Target Action Detection Method, Device and Terminal Device Based on WiFi Signal

Abstract

The present application relates to a target action detection method, apparatus, terminal device, and storage medium based on WiFi signals. There is a motion event of a preset target within a preset reception distance range of the WiFi signals. The target action detection method includes: obtaining channel state information in each path of WiFi received training signals as respective corresponding channel state signals, and obtaining a channel state real part matrix and a channel state imaginary part matrix; inputting the channel state real part matrix and the channel state imaginary part matrix into corresponding preset convolutional neural networks for processing respectively to obtain a first feature map and a second feature map; splicing the first feature map and the second feature map to obtain a spliced feature map; inputting the spliced feature map into a deformable convolutional network in a predetermined direction for feature extraction to obtain a final feature classification result; and finally obtaining a target action detection model based on WiFi signals to achieve the detection of the target motion posture in the absence of a camera.

Description

Technical Field

[0001] The present application relates to the field of target detection, and in particular, to a method, device, terminal device, and storage medium for detecting target actions based on WiFi signals. Background Art

[0002] Current machine learning algorithms for human action detection (such as fall detection, pose detection, etc.) are usually based on judging through images or multiple frames of images. For example, a detection model is combined with the TSN network to use the target detection results of the image sequence and the temporal combination of the corresponding targets to identify the pose of the target, or a key point model is combined with a graph convolutional neural network to judge the actions of the target.

[0003] However, the above methods usually require a camera with human tracking function. Summary of the Invention

[0004] The main purpose of the present application is to provide a method, device, terminal device, and storage medium for detecting target actions based on WiFi signals, aiming to solve the problem of detecting the motion posture of a target object in the absence of a camera.

[0005] A method for detecting target actions based on WiFi signals, which is applied to a signal receiving device. The signal receiving device includes at least three signal receiving units. There is a motion event of a preset target within the preset receiving distance range of the WiFi signal. The target action detection method includes:

[0006] Obtain the channel state information in each received WiFi training signal as the corresponding channel state signal;

[0007] Preprocess the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix;

[0008] Input the channel state real part matrix and the channel state imaginary part matrix into the corresponding preset convolutional neural network for processing, to obtain a first feature map corresponding to the channel state real part matrix and a second feature map corresponding to the channel state imaginary part matrix;

[0009] Stitch the first feature map and the second feature map to obtain a stitched feature map;

[0010] Input the stitched feature map into a deformable convolutional network in a predetermined direction for feature extraction, and perform a feature flattening operation to connect to a fully connected layer to obtain a final feature classification result. The predetermined direction is the time dimension direction;

[0011] Train a model according to the feature classification result, the preset motion posture information of the preset target, and the preset multi-class cross-entropy loss function to obtain a target action detection model based on WiFi signals.

[0012] In one embodiment, the above-mentioned target action detection method further includes:

[0013] Performing target object detection on the actual WiFi signal according to the target action detection model to identify the motion posture information of the preset target.

[0014] In one embodiment, the steps of preprocessing the channel state signal to obtain the channel state real part matrix and the channel state imaginary part matrix include:

[0015] Separating the real part and the imaginary part of the subcarrier packet included in the channel state signal respectively;

[0016] Extracting the real part of each path of the channel state signal and performing normalization processing to form the channel state real part matrix;

[0017] Extracting the imaginary part of each path of the channel state signal to form the channel state imaginary part matrix.

[0018] In one embodiment, the steps of extracting the real part of each path of the channel state signal and performing normalization processing to form the channel state real part matrix include:

[0019] Arranging the separated real parts in two directions of the signal reception order and the time sequence and performing normalization processing to obtain the channel state real part matrix.

[0020] In one embodiment, the preset convolutional neural network adopts a lightweight convolutional neural network, and the lightweight convolutional neural network includes a first asymmetric depthwise separable convolutional network and a second asymmetric depthwise separable convolutional network connected in series.

[0021] In one embodiment, the structures of the first asymmetric depthwise separable convolutional network and the second asymmetric depthwise separable convolutional network are the same.

[0022] In one embodiment, the first asymmetric depthwise separable convolutional network includes a depthwise separable convolutional network with an N*N dimension and a depthwise separable convolutional network with an N*N 2 dimension, where N is the number of paths of the signal receiving unit.

