Gesture recognition method and system based on mobile communication signal

Abstract

The invention belongs to the technical field of wireless communication, signal processing and intelligent sensing, and relates to a gesture recognition method and system based on a mobile communication signal, and the method comprises the steps: carrying out the scene judgment of a to-be-recognized gesture, determining a scene with a sensing range within a set threshold value as a close-distance scene, or else, determining the scene as a long-distance scene; for a short-distance scene, extracting CRS data in a single carrier of a mobile communication signal, and carrying out smooth denoising processing by adopting a sliding window filter; for a long-distance scene, extracting and superposing CRS data between different carriers of a mobile communication signal; using a long-time window filter to calculate a static component; calculating to obtain a dynamic component according to the static component; and performing short-time Fourier transform on the dynamic component, constructing a gesture spectrogram, and inputting the gesture spectrogram into a trained convolutional neural network to realize gesture recognition. According to the invention, mobile communication signals are adopted, the coverage range is wider, the network stability is higher, the method is suitable for application scenes with more distances and distances, and the recognition accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, signal processing and intelligent perception, and in particular to a gesture recognition method and system based on mobile communication signals. Background Art

[0002] As a natural human-computer interaction method, gesture recognition technology brings new interactive experience and convenient operation methods to users in the fields of smart home, smart driving, entertainment games, medical rehabilitation and public safety, and has wide application potential. The core principle of gesture recognition technology is to use methods to extract feature information related to gestures from data collected by sensors, cameras or wireless devices, and convert it into instructions that computers can understand. Existing gesture recognition technologies are mainly divided into three methods: vision-based, sensor-based and wireless signal-based. The vision-based gesture recognition method uses a camera to capture hand images or video sequences, combines image processing and deep learning methods for analysis, and recognizes and interprets the shape, position and dynamic changes of the hand; the sensor-based gesture recognition technology captures the physical information of hand movement (such as acceleration, angular velocity, orientation, etc.), converts it into digital signals, and processes and analyzes these signals through methods to identify specific gestures; the gesture recognition method based on wireless signals analyzes the changes of wireless signals such as millimeter waves and WiFi under the influence of gesture movements to perceive and recognize different gestures. With the development of integrated communication and perception technology, how to use communication signals (such as WiFi signals and mobile communication signals) for gesture perception while achieving efficient communication has become a research hotspot in the future. Among them, WiFi-based solutions are usually limited to indoor environments, and in order to achieve high sampling rates, communication resources need to be occupied, thereby interfering with normal data transmission. In contrast, mobile communication signals have a wider coverage range and higher network stability, and are suitable for more application scenarios. Therefore, the present invention chooses to use mobile communication signals as the carrier of gesture perception. However, this also has some challenges. For example, when gestures are performed in different positions, directions or long-distance scenes, the original spectrum diagram will be significantly different due to the influence of the scene, which in turn affects the recognition accuracy of the system.

[0003] In order to explain the principle more clearly, three typical cases are listed, such as Figure 1 As shown in (a), target one and target two are to perform the same gesture at different positions and directions in the perception area, and target three is to perform the same gesture at the boundary of the perception area. The present invention performs preliminary denoising on the original CRS (cell reference signal) information of different carriers collected. The original CRS information in different carriers and the preliminary denoising results are shown in FIG. Figure 1The initial denoising result is subjected to short-time Fourier transform (STFT, Short-Time Fourier Transform), and the original spectrum obtained is as follows Figure 1 As shown in (c) in the figure. It can be found that when target one and target two perform the same gesture at different positions and directions, the original spectrum graph corresponding to the CRS information in the mobile communication system is very different due to the inconsistent size of the dynamic components generated by the gesture in different directions and the different sizes of the static components after the signal is reflected from the environment at different positions. For target three, since the position where the same gesture is performed is far away from the transceiver, the system cannot directly capture the changes in the original CRS information caused by the gesture, resulting in the dynamic component information of the gesture being annihilated by noise, and the obtained original spectrum cannot directly reflect the gesture information. Based on the above phenomenon, it can be found that the original spectrum graph of the CRS information cannot represent the intrinsic characteristics of the gesture perception task. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the prior art and propose a gesture recognition method and system based on mobile communication signals.

