolution processing, the displacement situation between human joints can be analyzed through the phase change data of the RFID tags, and then the actions made by the human body can be recognized and classified.

[0050] S305. Use the trained graph convolution network to perform graph convolution processing on the initial human skeleton structure graph, classify and recognize human actions, and generate human action prediction classification labels.

[0051] Optionally, the S101 includes: Collect human skeleton phase data and perform channel hopping buffer processing on the human skeleton phase data.

[0052] In the original feature data matrix composed of the response records of all RFID tags attached to human joints during the action duration, extract the phase feature and represent it as:

[0053] Among them, M represents the total number of RFID signal propagation paths, m represents the index of the signal propagation path, and respectively represent the signal strength and distance of each multipath component, c represents the speed of light, is the current channel frequency, is the initial phase offset caused by the circuits of both the antenna and the tag in the channel s Since both the reader antenna and the RFID tag circuit have electronic components and circuit layouts, etc., these circuit characteristics will cause a fixed initial phase offset when the signal is transmitted or received. It is not affected by factors such as signal propagation distance during the propagation process and is only determined by the circuit structure, component parameters, etc. It will be superimposed on the phase of the received signal. represents the complex phase factor of the total phase on this propagation path. For each path, using the path length and the channel frequency product, combined with the speed of light and 2π, calculate the phase introduced by path propagation, and then add it to the initial phase to obtain the total phase of this path , and then use the total phase to calculate the complex phase factor , and finally multiply this factor by the signal strength of this path to obtain the signal component of this path. Finally, add the signal components of all paths to obtain the total signal .

[0054] The phase data can reflect the positional relationship between the tag and the reader. When the tag moves due to human actions, the relative position between the tag and the reader will also change, so there will be a phase change of the RFID tag. In each propagation environment, there is at least one dominant path, that is, the main line-of-sight (LOS) path distance between the RFID reader antenna and the RFID tag. The received phase data is expressed as:

[0055] where, R is the main line-of-sight (LOS) path distance between the RFID reader antenna and the RFID tag, s is the channel index, n is the number of channels. In the ultra-high frequency band (UHF band, 902–928 MHz), it is divided into 50 channels according to the regulations of the FCC (Federal Communications Commission of the United States), that is, n = 50, and the channel index sOn each channel, it changes from 1 to 50 every 200 ms. That is, when the reader receives the phase information of the RFID tag, it will sequentially traverse 50 channels within 200 ms to obtain the phase values of each channel, and use the difference between two adjacent phase samples on the same channel to cancel the same-channel phase offset component in the two samples. , cancel the phase offset component of the same channel, and obtain the s phase change of this channel, which is expressed as:

[0056] where n is the sample index on each channel, is the propagation distance corresponding to the s th n sample on channel . Since the phase change only depends on the change in the range of the main propagation path: , the sequence of phase changes is: , and the sequence of distance changes between the reader antenna and the RFID tag can be obtained as:

[0057] , which is used to represent the change in the propagation distance of the captured RFID tag, that is, the displacement of the RFID tag. The channel offset is related to the channel. When channel hopping is performed according to FCC rules, the channel offset also changes. When directly using the original phase data, since each phase contains a channel offset phase, the phase difference of each channel needs to be used to cancel the channel offset phase and perform channel hopping buffer processing, so that the extracted phase change only depends on the distance change caused by the tag movement, reducing the phase distortion caused by channel hopping.

[0058] Unwrap the phase change of the RFID tag into:

[0059] where represents the phase change amount after phase unwrapping processing, represents the difference between the original phases of two consecutive measurements, represents an integer multiple correction term of a , and its sign is the same as the sign of . When , it means that a periodic jump occurs in the phase change at this time. At this time, a is added or subtracted for correction. When , no correction is required at this time, and the value after phase unwrapping is equal to the original value. Since the original phase is a periodic function, the collected original phase data will be mapped by modulo operation in [0, within the range, when crossing 0 or during this time, the modulo operation will produce a sharp phase change. By expanding the phase change data through the above formula, identify and correct the integer multiple ambiguity, reconstruct a continuous phase sequence from the original, possibly discontinuous phase data, perform phase unwrapping processing on the phase data, and mitigate the sharp phase change caused by the modulo operation. The phase change tensor representation of the RFID tag can be obtained as:

[0060] where is a three-dimensional tensor, t represents the time slot, represents the total number of time slots read by each RFID reader antenna during a period of time, and are the indices of the RFID reader antenna and the RFID tag respectively, represents in the time slot t the calibration phase change of the RFID tag when sampled by the antenna during this time.