[0023] In addition, a target action detection device based on WiFi signals is further provided, which is applied to a signal receiving device. The signal receiving device includes at least three signal receiving units, and there is a motion event of a preset target within the preset receiving distance range of the WiFi signal. The target action detection device includes:

[0024] A signal acquisition unit, which acquires the channel state information in each path of the WiFi received training signal as the corresponding channel state signal;

[0025] A signal matrix generation unit, configured to preprocess a channel state signal to obtain a real part matrix of the channel state and an imaginary part matrix of the channel state;

[0026] A feature map generation unit, configured to input the real part matrix of the channel state and the imaginary part matrix of the channel state into corresponding preset convolutional neural networks for processing, to obtain a first feature map corresponding to the real part matrix of the channel state and a second feature map corresponding to the imaginary part matrix of the channel state;

[0027] A splicing unit, configured to splice the first feature map and the second feature map to obtain a spliced feature map;

[0028] A classification result generation unit, configured to input the spliced feature map into a deformable convolutional network in a predetermined direction for feature extraction, and perform a feature flattening operation to connect to a fully connected layer, to obtain a final feature classification result;

[0029] A model establishment unit, configured to perform model training according to the feature classification result and a preset multi-class cross-entropy loss function, to obtain a target action detection model based on WiFi signals.

[0030] In addition, a terminal device is further provided. The terminal device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the above-mentioned target action detection method are implemented.

[0031] In addition, a storage medium is further provided. A computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the above-mentioned target action detection method are implemented.

[0032] The above-mentioned target action detection method based on WiFi signals is applied to a signal receiving device. The signal receiving device includes at least three signal receiving units. There is a motion event of a preset target within the preset receiving distance range of the WiFi signal. The target action detection method includes obtaining the channel state information in each received WiFi training signal as the corresponding channel state signal, preprocessing the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix, inputting the channel state real part matrix and the channel state imaginary part matrix into corresponding preset convolutional neural networks for processing to obtain a first feature map corresponding to the channel state real part matrix and a second feature map corresponding to the channel state imaginary part matrix, splicing the first feature map and the second feature map to obtain a spliced feature map, inputting the spliced feature map into a deformable convolutional network in a predetermined direction for feature extraction, and performing a feature flattening operation to connect to a fully connected layer to obtain a final feature classification result. The predetermined direction is the time dimension direction. According to the feature classification result, the preset motion posture information of the preset target, and a preset multi-class cross-entropy loss function, model training is performed to obtain a target action detection model based on WiFi signals. Since there is a motion event of a preset target within the preset receiving distance range of the WiFi signal, the received signal is affected by the motion of the preset target, that is, the corresponding channel state information in the received WiFi training signal changes. By obtaining the channel state information in each received WiFi training signal as the corresponding channel state signal and preprocessing the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix, the motion posture information of the preset target is further extracted, and finally a target action detection model based on WiFi signals is trained to realize the detection of the motion posture of the preset target object. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the steps of the target action detection method based on WiFi signals in an embodiment of the present application;

[0034] Figure 2 Schematic diagram of the steps of the target action detection method based on WiFi signals in another embodiment of the present application;

[0035] Figure 3 Schematic diagram of the method steps for extracting the channel state real part matrix and the channel state imaginary part matrix in an embodiment of the present application;

[0036] Figure 4 Schematic block diagram of the structure of the target action detection device based on WiFi signals provided in an embodiment of the present application;

[0037] Figure 5 Schematic internal structure block diagram of a terminal device in an embodiment of the present application.

[0038] The realization of the purpose of this application, its functional features and advantages will be further described in conjunction with the embodiments with reference to the accompanying drawings. Detailed implementation manners

[0039] The embodiments of the present application will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals indicate the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and should not be construed as a limitation to the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0040] In addition, if the descriptions such as "first" and "second" are involved in the present application, they are only for descriptive purposes (such as for distinguishing the same or similar elements), and should not be construed as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments may be combined with each other, but it must be based on the fact that those of ordinary skill in the art can implement them. When the combination of the technical solutions appears to be contradictory or unable to be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present application.

