A method and system for identifying human behavior based on WiFi signal based on data packet compression network

Abstract

The present invention discloses a method and system for identifying human behavior based on a WiFi signal based on a data packet compression network, which is applied to the field of pattern recognition technology. The WiFi antenna collects data packets, each of which contains the channel state information of the subcarriers, and all the information is merged into a data matrix. After passing through a 3×3 convolutional layer, a normalization layer, and an activation layer, it is sent to a compression residual module; after passing through four stages of compression residual modules and downsampling compression residual module network combinations, a mean pooling layer and a full connection layer are performed, and finally the human behavior type is predicted. The compression residual module network adds a data packet compression network to the commonly used residual block, extracts the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weights the input data packet by multiplication to complete the recalibration of the original data. The present invention can effectively extract the characteristic information of the data packet and significantly improve the accuracy of human behavior recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of pattern recognition, and more specifically to a method and system for identifying human behavior using a WiFi signal based on a data packet compression network. Background Art

[0002] Human behavior recognition is a technology that identifies and classifies different behaviors by processing and analyzing human behavior information obtained by various sensors. With the rapid development of recognition technology, the recognition accuracy and speed have been significantly improved. Human behavior recognition is widely used in application fields such as smart security, smart home and health monitoring. Human behavior recognition technology can be divided into the following three types according to the type of sensor: camera-based, wearable sensor-based and wireless signal-based human behavior recognition system. The camera-based human behavior recognition system uses a high-definition camera to shoot human behavior scenes, and recognizes various human behaviors through image processing and feature extraction technology. Although this method can achieve a high recognition rate, the video or image captured is easily affected by the camera's field of view and the external brightness environment, and the captured image is prone to privacy leakage. The wearable sensor-based human behavior recognition system requires users to wear hardware equipment to collect motion data, and realizes human behavior recognition by detecting and processing motion data. Although this type of system also has a high recognition rate, it affects the user experience because it requires wearing various sensors and requires a lot of calibration work. The human behavior recognition system based on wireless signals mainly transmits electromagnetic waves to the receiving end through the signal source device. On the transmission path of the electromagnetic waves, different human behaviors will cause changes in the amplitude or phase of the electromagnetic waves at the receiving end, and the purpose of recognition is achieved by capturing and sensing the changes. Based on wireless signals, human behavior effectively protects personal privacy without collecting sensitive information of users, and improves user experience without wearing other hardware devices. Therefore, it has attracted widespread attention from academia and industry.

[0003] As the demand for connecting the Internet of Things continues to increase, the popularity of WiFi devices is increasing, and the human activity recognition technology based on WiFi signals has also developed rapidly. There are two main WiFi signal attributes that can be used to recognize human activities: received signal strength (RSS) and channel state information (CSI). Channel state information describes the signal transmission characteristics on each communication link, including the amplitude value, phase value and frequency response value of the signal, so it has higher robustness and accuracy than RSS. In order to accurately identify the changes in channel state information caused by different human activities, researchers have developed many recognition algorithms based on convolutional neural networks. The commonly used convolutional algorithm has a limited receptive field and can only extract effective information in a single space, ignoring the continuity of human activities, and the importance of different data packets cannot be reflected.

[0004] Therefore, in view of the defects of the above-mentioned human activity recognition algorithm based on WiFi signals and the singleness of effective information extraction in traditional convolutional neural networks, the present invention invents a method for human behavior recognition based on WiFi signals based on a data packet compression network. By compressing and extracting the directional dimension of the data packet, the weight of each data packet is extracted, and then the input data packet is weighted by multiplication to complete the recalibration of the original data. In the training stage, two different loss functions are used for training. The features after the pooling layer are trained using a ternary loss function, and the features after the fully connected layer are trained using a cross entropy loss function, which can not only accelerate the convergence speed of training, but also improve the generalization ability of the network. Summary of the invention

