A method for gesture recognition using millimeter-wave radar sensing

Abstract

A millimeter-wave radar sensing gesture recognition method based on few-sample learning, which designs action gestures and radar parameters, builds a millimeter-wave radar system platform, and a human body stands in front of the platform collection point to perform gesture changes. The millimeter-wave radar transmits a linear frequency modulation signal, and then receives an echo signal containing gesture information. The transmitted signal and the received signal are mixed to obtain an intermediate frequency signal; then the intermediate frequency signal is subjected to clutter suppression and 2D-FFT preprocessing to construct a distance-time and speed-time data set of the gesture action, and a number of representative distance and speed features are selected as a network input sample set; then the SVM model parameters are optimized using an improved grid search algorithm, and the optimal parameters are selected to implement a few-sample data set training and realize gesture recognition. The present invention achieves a gesture recognition effect with high accuracy. The present invention has broad application prospects and strong practicality.

Description

Technical Field

[0001] The present invention relates to the technical field of human-computer interaction, and particularly to a millimeter-wave radar perception gesture recognition method based on few-shot learning. Background Art

[0002] With the rapid development of artificial intelligence technologies such as neural networks and deep learning, gesture action recognition technology has become one of the research hotspots in fields such as wireless sensing, pattern recognition, computer vision, and signal processing. It has broad market prospects and profound social application value in fields such as security monitoring, human-computer interaction, and safe driving. For example, in home life, people can use air gestures to remotely adjust the volume and switch programs without touching the device, making the user experience more convenient and comfortable; in the field of automotive driving, drivers can use designated dynamic gestures for driving operations to improve safety; in the aspect of entertainment games, players can maximize innovative game operations, making the entertainment methods more diverse and promoting the optimization and upgrading of more entertainment products.

[0003] Most of the existing relatively mature gesture recognition technologies achieve gesture recognition by obtaining gesture feature information through wearable devices or optical images. Among them, the gesture recognition technology based on wearable devices mostly uses sensors or data gloves, etc. for information acquisition, which has problems of poor portability and high device cost; while the gesture recognition technology based on optical images mostly captures gesture information through cameras, which has the problem of high power consumption; in addition, the data of gesture recognition technology is stored in the form of images or videos, which easily leads to problems of user privacy and security, and the gesture recognition technology itself has low information extraction efficiency in complex environments.

[0004] In recent years, with the continuous development of millimeter-wave radar technology, gesture recognition technology based on millimeter-wave radar signals has received increasing attention. Millimeter-wave radar sensors can effectively solve the problem of recognition accuracy affected by conditions such as insufficient light, have a certain function of penetrating and propagating through obstacles, can effectively avoid the influence of occlusion, and at the same time greatly eliminate the hidden danger of user privacy leakage, and can be integrated into high-speed processing chips with low power consumption and small volume. Therefore, researching gesture recognition technology based on millimeter-wave radar has very great significance.

[0005] At present, the gesture recognition method of millimeter-wave radar is mainly based on deep learning algorithms, which has the dependence on large datasets and computational complexity. Gesture recognition based on deep learning is a data-driven algorithm, and it cannot effectively solve problems such as scarce training data and limited computational resources. Generally speaking, a high-performance recognition network often requires a deeper network structure. The larger the parameter scale of the model, the more training samples are needed, resulting in higher computational complexity and a greater risk of overfitting of the model. There are few publicly available datasets for radar-based gesture action recognition, mostly based on self-measured data or simulation data, and it is also difficult to obtain a large amount of radar echo data within a limited time in the experiment. Therefore, how to effectively draw on the transfer learning method in radar gesture action perception and study gesture action recognition training based on a small amount of sample data, so as to reduce the computational complexity and maintain a high recognition rate is one of the problems to be solved. Summary of the Invention

[0006] In order to overcome the deficiencies of the prior art, the present invention proposes a millimeter-wave radar sensing gesture recognition method based on few-shot learning, which can achieve high-accuracy recognition of human gesture actions with few samples and has good generalization ability.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] A millimeter-wave radar sensing gesture recognition method based on few-shot learning, comprising the following steps:

[0009] Step 1: Design action gestures and radar parameters, and build a millimeter-wave radar system platform;

[0010] Step 2: A person stands in front of the platform collection area and makes gesture action changes. The millimeter-wave radar emits a chirp signal, and then receives the echo signal containing gesture information. The transmitted signal and the received signal are mixed to obtain an intermediate-frequency signal;

[0011] Step 3: First perform clutter preprocessing on the digital intermediate-frequency signal, and then use the two-time FFT algorithm to process to obtain the distance and speed feature matrix of the gesture action;

[0012] Step 4: Select the speed and distance features within a set time to construct an action feature dataset for multiple gestures;

[0013] Step 5: Normalize the feature dataset of the action gestures to reduce the training complexity;

[0014] Step 6: Extract a part from the normalized action gesture feature dataset as the training set to train the SVM model;

[0015] Step 7: Optimize the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model using the improved grid search algorithm, repeat Step 6, and continuously repeat the training to obtain a better network model;

[0016] Step 8: Call the trained network model to perform classification and recognition on the validation set.