[0041] As Figure 1 shown, a target action detection method based on WiFi signals is provided, which is applied to a signal receiving device. The signal receiving device includes at least three signal receiving units. There is a motion event of a preset target within the preset receiving distance range of the WiFi signal. The target action detection method includes:

[0042] Step S110, obtaining the channel state information in each path of WiFi received training signal as the corresponding channel state signal for each path.

[0043] Since there is a motion event of a preset target within the preset receiving distance range of the WiFi signal, the received signal is affected by the motion of the preset target, that is, the corresponding channel state information in the WiFi received training signal changes. Therefore, it is necessary to obtain the channel state information in each path of WiFi received training signal as the corresponding channel state signal for each path.

[0044] The channel state signal is presented in the form of a carrier signal. Taking three antennas as an example, the first antenna x receives 100 sub-carrier packets, the second antenna y receives 100 sub-carrier packets, and the third antenna z receives 100 sub-carriers.

[0045] Among them, the signal receiving unit can be a device with signal receiving function, such as a device with a router AP, a wireless network card, etc. The three-way signal receiving unit can be integrated on the same device or dispersed on different devices.

[0046] In one embodiment, the channel state signal motion event represents a target posture made by someone within a preset receiving distance range, such as motion events like running, falling, and walking.

[0047] Step S120: Preprocess the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix.

[0048] Step S130: Input the channel state real part matrix and the channel state imaginary part matrix into corresponding preset convolutional neural networks for processing to obtain a first feature map corresponding to the channel state real part matrix and a second feature map corresponding to the channel state imaginary part matrix.

[0049] Step S140: Concatenate the first feature map and the second feature map to obtain a concatenated feature map.

[0050] After the preliminary feature extraction in step S130, use the concat operation to concatenate the first feature map and the second feature map to obtain a concatenated feature map.

[0051] Step S150: Input the concatenated feature map into a deformable convolutional network in a predetermined direction for feature extraction, and perform a feature flattening operation to connect to a fully connected layer to obtain a final feature classification result. The predetermined direction is the time dimension direction.

[0052] After obtaining the concatenated feature map, use a deformable convolution in a predetermined direction to perform further feature extraction on the feature map. Here, we propose the concept of a deformable convolution in a predetermined direction. The introduction of this convolution method is to better handle the signal relationship in the unknown time dimension. The traditional deformable convolution adds two offset variables on the basis of ordinary convolution, and these offset variables respectively output the offsets of each point in the x and y directions during the current convolution. Because in our application, we only focus on the offset in the time dimension, therefore, we still use a convolution of (N*N 2 ), combined with the offset amount of each point in the time dimension (N*N 2 ), to calculate a new feature map. However, performing an unconstrained convolution deformation operation in the time dimension may lead to problems with network convergence (such as the oscillation of the offset part of the deformation during training causing the features extracted by the convolution to be concentrated in a small part of the feature map). Therefore, during model initialization, we add a small offset to the initial value of the offset amount for lateral diffusion so that it can be effectively dispersed in the time dimension during training.

[0053] The specific implementation of deformable convolution in a predetermined direction is to add an offset variable during training. Specifically, the optimization method of traditional convolution is to optimize w, where x and y are the input and output respectively, that is, y = wx;

[0054] Deformable convolution in a predetermined direction processes x to form y = wx(p), where p represents the offset of x in the y direction, and this offset can be trained. Example: The meaning of the offset is how much to offset in the y direction when taking the value of a certain x0. If p is 1, then it becomes using x1 (the adjacent value x1 to the right of x0) to multiply with the w weight to obtain the output tensor y. If p is 0.5, then calculate the interpolation of x0 and x1, and then continue with the operation of y = wx.

[0055] In one embodiment, for example, the initial values of the offset can be set as follows:

[0056] 0, 0, 0, 0.05, 0.05, 0.05, 0.1, 0.1, 0.1

[0057] 0, 0, 0, 0.05, 0.05, 0.05, 0.1, 0.1, 0.1

[0058] 0, 0, 0, 0.05, 0.05, 0.05, 0.1, 0.1, 0.1

[0059] In this way, during the initial training of the deformable convolution in a predetermined direction, the convolution will tend to first try to extract the feature map in the direction of the time sequence, and finally stably converge to a globally optimal convolution form.