[0005] In order to achieve the above object, the present invention provides a gesture recognition method based on mobile communication signals, comprising:

[0006] Step 1) The scene of the gesture to be recognized is judged. If the perception range is within the set threshold, it is a close-range scene, otherwise it is a long-range scene;

[0007] Step 2) For close-range scenarios, extract the CRS data in a single carrier of the mobile communication signal and use a sliding window filter to perform smoothing and denoising; for long-range scenarios, extract and superimpose the CRS data between different carriers of the mobile communication signal;

[0008] Step 3) using a long time window filter to calculate the static component;

[0009] Step 4) Calculate the dynamic component based on the static component;

[0010] Step 5) Perform short-time Fourier transform on the dynamic component, construct a gesture spectrum diagram, and input it into the trained convolutional neural network to realize gesture recognition.

[0011] Preferably, the mobile communication system is an LTE system, comprising two omnidirectional antennas, one for transmitting and one for receiving.

[0012] Preferably, the threshold value set in step 1) is 30 square meters.

[0013] Preferably, in step 2), for a close-range scene, the signal after smoothing and denoising is Satisfy the following formula:

[0014]

[0015] Where n∈[1,N] represents the index of the sampling point, N represents the number of sampling points in the sample, is the first point of the sliding window, represents the last point of the sliding window, L(n)=k2-k1+1 represents the number of sampling points in the sliding window, and W is the window length parameter.

[0016] Preferably, the step 2) extracts and superimposes CRS data between different carriers of mobile communication signals for long-distance scenarios; including:

[0017] For long-distance scenarios, based on the correlation of information between multiple carriers of mobile communication signals, CRS data between different carriers are extracted and superimposed to achieve signal quality enhancement.

[0018] Preferably, the static component obtained in step 3) is Satisfy the following formula:

[0019]

[0020] in, is the first point of the long time window, Indicates the last point of the long time window, L long (n) = k long2 -k long1 +1 indicates the number of samples in the long time window, W long is the window length parameter.

[0021] Preferably, the dynamic component obtained in step 4) is Satisfy the following formula:

[0022]

[0023] Preferably, the convolutional neural network of step 5) includes four layers, the first three layers are composed of a convolutional layer, a batch normalization layer, a ReLU activation layer and a maximum pooling layer; the fourth layer is composed of a convolutional layer, a batch normalization layer and a ReLU activation layer; the number of convolution kernels in each layer is 16, 32, 64 and 256 respectively, the convolution kernel size is 3*3, the step size and padding size are both 1, and the classifier includes a first fully connected layer of 100 neurons, a second fully connected layer of 5 neurons, and a Softmax layer of 5 class labels.

[0024] Preferably, the input of the convolutional neural network is a gesture spectrum diagram, and the output is a gesture type, including:

[0025] The "push" gesture, the "rectangle" gesture, the "circle" gesture, the "push and pull" gesture, and the "X" gesture.

[0026] In another aspect, the present invention provides a gesture recognition system based on mobile communication signals, comprising:

[0027] The near and far scene judgment module is used to judge the scene of the gesture to be recognized. If the perception range is within the set threshold, it is a near scene, otherwise it is a far scene;

[0028] The CRS data processing module is used to extract the CRS data in a single carrier of a mobile communication signal for close-range scenarios, and use a sliding window filter for smoothing and denoising; it is used to extract and superimpose the CRS data between different carriers of a mobile communication signal for long-range scenarios;

[0029] A static component calculation module, used for calculating the static component using a long time window filter;

[0030] A dynamic component calculation module, used for calculating the dynamic component according to the static component;

[0031] The gesture recognition module is used to perform short-time Fourier transform on the dynamic component, construct a gesture spectrum diagram, and input the trained convolutional neural network to realize gesture recognition.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] 1. The present invention proposes a signal quality enhancement method based on LTE (Long Term Evolution) signals. The method enhances the signal quality by superimposing CRS information in multiple carriers, making the noise distribution more concentrated to improve the signal quality, thereby enhancing the perception range of the system and realizing long-distance gesture recognition;

[0034] 2. The present invention adopts a dynamic component capture method, which is based on the CRS signal. The method uses a sliding window filter to denoise the original CRS signal and extract the static component, and then subtracts the static component from the denoised signal to obtain the dynamic component information reflecting the intrinsic nature of the gesture action;