[0017] Furthermore, in Step 1, the action gesture is an action gesture that includes arm movement, and its motion characteristics are related to the entire hand and arm. The designed radar parameters include the number of transmitting antennas N tx , the number of receiving antennas N rx , the starting frequency f of the frequency modulation signal st , the frequency modulation slope K slope , the total number of frames N frame , the duration T of each frame frame , the number of chirps N within each frame's frequency modulation period chirp , the period T of each chirp chirp , the number of ADC sampling points N within each chirp adc , the ADC sampling frequency F sadc and the sampling time T of the ADC ramp_end , the settings of these parameters are determined based on the maximum measurable distance D_max, maximum measurable speed V_max, distance resolution D_res, and speed resolution V_res of the designed gesture as indicators. The calculation formulas are:

[0018]

[0019] In the formula, c represents the speed of light, λ represents the wavelength corresponding to the frequency modulation center frequency, B represents the frequency modulation bandwidth, and the direct relationships with the various parameters are:

[0020] B = K slope T adc

[0021]

[0022] T ramp_end is the total sampling time of the ADC, including the starting time and effective sampling time of the ADC sampling

[0023]

[0024] T chirp is the period time of each chirp, which consists of T ramp_end and a period of idle time. Therefore, when designing the parameters, it is necessary to follow:

[0025]

[0026] Further, in the step 2, the process of obtaining the intermediate frequency signal of the motion gesture is as follows: Using a millimeter-wave radar operating at 77 - 81 GHz, N tx transmitting antennas transmit FMCW signals with a period, setting the starting frequency f st of the frequency modulation signal, the frequency modulation slope K slope , the total number of frames N frame , the duration T frame of each frame, the number of chirps N chirp within each frame's frequency modulation period, the period T chirp of each chirp, the number of ADC sampling points N adc within each chirp, the ADC sampling frequency F sadc , the sampling time T ramp_end of the ADC and other parameters, N rx receiving antennas receive the reflected echo signals of the human motion gesture. There are a total of N tx ×N rx channels. After the transmitted signal and the reflected signal are mixed, an intermediate frequency signal is obtained, and then it is sampled by a data acquisition board to obtain a digital intermediate frequency signal.

[0027] Furthermore, in the step 3, the process of performing clutter preprocessing on the digital intermediate frequency signal and calculating the distance and speed feature matrices of the gesture motion is as follows:

[0028] a. According to the set radar parameters, the original intermediate frequency signal data collected by the acquisition card is divided into multiple channel data. Each channel data is denoised using the EMD decomposition method, and then each denoised channel data is divided into N frame frame periods, and each frame period is further divided into N chirp chirp signals. Each chirp signal contains N adc complex sampling data;

[0029] b. Perform a one-dimensional fast Fourier transform (i.e., 1D - FFT) on each chirp signal within any channel and any frame period to obtain the one-dimensional range image information of different frequency modulation periods. Then, using the spectral peak position search method and the known intermediate frequency signal frequency, a spectrum with different separated peaks is generated. Each peak represents the presence of an object target at a specific distance, and thus the distance - time characteristics of the gesture target to be measured can be obtained.

[0030] c. Then, perform an FFT on the phase of the signal after 1D - FFT processing at the same distance for each chirp, and the range - Doppler information of the gesture target can be obtained. Use the range image information for registration to obtain targets with different speeds, and finally obtain the speed - time characteristics of different actions.

[0031] In step 4, the process of selecting the speed and distance features of the set time and constructing the action feature dataset of multiple gestures is as follows: by setting a fixed sliding frame window with a length of W frame within each of the N frame frames of each gesture, the distance and speed information of the gesture action is selected every W frame fixed frames to form a distance feature vector and a speed feature vector, which are stored in the dataset; for any gesture feature, each gesture is repeated multiple times, so as to save multiple groups of distance and speed feature vectors of the gesture action, and thus establish a feature dataset of different action gestures for the training of the subsequent gesture classification model.