[0005] In view of this, the present invention provides a method and system for identifying human behavior based on a WiFi signal in a data packet compression network to solve the problems in the background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] On the one hand, the present invention discloses a method for identifying human behavior based on a WiFi signal in a data packet compression network, and the specific steps are as follows:

[0008] Build a behavior recognition system, with a fixed distance between the receiving end and the sending end, and establish a WiFi communication connection between the receiving end and the sending end; identify that the object is between the sending end and the receiving end, and perform corresponding actions;

[0009] The receiving end receives the data packet sent by the transmitting end, parses it, and obtains the channel state information of each subcarrier as a sample; after filtering and segmentation, all the channel information is merged into a data matrix;

[0010] Establish a human behavior recognition model, extract the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weight the input data packet by multiplication to complete the recalibration of the original data; obtain the optimal human behavior recognition model;

[0011] Obtain the data packet corresponding to the behavior to be tested, input the optimal human behavior recognition model, and output the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0012] Preferably, in the above-mentioned method for human behavior recognition based on WiFi signal of data packet compression network, a human behavior recognition model is established, and the specific steps are as follows:

[0013] A data packet compression network is constructed using Resnet18 as the basic network, wherein the data packet compression network includes a basic network and a compression network;

[0014] The compression network captures the data packets that play a decisive role in the final recognition by adjusting the size of the feature tensor and the number of output channels, reassigns each data packet according to its importance, and obtains the characteristic function of the data packet direction.

[0015] Preferably, in the above-mentioned method for identifying human behavior based on a WiFi signal of a data packet compression network, the specific steps are also as follows:

[0016] After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0017] Preferably, in the above-mentioned method for identifying human behavior using a WiFi signal based on a data packet compression network, the specific calculation process of the characteristic function of the data packet direction is as follows:

[0018] For a feature table with C channels, H height and W width, After lateral feature averaging, a 1×1×W tensor is obtained, which is calculated as follows:

[0019]

[0020] Where h = 1, 2, ..., H, and then pass through two fully connected layers to obtain features with strong horizontal importance. The calculation formula is as follows:

[0021]

[0022] Among them, T1 is the compression function, T2 is the expansion function, and ReLU is the activation function;

[0023] Then perform point multiplication on the input features to obtain the characteristic function of the data packet direction, namely:

[0024]

[0025] Where σ is the Sigmoid activation function, is the dot multiplication operation.

[0026] Preferably, in the above-mentioned method for human behavior recognition of WiFi signals based on data packet compression network, the features after the mean pooling layer are trained using a ternary loss function, and the formula is:

[0027] L tri =max(d p -d n +α,0)};

[0028] Among them, d p and d n They represent the feature distances of positive sample pairs and negative samples respectively, and α represents the predetermined margin.

[0029] Preferably, in the above-mentioned method for human behavior recognition of WiFi signals based on data packet compression network, a cross entropy loss function is used for training after the fully connected layer, and its formula is:

[0030]

[0031] Where n is the number of different human behavior categories, y i is the true label, p i is the predicted label.

[0032] On the other hand, the present invention discloses a WiFi signal human behavior recognition system based on a data packet compression network, comprising: a transmitting end, a receiving end and a data processing module;

[0033] The receiving end and the sending end are at a fixed distance, and a wifi communication connection is established between the receiving end and the sending end; the object is identified to be between the sending end and the receiving end, and a corresponding action is performed;

[0034] The receiving end receives the data packet sent by the transmitting end, and the data processing module parses the data packet to obtain the channel state information of each subcarrier as a sample; after filtering and segmentation, all channel information is merged into a data matrix; the optimal human behavior recognition model is input, and the output is the probability of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down and standing up, and the behavior with the largest probability value is taken as the final prediction result.