[0035] 3. Since the amplitude and trend of the dynamic component are related to the position and direction of human gestures, the present invention performs short-time Fourier transform on the obtained dynamic component to generate a gesture intrinsic spectrum diagram, and then recognizes it through a convolutional neural network to achieve position and direction-independent gesture recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the difference in the spectrogram of the original CRS information of the same gesture at different positions, directions or distances;

[0037] Figure 2 is a flow chart of the method of the present invention;

[0038] Figure 3 is a comparison of the effects of dynamic component capture methods, where Figure 3(a) is Figure 1 The initial denoising results of the original CRS information of target 1 and target 2 performing the "push" gesture action. Figure 3(b) is the result after processing by the dynamic component capture method;

[0039] Figure 4 is a visual comparison of the original spectrogram and the gesture spectrogram;

[0040] FIG5 shows the noise statistical characteristics and CRS amplitude information before and after the perceived distance enhancement method, wherein FIG5(a) shows the noise statistical characteristics and FIG5(b) shows the CRS amplitude information;

[0041] FIG6 is a comparison of the effects of the original spectrum diagram and the gesture spectrum diagram at a long distance, where FIG6(a) is the original spectrum diagram at a long distance, and FIG6(b) is the gesture spectrum diagram. DETAILED DESCRIPTION

[0042] The present invention proposes a signal quality enhancement method based on mobile communication signals and a method for constructing a gesture spectrum diagram, which solves the technical problem of inconsistency of the original spectrum diagram of the gesture recognition system based on mobile communication signals in different gesture positions, directions or long-distance scenes, and realizes the consistency of the features of the same gesture at any position, direction or long distance, and the difference of the features of different gestures. The core idea of ​​the present invention is to enhance the signal quality of the dynamic component of the gesture by superimposing the CRS information of multiple carriers in the mobile communication system, thereby increasing the gesture perception range to realize the long-distance perception task. Then, the dynamic component related to the gesture is extracted from the CRS information, and a gesture spectrum diagram is constructed to represent the intrinsic motion pattern of the gesture that is independent of the scene.

[0043] In daily gesture recognition tasks, the system's perception scenarios can be divided into two situations: close distance and long distance. The close distance scenario refers to the target being within the system's initial perception range (after testing and evaluation, the initial perception range of this system covers an area of ​​30 square meters); the long distance scenario refers to the target being beyond the initial perception range (after testing and evaluation, the system's perception range covers an area of ​​up to 228 square meters after signal quality enhancement). In view of these two different scenarios, the present invention proposes a robust gesture recognition method based on gesture spectrum diagram, as follows.

[0044] The workflow of the present invention is as follows Figure 2As shown, it is generally divided into a signal quality enhancement stage and a gesture spectrum diagram construction stage. In the signal quality enhancement stage, for gesture recognition tasks at any position and direction in close-range scenes, the present invention can complete the perception task by directly using the CRS information in a single carrier. For gesture recognition tasks at any position and direction in long-range scenes, the present invention chooses to superimpose the CRS amplitude information of 30 different subcarriers in the mobile communication system. According to the Wiener-Khinchin law of large numbers, the distribution of noise can be significantly reduced based on the superposition of multiple independent and identically distributed samples, highlighting the CRS information related to gestures. In the gesture spectrum construction stage, the present invention first uses a smoothing filter with a window length of 100 to further eliminate the interference of noise based on the CRS information processed in the signal quality enhancement stage, and then uses a long-term sliding window filter with a window length of 400 to calculate the static component related to the environment, and then subtracts the static component from the denoised data to obtain the dynamic component of the gesture action, and then obtains the CRS information containing the dynamic component related to the gesture movement process, and finally performs a short-time Fourier transform based on the dynamic component of the CRS to obtain a gesture spectrum diagram that can represent the intrinsic characteristics of the gesture, and inputs it into a pre-established and trained convolutional neural network to obtain the gesture type to complete the gesture recognition task.

[0045] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0046] Example 1

[0047] Example 1: Gesture recognition at any position and direction in a close-range scene

[0048] The system configuration is as follows:

[0049] 1. The system works on the YunSDR Y750 software radio platform;

[0050] 2. The bandwidth of the LTE system is 20MHz, and the system sampling rate is 30.72MHz. Only CRS data in a single carrier is collected, and the number of CRS information sample points per second is 2000. LTE Long Term Evolution (Long Term Evolution,) is the long-term evolution of the Universal Mobile Telecommunications System (UMTS) technical standard developed by the 3rd Generation Partnership Project (3GPP).