[0032] In step 6, when using the training set to train the support vector machine SVM network, in order to reduce the training complexity and achieve non-linear mapping, the RBF kernel function in the support vector machine SVM is adopted, and the function parameters are c and g. c is used to control the penalty coefficient of the loss function and needs to be set before training the model; g is the setting of the gamma function in the kernel function, and gamma is equivalent to adjusting the complexity of the model.

[0033] In step 7, the process of optimizing the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model by using the improved grid search algorithm so that the model can have good generalization ability and high accuracy is as follows:

[0034] a. Give the value ranges of c and g according to experience;

[0035] b. Discretize the parameter value grid, set a determined search step size, and create a network along different growth directions of the parameters. The nodes in the network are related parameter pairs;

[0036] c. In the samples to be searched, several discrete values are selected for each parameter, and all possible combinations of c and g are selected for training the model; after searching, the optimal parameters are the best parameters after training;

[0037] d. Substitute the optimal c and g into the training function in the support vector machine SVM, and use the training set divided by the feature dataset to train the network model.

[0038] The beneficial effects of the present invention are as follows:

[0039] 1. The present invention has good generalization ability and strong anti-interference ability, and can be applied to remote control, intelligent driving and other aspects;

[0040] 2. The present invention simultaneously uses distance and speed information as the recognition features of gesture actions, providing a new research direction for the subsequent research;

[0041] 3. The support vector machine algorithm with improved grid search is mainly applied, which can achieve high recognition accuracy in the case of few samples. Brief Description of the Drawings

[0042] Figure 1 is the flowchart of the implementation process of the gesture recognition method of the present invention;

[0043] Figure 2 is the original data format of the radar output signal of the present invention;

[0044] Figure 3 are the schematic diagrams of six gestures defined in the embodiments of the present invention. Among them, (a) forward push, (b) backward swing, (c) wave, (d) backward pull, (e) knock on the door, (f) forward swing;

[0045] Figure 4 are the schematic diagrams of the distance and speed characteristics of six gesture actions defined in the embodiments of the present invention. Among them, (a) forward push, (b) backward swing, (c) wave, (d) backward pull, (e) knock on the door, (f) forward swing;

[0046] Figure 5 is the experimental prediction result diagram of the present invention under the embodiments;

[0047] Figure 6 is the gesture recognition confusion matrix diagram of the present invention under the embodiments;

[0048] Figure 7 is the flowchart of the millimeter wave radar sensing gesture recognition method based on few-shot learning. Detailed Embodiment

[0049] The following elaborates on the preferred embodiments of the present invention in conjunction with the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0050] Refer to Figures 1 to 7 , a millimeter wave radar sensing gesture recognition method based on few-shot learning, includes the following steps:

[0051] Step 1: Design gesture actions and radar parameters, and build a millimeter wave radar system platform;

[0052] Step 2: A person stands in front of the platform collection area, makes gesture action changes, uses the millimeter wave radar to transmit a chirp signal, then receives the echo signal containing gesture information, and performs mixing processing on the transmitted signal and the received signal to obtain an intermediate frequency signal. The original data format of the signal is as Figure 2 shown;

[0053] In step 3, the digital intermediate frequency signal is first preprocessed for clutter using the EMD decomposition method, and then the distance and speed feature matrices of the gesture actions are obtained by processing with the two - time FFT algorithm;

[0054] In step 4, the speed and distance features within a set time are selected to construct the action feature datasets of various gestures;

[0055] In step 6, the feature datasets of the action gestures are normalized to reduce the training complexity;

[0056] In step 9, a part is extracted from the normalized action gesture feature datasets as the training set, and the SVM model is trained using K - fold cross - validation;

[0057] In step 10, the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model are optimized using the improved grid search algorithm, and the process (6) is repeated, and continuous training is carried out to obtain a better network model;

[0058] In step 11, the trained network model is called to perform classification and recognition on the validation set;

[0059] Furthermore, in step 1, the action gesture is an action gesture including arm movement, and its motion characteristics are related to the entire hand and arm. The designed radar parameters include the number of transmitting antennas N tx , the number of receiving antennas N rx , the starting frequency f of the frequency - modulated signal st , the frequency - modulation slope K slope , the total number of frames N frame , the duration T of each frame frame , the number of chirps N within each frame's frequency - modulation period chirp , the period T of each chirp chirp , the number of ADC sampling points N within each chirp adc , the ADC sampling frequency F sadc and the sampling time T of the ADC ramp_end , the settings of these parameters are determined based on the maximum measurable distance D_max, the maximum measurable speed V_max, the distance resolution D_res, and the speed resolution V_res of the designed gesture as indicators. The calculation formulas are as follows:

[0060]

[0061] In the formula, c represents the speed of light, λ represents the wavelength corresponding to the frequency - modulation center frequency, B represents the frequency - modulation bandwidth, and the direct relationships with each parameter are:

[0062] B = K slope T adc

[0063]

[0064] T ramp_end is the total sampling time of the ADC, including the start time and the effective sampling time of the ADC sampling

[0065]

[0066] T chirp is the cycle time of each chirp, which consists of T ramp_end and an idle time. Therefore, when designing parameters, the following should be followed:

[0067]

[0068] Furthermore, in step 2, the process of obtaining the intermediate frequency signal of the action gesture is as follows: Using a millimeter-wave radar operating at 77 - 81 GHz, N tx transmitting antennas transmit FMCW signals with a period, setting the starting frequency f st of the frequency modulation signal, the frequency modulation slope K slope , the total number of frames N frame , the duration T frame of each frame, the number of chirps N chirp within each frame's frequency modulation period, the period T chirp of each chirp, the number of ADC sampling points N adc within each chirp, the ADC sampling frequency F sadc , the sampling time T ramp_end of the ADC and other parameters. After that, N rx receiving antennas receive the reflected echo signals of the human body's action gesture. There are a total of N tx ×N rx channels. After mixing the transmitted signal and the reflected signal, an intermediate frequency signal is obtained, and then it is sampled by a data acquisition board to obtain a digital intermediate frequency signal.

[0069] Even further, in step 3, the process of performing clutter preprocessing on the digital intermediate frequency signal and calculating the distance and speed feature matrix of the gesture action is as follows:

[0070] a. According to the set radar parameters, the original intermediate frequency signal data collected by the acquisition card is divided into multiple channel data. Each channel data is denoised using the EMD decomposition method, and then each denoised channel data is further divided into N frame frame periods. Each frame period is further divided into N chirp chirp signals, and each chirp signal contains N adc complex sampling data;

[0071] b. Perform one-dimensional fast Fourier transform (i.e., 1D-FFT) on each chirp signal within any channel and any frame period to obtain one-dimensional range image information for different frequency modulation periods. Then, using the spectral peak position search method and the known intermediate frequency signal frequency, generate a spectrum with different separated peaks, where each peak indicates the presence of an object target at a specific distance. Thus, the distance-time characteristics of the gesture target to be measured can be obtained.

[0072] c. Then, perform FFT on the phase of the signal after 1D-FFT processing at the same distance for each chirp, and the distance-Doppler information of the gesture target can be obtained. Use the range image information for registration to obtain targets with different speeds, and finally obtain the speed-time characteristics of different actions.

[0073] In step 4, the process of selecting the speed and distance characteristics within the set time and constructing the action feature dataset for multiple gestures is as follows: By setting a fixed sliding frame window with a length of W frame and selecting the distance and speed information of the gesture action every W frame fixed frames within each of the N frame frames of each gesture, form the distance feature vector and the speed feature vector, and store them in the dataset; for any gesture feature, repeat each gesture multiple times, thus saving multiple sets of distance and speed feature vectors of the gesture action, and thus establishing a feature dataset for different action gestures for use in the training of subsequent gesture classification models.

[0074] In step 6, when using the training set to train the support vector machine SVM network, in order to reduce the training complexity and achieve non-linear mapping, the RBF kernel function in the support vector machine SVM is used, and the function parameters are c and g. c is used to control the penalty coefficient of the loss function and needs to be set before training the model; g is the setting of the gamma function in the kernel function, and gamma is equivalent to adjusting the complexity of the model.

[0075] In step 7, the process of using the improved grid search algorithm to optimize the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model so that the model can have good generalization ability and high accuracy is as follows:

[0076] a. Give the value ranges of c and g according to experience;

[0077] b. Discretize the parameter value grid, set a definite search step size, and create a network along different growth directions of the parameters. The nodes in the network are related parameter pairs;

[0078] c. In the samples to be searched, select several discrete values for each parameter, and select all possible combinations of c and g for training the model; after searching, the optimal parameters are the best parameters after training.

[0079] d. Substitute the optimal c and g into the training function in the support vector machine (SVM), and use the training set divided by the feature data set to train the network model.