[0035] Preferably, in the above-mentioned WiFi signal human behavior recognition system based on data packet compression network, the human behavior recognition model includes: a basic network and a compression network;

[0036] Use Resnet18 as the basic network;

[0037] The compression network includes a first compression residual module, a second compression residual module, a downsampling compression residual module and a third compression residual module which are connected in sequence; by adjusting the size of the feature tensor and the number of output channels, the data packets that play a decisive role in the final recognition are captured, and each data packet is reassigned according to its importance to obtain a characteristic function of the data packet direction.

[0038] Preferably, in the above-mentioned WiFi signal human behavior recognition system based on data packet compression network, it also includes a mean pooling layer, a normalization layer and a fully connected layer;

[0039] After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0040] Preferably, in the above-mentioned WiFi signal human behavior recognition system based on data packet compression network, the first compression residual module, the second compression residual module, the down-sampling compression residual module and the third compression residual module all include a data packet compression module;

[0041] The data packet compression module takes as input a feature table with C channels, H height and W width. After lateral feature averaging, a 1×1×W tensor is obtained; then it passes through two fully connected layers, and then the input features are multiplied to obtain the feature function in the direction of the data packet.

[0042] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for identifying human behavior based on a WiFi signal based on a data packet compression network, which extracts the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weights the input data packet by multiplication to complete the recalibration of the original data. In the training stage, two different loss functions are used for training, the features after the pooling layer are trained using a ternary loss function, and the features after the fully connected layer are trained using a cross entropy loss function, which can not only accelerate the convergence speed of training, but also improve the generalization ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0044] Figure 1 is a flow chart of the method of the present invention;

[0045] Figure 2 It is the overall network framework diagram of the present invention;

[0046] Figure 3 This is a framework diagram of the compression residual module of the present invention;

[0047] Figure 4 This is a framework diagram of the downsampling compression residual module of the present invention;

[0048] Figure 5 This is a framework diagram of the data packet compression module of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The embodiment of the present invention discloses a method for identifying human behavior using a WiFi signal based on a data packet compression network, and the specific steps are as follows:

[0051] Build a behavior recognition system, with a fixed distance between the receiving end and the sending end, and establish a WiFi communication connection between the receiving end and the sending end; identify that the object is between the sending end and the receiving end, and perform corresponding actions;

[0052] The receiving end receives the data packet sent by the transmitting end, parses it, and obtains the channel state information of each subcarrier as a sample; after filtering and segmentation, all the channel information is merged into a data matrix;

[0053] Establish a human behavior recognition model, extract the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weight the input data packet by multiplication to complete the recalibration of the original data; obtain the optimal human behavior recognition model;

[0054] Obtain the data packet corresponding to the behavior to be tested, input the optimal human behavior recognition model, and output the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0055] Specifically, a TP-LINK N750 router is used as the transmitting end, and a computer equipped with an Intell 5300 network card is used as the receiving end. The distance between the receiving and transmitting antennas is about 2.5 meters. The three antennas of the receiving network card are externally connected and arranged side by side, and the number of subcarriers of each antenna is 30.

[0056] The sending speed of the data packets on the sending end is set to 100 per second. The tester demonstrates the corresponding behavioral actions in the middle. Each action lasts for 2 to 3 seconds, corresponding to 200 to 300 data packets. The maximum value of 300 is taken as the number of collected data packets.

[0057] 3×30×300 data packets are collected each time.

[0058] In order to further optimize the above technical solution, the CSI-TOOL tool is used to parse the data packets and obtain the CSI amplitude value of each subcarrier as a sample. The sample is filtered by a Butterworth low-pass filter, and then the 25 data packets before and after are divided again, and the three antenna data packets of the CSI sample are merged to obtain a 1×90×250 data matrix.

[0059] In order to further optimize the above technical solution, a human behavior recognition model is established. The specific steps are as follows:

[0060] A data packet compression network is constructed using Resnet18 as the basic network, wherein the data packet compression network includes a basic network and a compression network;

[0061] The compression network captures the data packets that play a decisive role in the final recognition by adjusting the size of the feature tensor and the number of output channels, reassigns each data packet according to its importance, and obtains the characteristic function of the data packet direction.