[0051] 3. The system uses two omnidirectional antennas, one for transmitting and one for receiving. The initial sensing capability of the device covers a range of 30m. 2 .

[0052] Task: Construct a gesture spectrum graph that is independent of the position and direction of the gesture in close-range scenarios.

[0053] For gesture recognition, the present invention uses the dynamic component generated only by the target gesture action as the intrinsic information of the gesture spectrum. Specifically, in the wireless communication channel, in addition to LOS propagation, there is also a multipath effect caused by environmental diffraction, reflection, refraction and scattering. When there is a human gesture in the propagation environment, the signal reflected or scattered by the arm and palm will form a dynamic component. Therefore, the signal at the receiver can be further written as:

[0054] h(n)=h s (n)+h d (n)+h n (n)

[0055] where h s (n) represents the static component due to environmental reflection, h d The term (n) represents the dynamic component due to the reflection of human gestures, h n Item (n) represents noise. Since the length of the signal environment reflection path is inconsistent in different positions and directions of the same gesture, the waveform of the CRS information and the original spectrum of the waveform will be different after the dynamic components generated by the noise and gesture are superimposed.

[0056] In order to solve this problem, the present invention adopts a dynamic component capture method to filter out the noise component h n (n) and the static component h s (n) interference, and only focus on the dynamic component h generated by the gesture action d Specifically, the present invention first performs smoothing and denoising on the original CRS data h(n) to obtain a signal after denoising. The formula is as follows:

[0057]

[0058] Where n∈[1,N] represents the index of the sampling point, N represents the number of sampling points in the sample,

[0059] is the first point of the sliding window, represents the last point of the sliding window, L(n)=k2-k1+1 represents the number of sampling points in the sliding window, and W is the window length parameter, which is set to 100 in the present invention. Figure 1 The results of preliminary denoising of the original CRS information collected by target 1 and target 2 are shown in Figure 3(a). It can be seen that for the same gestures performed in different positions and directions, their differences are very large.

[0060] For the static component The present invention uses a sliding window filter to For filtering, the calculation formula is as follows:

[0061]

[0062] in is the first point of the long time window, Indicates the last point of the long time window, L long (n) = k long2 -k long1 +1 indicates the number of samples in the long time window,

[0063] W long is the window length parameter, which is set to 400 sampling points in the present invention. Subtract the static component To get the dynamic component of the gesture

[0064]

[0065] The result after processing by the dynamic component capture method is shown in Figure 3(b). Since the "push" gesture has a motion phase, the CRS information also has a dynamic component fluctuation phase after processing by the method. Therefore, this method can effectively extract the intrinsic motion information of the gesture, so that the same gesture has similar characteristic rules in different positions and directions.

[0066] Since the amplitude and trend of the dynamic component are related to the size, position, and direction of the human gesture, the dynamic component of the same gesture in different scenes will be different, thus affecting the accuracy of gesture recognition. Perform a short-time Fourier transform, setting the time window length to N s = 1500, a frequency resolution of about 1.34 Hz ​​is obtained, and finally a gesture spectrum diagram related to the gesture action is obtained. Figure 4As shown, the upper row of figures are the original spectrum graphs of gestures, and the lower row of figures are the spectrum graphs of gestures obtained by the method proposed in the present invention. The gesture recognition method adopted is through a convolutional neural network, and the obtained gesture spectrum graph is used as the input of the network. Specifically, we designed a convolutional neural network based on deep learning. The network consists of four layers, and the number of convolution kernels in each layer is 16, 32, 64 and 256 respectively, the convolution kernel size is 3*3, and the step size and padding size are 1. The first three layers are composed of a convolution layer, a batch normalization layer, a ReLU activation layer and a maximum pooling layer. The fourth layer is composed of a convolution layer, a batch normalization layer, and a ReLU activation layer. The classifier consists of a fully connected layer 1 with 100 neurons, a fully connected layer 2 with 5 neurons, and a Softmax layer of 5 class labels.

[0067] The currently selected gesture types include but are not limited to the following five: "push" gesture, "rectangle" gesture, "circle" gesture, "push and pull" gesture and "X" gesture.