[0061] In this embodiment, the collected phase data is preprocessed through channel hopping buffer processing and phase unwrapping, mitigating the phase distortion caused by channel hopping and the sharp phase change caused by the modulo operation, and forming an RFID phase change tensor that can be used in the deep learning model. The preprocessed data can construct a human motion graph according to the position of the RFID tag on the human joint, and embed the phase change data of the RFID tag into the nodes at the corresponding joint positions in the graph, generating an initial human skeleton structure graph for the graph convolutional network, reducing the interference information in the phase data, and improving the classification accuracy of the model.

[0062] Embodiment 4 Figure 6 is a schematic structural diagram of a human motion classification device based on RFID according to Embodiment 4 of the present invention. In this embodiment, the human motion classification device based on RFID includes: A data acquisition module 810, configured to acquire RFID tag phase data and perform preprocessing; A graph structure establishment module 820, configured to embed RFID skeleton activity data into a human motion graph according to the position correspondence between the RFID tag and the human skeleton; A graph convolution module 830, configured to perform graph convolution processing on the initial human skeleton structure graph.

[0063] In this embodiment, the data acquisition module collects and preprocesses the phase data of the RFID tags placed at the human joints. The graph structure building module embeds the phase data of the RFID tags into the nodes at the corresponding positions in the corresponding human action graph. The graph convolution module performs convolution processing on the human skeleton structure graph to identify and classify the actions of the human skeleton and output classification labels. By adding an improved graph structure formed by the connection between the hands and feet in the human structure graph, the topological distance between nodes is reduced, enabling nodes to aggregate more neighbor node information, enhancing the aggregation ability of joint features. When recognizing actions involving both hands and the coordination of hands and feet, the aggregation features between non-physically connected joints can capture richer action feature information, improving the recognition accuracy of the model for complex postures.

[0064] The RFID-based human action classification device provided by the embodiment of the present invention can execute the RFID-based human action classification method provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the method.

[0065] Embodiment 5 Figure 7 It is a structural diagram of a device according to Embodiment 5 of the present invention. Figure 7 It shows a block diagram of an exemplary device 12 suitable for implementing the embodiments of the present invention. Figure 7 The shown device 12 is only an example and should not impose any limitation on the functions and usage scope of the embodiments of the present invention.

[0066] As Figure 7 shown, the device 12 is presented in the form of a general-purpose computing device. The components of the device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 connecting different system components (including the system memory 28 and the processing unit 16).

[0067] The bus 18 represents one or more of several types of bus structures, including a memory bus or a memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the multiple bus structures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0068] The device 12 typically includes a variety of computer system-readable media. These media can be any available media accessible by the device 12, including volatile and non-volatile media, removable and non-removable media.

[0069] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used for reading and writing on non-removable, non-volatile magnetic media ( Figure 7 not shown, commonly referred to as a "hard disk drive"). Although Figure 7 not shown in the figure, a disk drive for reading and writing on removable non-volatile disks (such as "floppy disks"), and an optical disk drive for reading and writing on removable non-volatile optical disks (such as CD-ROM, DVD-ROM or other optical media) may be provided. In these cases, each drive may be connected to bus 18 through one or more data media interfaces. Memory 28 may include at least one program product having a set (such as at least one) of program modules configured to perform the functions of various embodiments of the present invention.

[0070] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. An implementation of a network environment may be included in each or some combination of these examples. Program modules 42 generally perform the functions and / or methods in the embodiments described in the present invention.