[0062] Specifically, the present invention uses Resnet18 as the basic network, and its input feature tensor is 3×32×32. Therefore, the CSI sample matrix passes through a convolution layer with an output channel of 3, a kernel size of 7×7, and a step size of 1 and 3 to obtain a 3×81×83 tensor, and then passes through a ReLU activation layer and a 2×2 maximum pooling layer, and finally a convolution layer with a kernel size of 10×11 and a step size of 1 to obtain a 3×32×32 feature tensor suitable for Resnet18 network input.

[0063] Furthermore, if the feature tensor size and the number of output channels remain unchanged after passing through the compressed residual module, if the feature tensor size is halved after passing through the downsampling compressed residual module, the number of output channels becomes twice as much, thus ensuring that the computational effort of each module is equal.

[0064] In the first stage, the feature tensor size of the convolutional layer in the compressed residual module is 32×32, and the number of output channels is 16;

[0065] In the second stage, after the downsampling compression residual module, the feature tensor is 16×16, and the number of output channels is 32;

[0066] Similarly, after three stages of downsampling compressed residual modules and compressed residual module network combinations, a 128×4×4 feature tensor is obtained, as shown in Figure 2 shown.

[0067] The compressed residual module consists of two convolutional layers of size 3×3 and stride 1, two normalization layers, and two ReLU activation layers. The packet compression network and shortcut connection are added to the second activation layer, such as Figure 3 shown.

[0068] The downsampling compressed residual module uses a downsampling convolution layer with a kernel size of 1×1 and a step size of 2 at the input and quick connection. The downsampling method can effectively reduce the network parameters and reduce the amount of calculation, preventing overfitting. Figure 4 shown.

[0069] In order to further optimize the above technical solution, the specific steps are as follows:

[0070] After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0071] In order to further optimize the above technical solution, the specific calculation process of the characteristic function of the data packet direction is as follows:

[0072] The packet compression network captures the data packets that play a decisive role in the final recognition through the packet compression method, reassigns each data packet according to its importance, and takes the feature table with C channels, H height and W width as input. After lateral feature averaging, we get a 1×1×W tensor, which is calculated as follows:

[0073]

[0074] Where h = 1, 2, ..., H, and then pass through two fully connected layers with the same compression rate and expansion rate of r = 8 to obtain features with strong horizontal importance. The calculation formula is as follows:

[0075]

[0076] Among them, T1 is the compression function, T2 is the expansion function, and ReLU is the activation function. Then the input features are multiplied to obtain the feature function of the data packet direction (horizontal), that is:

[0077]

[0078] Where σ is the Sigmoid activation function, is the dot multiplication operation, the process is as follows Figure 5 shown.

[0079] After the mean pooling layer, a 128×4×4 feature tensor is obtained. Unlike the commonly used Resnet18 network, the present invention adds a normalization layer before the pooling layer and the fully connected layer, so that the features are more uniform and the training converges faster. The output of the fully connected layer is the probability of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0080] Using 3977 samples in UT-HAR as the training set and 996 samples as the test set, the network is trained using two loss functions. The features after the pooling layer are trained using the ternary loss function, and the formula is:

[0081] L tri =max(d p -d n +α,0)}

[0082] Among them, d p and d n Represent the characteristic distance of positive sample pairs and negative samples respectively, α represents the predetermined margin, which is set to 0.3 during training. The cross entropy loss function is used for training after the fully connected layer, and its formula is:

[0083]

[0084] Where n is the number of different human behavior categories, which is set to 7 here, and y i is the true label, p i To predict the label. The size of each batch of training data is 64, and the parameter learning rate is 10 -3 , a total of 200 rounds of training.