[0068] in, Figure 4 (a) and (b) are the original frequency spectrum diagram and gesture spectrum diagram of target one and target two performing the “push” gesture action, respectively. It can be seen that the proposed method of the present invention can significantly reduce the influence of gesture position and direction, making the same gesture features consistent. Figure 4 (c)-(f) are the original frequency spectrograms and gesture spectrograms of drawing the “rectangle”, “circle”, “push and pull” and “X” gesture actions respectively. It can be seen that the characteristics of the method proposed in the present invention are significantly different for different gesture actions.

[0069] Example 2: Gesture recognition at any position and direction in a long-distance scene

[0070] The system configuration is as follows:

[0071] 1. The system works on the YunSDR Y750 software radio platform;

[0072] 2. The bandwidth of the built LTE system is 20MHz, and the system sampling rate is 30.72MHz. The subcarrier frequency interval is 15kHz, with a total of 1200 subcarriers. CRS data from 30 carriers is collected, and the number of CRS information sample points per second is 2000.

[0073] 3. The system uses two omnidirectional antennas, one for transmitting and one for receiving. After the signal quality is enhanced, the system's sensing coverage range is 228m 2 .

[0074] Task: Construct a gesture spectrum graph that is independent of the position and direction of the gesture within the enhanced perception coverage of the system.

[0075] As we all know, the measurement noise of CRS information follows the rule of zero mean and variance σ 2 (f) Additive white Gaussian distribution. When the position of the gesture is far away from the transceiver, the noise power will be greater than the dynamic power. At this time, the change of the dynamic component caused by the human gesture will be buried in the noise, so it is difficult to extract the change of CRS caused by the human gesture. Due to the influence of noise, the directly obtained CRS change cannot accurately reflect the gesture.

[0076] In order to solve the above problems, the present invention proposes a signal quality enhancement method. The idea of ​​this method is that multiple samplings of independent and identically distributed noises and then superposition can significantly reduce the noise level. And this idea conforms to the Wiener-Khinchin large number theorem: Let X n , n=1,2,... are independent and identically distributed random variables. As the number of samples n increases, the average value of all samples will converge to the expectation, as shown in the following formula:

[0077]

[0078] Since the LTE system can divide the signal into multiple subcarriers in the frequency domain, the present invention uses the correlation between the information of multiple carriers to superimpose the CRS information between different carriers, increase the number of data samples, and achieve the purpose of enhancing the quality of the CRS signal.

[0079] Specifically, in the signal quality enhancement stage, the present invention selects to use 30 subcarriers. First, collect the CRS information of each carrier. Second, superimpose the CRS information of different subcarriers to make the noise distribution narrower. Figure 5(a) shows the Figure 1 The statistical characteristics of the noise before and after processing of the CRS information obtained when the target is at a long distance are shown in Figure 5(b). It can be seen that the noise distribution is narrower after processing. Figure 5(b) shows the CRS amplitude information before and after processing by this method. It can be seen that this method can effectively enhance the dynamic component and can effectively obtain gesture information at a long distance.

[0080] In the gesture spectrum construction stage, firstly, based on the CRS information obtained by the signal quality enhancement method, the dynamic component extraction method is used to extract the dynamic component related to the gesture action, and the static component related to the gesture position, direction and environment is filtered out. Secondly, a gesture spectrum that can represent the intrinsic characteristics of the gesture is constructed, as shown in Figure 6(b).

[0081] The gesture recognition method adopted is through a convolutional neural network, and the obtained gesture spectrum diagram is used as the input of the network. The specific network parameters are the same as those in Example 1.

[0082] Compared with the original spectrum graph 6(a) of target three, the method of the present invention can extract the essential features representing the gesture action at a long distance, and the extracted gesture spectrum is consistent with Figure 4 (a) and (b) also have similar "push" gesture features, reflecting the consistency of the same gesture features.