[0060] Among them, the understanding of the offset variable in the y direction: Because when we reprocess the channel state signal into a matrix form, the y-axis represents the time dimension. Therefore, a group of channel state signals becomes 300 feature points distributed in time (taking three receiving antennas as an example). And p in the above description is an operation on the position. If the original 3*9 convolution covers a certain 27 points, through p (since it is a matrix and also has 27 values), the offset x for each point is recalculated, and a new set of x can be obtained. Then, the convolution weight w is used to calculate with this set of x to obtain the final y = wx(p). The y direction here represents the time axis direction.

[0061] Step S160, perform model training according to the feature classification result, the preset motion posture information of the preset target, and the preset multi-class cross-entropy loss function to obtain a target action detection model based on WiFi signals.

[0062] In this embodiment, since there is a motion event of a preset target within the preset reception distance range of the WiFi signal, the received signal is affected by the motion of the preset target, that is, the corresponding channel state information in the WiFi received training signal changes. By obtaining the channel state information in each path of the WiFi received training signal as the corresponding channel state signal, preprocessing the channel state signal to obtain the channel state real part matrix and the channel state imaginary part matrix, and then extracting the motion posture information of the preset target, finally training a target action detection model based on the WiFi signal to detect the motion posture of the preset target object.

[0063] In one embodiment, as Figure 2 shown, the above target action detection method further includes:

[0064] Step S170, detecting the target object for the actual WiFi signal according to the target action detection model to identify the motion posture information of the preset target.

[0065] In one embodiment, as Figure 3 shown, step S120 includes:

[0066] S121, respectively splitting the real part and the imaginary part of the subcarrier packet included in the channel state signal.

[0067] S122, extracting the real part of each path of the channel state signal and performing normalization processing to form the channel state real part matrix.

[0068] S123, extracting the imaginary part of each path of the channel state signal to form the channel state imaginary part matrix.

[0069] In one embodiment, step S122 includes:

[0070] Arranging the split real parts in two directions of the signal reception order and the time sequence and performing normalization processing to obtain the channel state real part matrix.

[0071] In this embodiment, further split the real part and the imaginary part of the subcarrier packet and arrange them according to the signal reception order and the time sequence. For example, the real parts of 100 subcarrier packets received by the first antenna x, the real parts of 100 subcarrier packets received by the second antenna y, and the real parts of 100 subcarriers received by the third antenna z: x1, x2, x3...x 100 ; y1, y2...y 100 ; z1, z2...z 100 ; Similarly, the subcarrier imaginary parts are xi1, xi2, xi3...xi 100 ; yi1, yi2...yi 100 ; zi1, zi2...zi 100Normalize the real part of the channel state signal received by each antenna to form a relative amplitude, such as xi_norm = x / max(xi). The imaginary part does not need to be normalized. Arrange the processed channel state signals in two forms of 3 rows and 100 columns according to the real part and the imaginary part to form a channel state real part matrix and a channel state imaginary part matrix. Therefore, the shape of each channel state real part matrix and channel state imaginary part matrix is 3 * 100.

[0072] In this embodiment, the signal receiving device includes at least three antennas (illustrated with three antennas here), and the channel state signal data format is to obtain 1 * 3 * 100 subcarrier packets at every preset time interval (taking three seconds as an example). 1 is the number of antennas transmitting signals, 3 is the number of antennas receiving signals, and the subcarrier packets are represented in complex form, where the real part represents the signal strength and the imaginary part represents the phase.

[0073] In one embodiment, the preset convolutional neural network adopts a lightweight convolutional neural network, and the lightweight convolutional neural network includes a first asymmetric depthwise separable convolutional network and a second asymmetric depthwise separable convolutional network connected in series.

[0074] In this embodiment, the reason for adopting the asymmetric depthwise separable convolutional network is that the signal relationship established based on time series reflected between different columns in the channel state real part matrix and the channel state imaginary part matrix. Therefore, it is necessary to use asymmetric convolution to more fully abstract the temporal correlation between features.

[0075] In one embodiment, the first asymmetric depthwise separable convolutional network and the second asymmetric depthwise separable convolutional network have the same structure.