[0071] Device 12 may also communicate with one or more external devices 14 (such as a keyboard, a pointing device, a display 24, etc.), and may also communicate with one or more devices that enable a user to interact with this device 12 / server / computer, and / or communicate with any device that enables this device 12 to communicate with one or more other computing devices (such as a network card, a modem, etc.). Such communication may be carried out through an input / output (I / O) interface 22. Also, device 12 may communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through a network adapter 20. As shown in the figure, network adapter 20 communicates with other modules of device 12 through bus 18. It should be understood that although not shown in the figure, other hardware and / or software modules may be used in conjunction with device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0072] Processing unit 16 performs various functional applications and data processing by running programs stored in system memory 28, such as implementing the RFID-based human motion classification method provided by the embodiments of the present invention.

[0073] Example 6 Example 6 of the present invention further provides a storage medium containing computer-executable instructions, and the computer-executable instructions are used to execute the RFID-based human motion classification method provided in the above embodiments when executed by a computer processor.

[0074] The computer storage medium of the embodiments of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0075] The computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, in which the computer-readable program code is carried. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, and the computer-readable medium can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0076] The program code contained on the computer-readable medium can be transmitted by any appropriate medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0077] Computer program code for performing the operations of the present invention may be written in one or more programming languages or combinations thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and also including conventional procedural programming languages such as the "C" language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0078] Note that the above is only the preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, re-adjustments, and substitutions can be made by those skilled in the art without departing from the scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments may be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for classifying human body movements based on RFID, characterized in that, Including: S101, collecting human skeleton phase data composed of RFID tags by using the built RFID antenna array, and preprocessing the human skeleton phase data to form RFID skeleton activity data; S102, constructing a human action graph by using the RFID skeleton activity data, connecting the hand joint nodes and the foot joint nodes in the human action graph to form an initial human skeleton structure graph; S103, performing graph convolution processing on the initial human skeleton structure graph by using the trained graph convolutional network, classifying and recognizing human actions, and generating human action prediction classification labels.

2. The method according to claim 1, wherein The S103 further includes: When performing graph convolution on the initial human skeleton structure graph, introducing a self-connection term into the adjacency matrix through a self-connection identity matrix, calculating the similarity of the adjacency matrix, generating a connection strengthening matrix, and adding an adaptive term to the adjacency matrix.

3. The method according to claim 1, wherein The S102 includes: Establishing a human action graph based on RFID tags according to the human skeleton structure; Connecting the hand joint nodes and the foot joint nodes in the human action graph; Embedding the RFID skeleton activity data into the human action graph according to the corresponding relationship between the human joint positions where the RFID tags are located and the human skeleton structure in the RFID skeleton activity data to form an initial human skeleton structure graph.

4. The method according to claim 1, characterized in that The S101 includes: Collecting human skeleton phase data and performing channel hopping buffer processing on the human skeleton phase data; Performing phase unwrapping processing on the human skeleton phase data to form RFID skeleton activity data.

5. The method according to claim 1, characterized in that: The trained graph convolutional network is trained by adopting a model-agnostic meta-learning MAML training strategy based on meta-learning during training.

6. The method according to claim 1, characterized in that: The trained graph convolutional network includes: a batch normalization layer, multiple graph convolutional layers, an average pooling layer, and an activation function layer.

7. The method according to claim 6, wherein: There are six graph convolutional layers, and the number of input channels and output channels of each graph convolutional layer increases sequentially.

8. An RFID-based human motion classification device, characterized in that Including: A data acquisition module for collecting RFID tag phase data and performing preprocessing; A graph structure establishment module for embedding RFID skeleton activity data into a human action graph according to the position corresponding relationship between RFID tags and the human skeleton; A graph convolution module for performing graph convolution processing on the initial human skeleton structure graph.

9. A device, characterized in that, The device includes: One or more processors; A storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, enabling the one or more processors to implement the RFID-based human action classification method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, where the computer-executable instructions are used to execute the RFID-based human action classification method as described in any one of claims 1-7 when executed by a computer processor.

Images