[0085] Another embodiment of the present invention discloses a WiFi signal human behavior recognition system based on a data packet compression network, comprising: a transmitting end, a receiving end and a data processing module;

[0086] The receiving end and the sending end are at a fixed distance, and a wifi communication connection is established between the receiving end and the sending end; the object is identified to be between the sending end and the receiving end, and a corresponding action is performed;

[0087] The receiving end receives the data packet sent by the transmitting end, and the data processing module parses the data packet to obtain the channel state information of each subcarrier as a sample; after filtering and segmentation, all channel information is merged into a data matrix; the optimal human behavior recognition model is input, and the output is the probability of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down and standing up, and the behavior with the largest probability value is taken as the final prediction result.

[0088] In order to further optimize the above technical solution, the human behavior recognition model includes: a basic network and a compression network;

[0089] Use Resnet18 as the basic network;

[0090] The compression network includes a first compression residual module, a second compression residual module, a downsampling compression residual module and a third compression residual module which are connected in sequence; by adjusting the size of the feature tensor and the number of output channels, the data packets that play a decisive role in the final recognition are captured, and each data packet is reassigned according to its importance to obtain a characteristic function of the data packet direction.

[0091] In order to further optimize the above technical solution, mean pooling layer, normalization layer and fully connected layer are also included;

[0092] After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

[0093] In order to further optimize the above technical solution, the first compressed residual module, the second compressed residual module, the down-sampling compressed residual module and the third compressed residual module all include a data packet compression module;

[0094] The data packet compression module takes as input a feature table with C channels, H height and W width. After lateral feature averaging, a 1×1×W tensor is obtained; then it passes through two fully connected layers, and then the input features are multiplied to obtain the feature function in the direction of the data packet.

[0095] Furthermore, the first compressed residual module, the second compressed residual module and the third compressed residual module are composed of convolutional layers, normalization layers and activation layers. A shortcut connection is added to the second activation layer to achieve identity mapping, so that network training converges faster and eliminates the degradation problem caused by the deepening of the network depth. On this basis, the compressed residual module adds a data packet compression module to further improve the recognition accuracy of the network.

[0096] Furthermore, downsampling in the downsampling compressed residual module enables the network to capture feature information at different resolutions. In order to solve the dimension mismatch problem of identity mapping, a downsampling convolution layer is added before the shortcut link. The downsampling convolution layer is implemented by using a 1×1 convolution with a step size of 2.

[0097] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0098] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying human behavior based on WiFi signals in a data packet compression network, characterized in that: The specific steps are as follows: Build a behavior recognition system, with a fixed distance between the receiving end and the sending end, and establish a WiFi communication connection between the receiving end and the sending end; identify that the object is between the sending end and the receiving end, and perform corresponding actions; The receiving end receives the data packet sent by the sending end, parses it, and obtains the channel state information of each subcarrier as a sample; After filtering and segmentation, all channel information is merged into a data matrix; Establish a human behavior recognition model, extract the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weight the input data packet by multiplication to complete the recalibration of the original data; obtain the optimal human behavior recognition model; establish a human behavior recognition model, the specific steps are as follows: A data packet compression network is constructed using Resnet18 as the basic network, wherein the data packet compression network includes a basic network and a compression network; The compression network captures the data packets that play a decisive role in the final recognition by adjusting the feature tensor size and the number of output channels, and reassigns each data packet according to its importance to obtain the feature function of the data packet direction; the specific calculation process of the feature function of the data packet direction is as follows: For a feature table with C channels, H height and W width, After lateral feature averaging, we get a 1×1×W tensor, which is calculated as follows: Where h = 1, 2, ..., H, and then pass through two fully connected layers to obtain features with strong horizontal importance. The calculation formula is as follows: Among them, T1 is the compression function, T2 is the expansion function, and ReLU is the activation function; Then perform point multiplication on the input features to obtain the characteristic function of the data packet direction, namely: Where σ is the Sigmoid activation function, is the dot multiplication operation; After obtaining the data packet corresponding to the behavior to be tested and obtaining the characteristic function of the data packet direction, the following steps are also included: After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

2. According to claim 1, a method for identifying human behavior based on a WiFi signal in a data packet compression network is characterized in that: The features after the mean pooling layer are trained using the ternary loss function, and the formula is: L tri =max(d p -d n +α,0)}; Among them, d p and d n They represent the feature distances of positive sample pairs and negative samples respectively, and α represents the predetermined margin.