[0076] In this embodiment, the asymmetric depthwise separable convolutional network is used twice continuously for feature extraction, and then a padding operation is performed when using the convolutional kernel to ensure that the shape of the feature map remains unchanged.

[0077] In one embodiment, the first asymmetric depthwise separable convolutional network includes a depthwise separable convolutional network with an N * N dimension and a depthwise separable convolutional network with an N * N 2 dimension, where N is the number of paths of the signal receiving unit.

[0078] In addition, as Figure 4 shown, there is also provided a target action detection device 200 based on WiFi signals, which is applied to a signal receiving device. The signal receiving device includes at least three signal receiving units. There is a motion event of a preset target within the preset receiving distance range of the WiFi signal. The target action detection device 200 includes:

[0079] A signal acquisition unit 210, which acquires the channel state information in each WiFi received training signal as the corresponding channel state signal.

[0080] A signal matrix generation unit 220 is configured to preprocess the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix.

[0081] A feature map generation unit 230 is configured to respectively input the channel state real part matrix and the channel state imaginary part matrix into corresponding preset convolutional neural networks for processing, to obtain a first feature map corresponding to the channel state real part matrix and a second feature map corresponding to the channel state imaginary part matrix.

[0082] A splicing unit 240 is configured to splice the first feature map and the second feature map to obtain a spliced feature map.

[0083] A classification result generation unit 250 is configured to input the spliced feature map into a deformable convolutional network in a predetermined direction for feature extraction, and perform a feature flattening operation to connect to a fully connected layer, to obtain a final feature classification result.

[0084] A model establishment unit 260 is configured to perform model training according to the feature classification result and a preset multi-class cross-entropy loss function to obtain a target action detection model based on WiFi signals.

[0085] In addition, a terminal device is further provided. The terminal device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the above-mentioned target action detection method are implemented.

[0086] In addition, a storage medium is further provided. A computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the above-mentioned target action detection method are implemented.

[0087] In addition, in an embodiment of the present application, a terminal device is further provided. The internal structure of the terminal device may be as Figure 5 shown. The terminal device includes a processor, a memory, a communication interface, and a database connected through a system bus. Wherein, the processor is configured to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the terminal device is configured to store data called by the computer program. The communication interface of the terminal device is configured to perform data communication with an external terminal. The input device of the terminal device is configured to receive signals input by an external device. When the computer program is executed by the processor, a method as described in the above embodiments is implemented.

[0088] Those skilled in the art can understand, Figure 5The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the terminal device to which the solution of this application is applied.

[0089] In addition, this application also provides a readable storage medium, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the target action detection method as described in the above embodiments. It can be understood that the readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0090] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided in this application and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be obtained in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0091] It should be noted that in this article, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, device, article or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, device, article or method including that element.

[0092] The above are only the preferred embodiments of the present application, and do not thereby limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall similarly be included within the patent protection scope of the present application.

Claims

1. A target motion detection method based on WiFi signals, characterized in that: Applied to a signal receiving device, the signal receiving device includes at least three signal receiving units, there is a motion event of a preset target within a preset receiving distance range of the WiFi signal, and the target motion detection method includes: Acquire channel state information in each WiFi receiving training signal as the corresponding channel state signal; Preprocessing the channel state signal to obtain a channel state real matrix and a channel state imaginary matrix; The step of preprocessing the channel state signal to obtain a channel state real matrix and a channel state imaginary matrix comprises: Separating the real part and the imaginary part of the subcarrier packet contained in the channel state signal respectively; Extracting the real part of each channel state signal and performing normalization processing to form a channel state real part matrix; Extracting the imaginary part of each channel state signal to form a channel state imaginary part matrix; The step of extracting the real part of each channel state signal and performing normalization processing to form a channel state real part matrix includes: Arrange the split real parts according to the signal receiving order and timing direction and perform normalization processing to obtain the channel state real part matrix; Inputting the channel state real matrix and the channel state imaginary matrix into the corresponding preset convolutional neural network for processing, respectively, to obtain a first feature map corresponding to the channel state real matrix and a second feature map corresponding to the channel state imaginary matrix, wherein the preset convolutional neural network adopts a lightweight convolutional neural network, and the lightweight convolutional neural network includes a first asymmetric depth-separable convolutional network and a second asymmetric depth-separable convolutional network connected in series; Splicing the first feature map and the second feature map to obtain a spliced ​​feature map; The spliced ​​feature map is input into a deformable convolutional network in a predetermined direction for feature extraction, and a feature flattening operation is performed to connect the fully connected layer to obtain a final feature classification result, wherein the predetermined direction is a time dimension direction, and the deformable convolutional network in the predetermined direction is implemented by adding an offset only in the time dimension direction when training the corresponding convolutional network, so that the deformable convolutional network can be effectively dispersed in the time dimension direction during training and preferentially converges stably to a global optimal convolutional form in the time dimension direction; Performing model training according to the feature classification result, the preset motion posture information of the preset target, and the preset multi-classification cross entropy loss function to obtain a target motion detection model based on the WiFi signal; Performing target object detection on the actual WiFi signal according to the target motion detection model to identify and obtain motion posture information of the preset target; The posture detection unit is used to perform target object detection on the actual WiFi signal according to the target motion detection model to identify and obtain the motion posture information of the preset target.