[0080] In this embodiment, the TI IWR1443BOOST radar and the DCA1000EVM data acquisition card are used to collect echo data. The radar includes 3 transmitting antennas and 4 receiving antennas, but only 1 transmitting antenna and 4 receiving antennas are used in this example.

[0081] In this embodiment, the maximum measurable distance D_max is set to 6.4 m, the maximum measurable speed V_max is 9.5 m / s, and the gesture action time is about 2 s. The designed radar parameters are as follows: the starting operating frequency f st is 77 GHz, the frequency modulation slope K slope is 40 MHz / μs, the total number of frames N frame is 50 frames, the duration of each frame T frame is 40 ms, the frequency modulation period T chirp is 100 μs, the number of chirps N within each frame's frequency modulation period chirp is 128, the number of ADC sampling points N within each chirp adc is 128, the ADC sampling frequency F sadc is 2000 MHz, the sampling time T of the ADC ramp_end is 75 μs. From this, the range resolution D_res can be calculated as 5 cm, and the velocity resolution V_res as 0.15 m / s.

[0082] The gesture actions defined in this embodiment are as Figure 3 shown, including 6 gesture actions: pushing forward, swinging backward, waving, pulling backward, knocking, and swinging forward. The human body stands at a position about 80 cm from the platform acquisition area, performs gesture action changes, and obtains the radar echo intermediate frequency signal.

[0083] The range-time and velocity-time characteristics of the 6 gesture actions obtained by the millimeter-wave radar in this embodiment are as Figure 4 shown. Select the action range and velocity information within a certain time interval, and 50 sets of feature vector data are obtained for the range-time and velocity-time of each action.

[0084] In the network training stage of this embodiment, a support vector machine network is mainly used for training. First, the obtained gesture distance and speed features are converted into a matrix form as the gesture data set, and then the data set is normalized to reduce the computational complexity. According to the predetermined labels, that is, the backswing corresponds to the number "1", the back pull corresponds to the number "2", the wave corresponds to the number "3", the forward swing corresponds to the number "4", the forward push corresponds to the number "5", and the knock corresponds to the number "6", the distance and speed feature information of each gesture is mapped to the training labels one by one. And it is divided into a training set, a validation set, and a test set. The network parameters are optimized through a grid search algorithm to train the training set, a set of optimal parameters is selected, and then network training is performed to save the trained network model.

[0085] Finally, this embodiment tests six gestures, with 5 samples for each gesture. The test results are as Figure 5 shown. The confusion matrix of the recognition results is as Figure 6 shown. In the case of only 5 gesture samples, the recognition rate of most gestures reaches 100%, and the average recognition accuracy reaches 96.7%, verifying the effectiveness of the millimeter-wave radar-based gesture recognition method for few-shot learning proposed by the present invention.

[0086] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for gesture recognition using millimeter-wave radar sensing, characterized in that, the method comprises the following steps: Step 1: Design gesture actions and radar parameters, and build a millimeter-wave radar system platform; Step 2: A person stands in front of the platform acquisition area, performs gesture action changes, uses the millimeter-wave radar to transmit a chirp signal, then receives an echo signal containing gesture information, and mixes the transmitted signal and the received signal to obtain an intermediate-frequency signal; Step 3: First perform clutter preprocessing on the digital intermediate-frequency signal, and then use the two-time FFT algorithm to process to obtain the distance and speed feature matrix of the gesture action; Step 4: Select the speed and distance features within a set time, and construct an action feature dataset for various gestures; Step 5: Normalize the feature dataset of the gesture actions to reduce the training complexity; Step 6: Extract a part from the normalized gesture action feature dataset as a training set to train the SVM model; Step 7: Use an improved grid search algorithm to optimize the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model, repeat Step 6, and continuously repeat the training to obtain a better network model; Step 8: Call the trained network model to perform classification and recognition on the validation set; In the step 1, the action gesture is an action gesture including arm movement, and its movement characteristics are related to the entire hand and arm. The designed radar parameters include the number of transmitting antennas N tx , the number of receiving antennas N rx , the starting frequency f of the frequency-modulated signal st , the frequency modulation slope K slope , the total number of frames N frame , the duration T of each frame frame , the number of chirps N within the frequency modulation period of each frame chirp , the period T of each chirp chirp , the number of ADC sampling points N within each chirp adc , the ADC sampling frequency F sadc and the sampling time T of the ADC ramp_end , the settings of these parameters are determined based on the maximum measurable distance D_max, the maximum measurable speed V_max, the distance resolution D_res, and the speed resolution V_res of the designed gesture as indicators. The calculation formula is as follows: In the formula, c represents the speed of light, λ represents the wavelength corresponding to the center frequency of frequency modulation, B represents the frequency modulation bandwidth, and the direct relationship with each parameter is: B = K slope T adc T ramp_end is the total sampling time of the ADC, including the start time and the effective sampling time adopted by the ADC T chirp is the period time of each chirp, consisting of T ramp_end and an idle time. Therefore, when designing parameters, the following should be followed:

2. A method for gesture recognition using millimeter-wave radar sensing according to claim 1, characterized in that, In the said step 2, the process of obtaining the intermediate frequency signal of the motion gesture is as follows: Using a millimeter-wave radar operating at 77 - 81 GHz, N tx transmitting antennas transmit FMCW signals in a cycle, and set the starting frequency f st of the frequency modulation signal, the frequency modulation slope K slope , the total number of frames N frame , the duration T frame of each frame, the number of chirps N chirp in each chirp period of each frame, the period T chirp of each chirp, the number of ADC sampling points N adc in each chirp, the ADC sampling frequency F sadc , the sampling time T ramp_end of the ADC. After setting these parameters, N rx receiving antennas receive the reflected echo signals of the human motion gesture. There are a total of N tx ×N rx channels. After the transmitted signal and the reflected signal are mixed, an intermediate frequency signal is obtained, and then it is sampled by a data acquisition board to obtain a digital intermediate frequency signal.

3. A method for gesture recognition using millimeter-wave radar sensing according to claim 1, characterized in that, In the said Step 3, the process of performing clutter preprocessing on the digital intermediate-frequency signal and calculating the distance and speed feature matrix of the gesture action is as follows: a. According to the set radar parameters, the original intermediate frequency signal data collected by the acquisition card is divided into multiple channel data. Each channel data is denoised using the EMD decomposition method, and then each denoised channel data is divided into N frame frame periods, and each frame period is further divided into N chirp chirp signals, and each chirp signal contains N adc complex sampling data; b. Perform one-dimensional fast Fourier transform on each chirp signal within any channel and any frame period to obtain one-dimensional range image information of different frequency modulation periods. Then use the spectral peak position search method, use the known intermediate-frequency signal frequency to generate a spectrum with different separated peaks, and each peak represents the existence of an object target at a specific distance, thereby obtaining the distance-time feature of the gesture target to be measured; c. Then perform FFT on the phase of the signal after 1D-FFT processing at the same distance for each chirp, then the range-Doppler information of the gesture target can be obtained, use the range image information for registration to obtain targets with different speeds, and finally obtain the speed-time feature of different actions.

4. A method for gesture recognition using millimeter-wave radar sensing according to claim 1, characterized in that, In the step 4, the process of selecting the speed and distance features of the set time and constructing the action feature dataset of multiple gestures is as follows: by setting a fixed sliding frame window with a length of W frame and selecting the distance and speed information of the gesture action every W frame frames within every N frame frames of each gesture to form a distance feature vector and a speed feature vector, and storing them in the dataset; for any gesture feature, each gesture is repeated multiple times, so as to save multiple groups of distance and speed feature vectors of the gesture action, thereby establishing a feature dataset of different action gestures for the training of the subsequent gesture classification model.

5. A method for gesture recognition using millimeter-wave radar sensing according to claim 1, characterized in that, In the said Step 6, use the training set to train the support vector machine SVM network, In order to reduce the training complexity and achieve non-linear mapping, the RBF kernel function in the support vector machine SVM is adopted, and the function parameters are c and g. c is used to control the penalty coefficient of the loss function and needs to be set before training the model; g is the setting of the gamma function in the kernel function, and gamma is equivalent to adjusting the complexity of the model.

6. A method for gesture recognition using millimeter-wave radar sensing according to claim 5, It is characterized in that In step 7, the penalty factor parameter c and the radial basis kernel function parameter g in the SVM model are optimized by using the improved grid search algorithm, so that the model can have good generalization ability and high accuracy. The process is as follows: a. Give the value ranges of c and g according to experience; b. Discretize the parameter value grid, set a determined search step size, and create a network along different growth directions of the parameters. The nodes in the network are related parameter pairs; c. In the samples to be searched, select several discrete values for each parameter, and select all possible combinations of c and g for training the model; after searching, the optimal parameters are the best parameters after training; d. Substitute the optimal c and g into the training function in the support vector machine SVM, and train the network model with the training set divided by the feature data set.

Images