3. According to the method of claim 1, the method is characterized in that: The cross entropy loss function is used for training after the fully connected layer, and its formula is: Where n is the number of different human behavior categories, y i is the true label, p i is the predicted label.

4. A WiFi signal human behavior recognition system based on data packet compression network, characterized in that: include: Transmitter, receiver and data processing module; The receiving end and the sending end are at a fixed distance, and a wifi communication connection is established between the receiving end and the sending end; the object is identified to be between the sending end and the receiving end, and a corresponding action is performed; The receiving end receives the data packet sent by the transmitting end, and the data processing module parses the data packet to obtain the channel state information of each subcarrier as a sample; after filtering and segmentation, all channel information is merged into a data matrix; the optimal human behavior recognition model is input, and the output is the probability of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down and standing up, and the behavior with the largest probability value is taken as the final prediction result; Establish a human behavior recognition model, extract the weight of each data packet by compressing and extracting the directional dimension of the data packet, and then weight the input data packet by multiplication to complete the recalibration of the original data; obtain the optimal human behavior recognition model; establish a human behavior recognition model, the specific steps are as follows: A data packet compression network is constructed using Resnet18 as the basic network, wherein the data packet compression network includes a basic network and a compression network; The compression network captures the data packets that play a decisive role in the final recognition by adjusting the feature tensor size and the number of output channels, and reassigns each data packet according to its importance to obtain the feature function of the data packet direction; the specific calculation process of the feature function of the data packet direction is as follows: For a feature table with C channels, H height and W width, After lateral feature averaging, we get a 1×1×W tensor, which is calculated as follows: Where h = 1, 2, ..., H, and then pass through two fully connected layers to obtain features with strong horizontal importance. The calculation formula is as follows: Among them, T1 is the compression function, T2 is the expansion function, and ReLU is the activation function; Then perform point multiplication on the input features to obtain the characteristic function of the data packet direction, namely: Where σ is the Sigmoid activation function, is the dot multiplication operation; After obtaining the data packet corresponding to the behavior to be tested and obtaining the characteristic function of the data packet direction, the following steps are also included: After the feature function of the data packet direction passes through the mean pooling layer, a normalization layer is added before the mean pooling layer and the fully connected layer. The features after the mean pooling layer are trained using the ternary loss function; the cross entropy loss function is used for training after the fully connected layer; the output of the fully connected layer is the probabilities of seven human behaviors: lying down, falling, walking, picking up things, running, sitting down, and standing up. The behavior with the largest probability value is taken as the final prediction result.

5. The WiFi signal human behavior recognition system based on data packet compression network according to claim 4 is characterized in that: The human behavior recognition model includes: basic network and compression network; Use Resnet18 as the basic network; The compression network includes a first compression residual module, a second compression residual module, a downsampling compression residual module and a third compression residual module which are connected in sequence; by adjusting the size of the feature tensor and the number of output channels, the data packets that play a decisive role in the final recognition are captured, and each data packet is reassigned according to its importance to obtain a characteristic function of the data packet direction.

6. The WiFi signal human behavior recognition system based on data packet compression network according to claim 5 is characterized in that: The first compressed residual module, the second compressed residual module, the down-sampled compressed residual module and the third compressed residual module all include a data packet compression module; The data packet compression module takes as input a feature table with C channels, H height and W width. After lateral feature averaging, a 1×1×W tensor is obtained; then it passes through two fully connected layers, and then the input features are multiplied to obtain the feature function in the direction of the data packet.

Images