2. The target motion detection method according to claim 1, characterized in that: The first asymmetric depth-wise separable convolutional network and the second asymmetric depth-wise separable convolutional network have the same structure.

3. The target motion detection method according to claim 2, characterized in that: The first asymmetric depth-wise separable convolutional network includes a depth-wise separable convolutional network of N*N dimensions and a N*N 2 A depth-wise separable convolutional network, where N is the number of signal receiving units.

4. A target motion detection device based on WiFi signals, characterized in that: Applied to a signal receiving device, the signal receiving device includes at least three signal receiving units, there is a motion event of a preset target within a preset receiving distance range of the WiFi signal, and the target motion detection device includes: The signal acquisition unit is used to obtain the channel state information of each WiFi receiving training signal as the corresponding channel state signal; A signal matrix generating unit, used for preprocessing the channel state signal to obtain a channel state real part matrix and a channel state imaginary part matrix; The signal matrix generating unit comprises: A splitting subunit, used to split the real part and the imaginary part of the subcarrier packet contained in the channel state signal respectively; A real part matrix generating unit, used for extracting the real part of each channel state signal and performing normalization processing to form a channel state real part matrix; An imaginary part matrix generating unit, used for extracting the imaginary part of each channel state signal to form a channel state imaginary part matrix; The real part matrix generating unit comprises: arranging the split real parts according to the signal receiving order and the timing direction and performing normalization processing to obtain the channel state real part matrix; A feature graph generating unit, used for inputting the channel state real matrix and the channel state imaginary matrix into corresponding preset convolutional neural networks for processing, to obtain a first feature graph corresponding to the channel state real matrix and a second feature graph corresponding to the channel state imaginary matrix, wherein the preset convolutional neural network adopts a lightweight convolutional neural network, and the lightweight convolutional neural network includes a first asymmetric depth-separable convolutional network and a second asymmetric depth-separable convolutional network connected in series; A splicing unit, used for splicing the first feature map and the second feature map to obtain a spliced ​​feature map; A classification result generating unit, used for inputting the spliced ​​feature map into a deformable convolutional network in a predetermined direction for feature extraction, and performing a feature flattening operation to connect the fully connected layer to obtain a final feature classification result, wherein the predetermined direction is a time dimension direction, and the deformable convolutional network in the predetermined direction is implemented by adding an offset only in the time dimension direction when training the corresponding convolutional network, so that the deformable convolutional network can be effectively dispersed in the time dimension direction during training and preferentially converges stably to a global optimal convolutional form in the time dimension direction; A model building unit, used for performing model training according to the feature classification result and a preset multi-classification cross entropy loss function to obtain a target motion detection model based on the WiFi signal; The posture detection unit is used to perform target object detection on the actual WiFi signal according to the target motion detection model to identify and obtain the motion posture information of the preset target.

5. A terminal device, characterized in that: The terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the target motion detection method according to any one of claims 1 to 3 are implemented.

6. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the target motion detection method according to any one of claims 1 to 3 are implemented.

Images