[0083] Example 3

[0084] Embodiment 3 of the present invention provides a gesture recognition system based on mobile communication signals, which is implemented based on the methods of Embodiment 1 and Embodiment 2, and includes:

[0085] The near and far scene judgment module is used to judge the scene of the gesture to be recognized. If the perception range is within the set threshold, it is a near scene, otherwise it is a far scene;

[0086] The CRS data processing module is used to extract the CRS data in a single carrier of a mobile communication signal for close-range scenarios, and use a sliding window filter for smoothing and denoising; it is used to extract and superimpose the CRS data between different carriers of a mobile communication signal for long-range scenarios;

[0087] A static component calculation module, used for calculating the static component using a long time window filter;

[0088] A dynamic component calculation module, used for calculating the dynamic component according to the static component;

[0089] The gesture recognition module is used to perform short-time Fourier transform on the dynamic component, construct a gesture spectrum diagram, and input the trained convolutional neural network to realize gesture recognition.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A gesture recognition method based on mobile communication signals, comprising: Step 1) The scene of the gesture to be recognized is judged. If the perception range is within the set threshold, it is a close-range scene, otherwise it is a long-range scene; Step 2) For close-range scenarios, extract the CRS data in a single carrier of the mobile communication signal and use a sliding window filter to perform smoothing and denoising; for long-range scenarios, extract and superimpose the CRS data between different carriers of the mobile communication signal; Step 3) using a long time window filter to calculate the static component; Step 4) Calculate the dynamic component based on the static component; Step 5) Perform short-time Fourier transform on the dynamic component, construct a gesture spectrum diagram, and input it into the trained convolutional neural network to realize gesture recognition.

2. The method for gesture recognition based on mobile communication signals according to claim 1, characterized in that: The mobile communication system is an LTE system, including two omnidirectional antennas, one for transmitting and one for receiving.

3. The gesture recognition method based on mobile communication signals according to claim 1, characterized in that: The threshold set in step 1) is 30 square meters.

4. The method for gesture recognition based on mobile communication signals according to claim 1, characterized in that: In step 2), for a close-range scene, the signal after smoothing and denoising is: Satisfy the following formula: Where n∈[1,N] represents the index of the sampling point, N represents the number of sampling points in the sample, is the first point of the sliding window, represents the last point of the sliding window, L(n)=k2-k1+1 represents the number of sampling points in the sliding window, and W is the window length parameter.

5. The method for gesture recognition based on mobile communication signals according to claim 1, characterized in that: The step 2) extracts and superimposes CRS data between different carriers of mobile communication signals for long-distance scenarios; including: For long-distance scenarios, based on the correlation of information between multiple carriers of mobile communication signals, CRS data between different carriers are extracted and superimposed to achieve signal quality enhancement.

6. The method for gesture recognition based on mobile communication signals according to claim 4, characterized in that: The static component obtained in step 3) is Satisfy the following formula: in, is the first point of the long time window, Indicates the last point of the long time window, L long (n) = k long2 -k long1 +1 indicates the number of samples in the long time window, W long is the window length parameter.

7. The method for gesture recognition based on mobile communication signals according to claim 6, characterized in that: The dynamic component obtained in step 4) is: Satisfy the following formula:

8. The method for gesture recognition based on mobile communication signals according to claim 1, characterized in that: The convolutional neural network of step 5) includes four layers, the first three layers are composed of a convolution layer, a batch normalization layer, a ReLU activation layer and a maximum pooling layer; the fourth layer is composed of a convolution layer, a batch normalization layer and a ReLU activation layer; the number of convolution kernels in each layer is 16, 32, 64 and 256 respectively, the convolution kernel size is 3*3, the step size and padding size are both 1, and the classifier includes a first fully connected layer of 100 neurons, a second fully connected layer of 5 neurons, and a Softmax layer of 5 class labels.

9. The method for gesture recognition based on mobile communication signals according to claim 1, characterized in that: The input of the convolutional neural network is a gesture spectrum diagram, and the output is a gesture type, including: a "push" gesture, a "rectangle" gesture, a "circle" gesture, a "push-pull" gesture, and an "X" gesture.

10. A gesture recognition system based on mobile communication signals, characterized in that: include: The near and far scene judgment module is used to judge the scene of the gesture to be recognized. If the perception range is within the set threshold, it is a near scene, otherwise it is a far scene; The CRS data processing module is used to extract the CRS data in a single carrier of a mobile communication signal for close-range scenarios, and use a sliding window filter for smoothing and denoising; it is used to extract and superimpose the CRS data between different carriers of a mobile communication signal for long-range scenarios; A static component calculation module, used for calculating the static component using a long time window filter; A dynamic component calculation module, used for calculating the dynamic component according to the static component; and The gesture recognition module is used to perform short-time Fourier transform on the dynamic component, construct a gesture spectrum diagram, and input the trained convolutional neural network to realize gesture recognition.

Images