Millimeter Wave Radar Human Fall Behavior Recognition Method and System Based on Multi-Class 3D Features and Transformer

Abstract

A millimeter-wave radar human fall behavior recognition method and system based on multi-class three-dimensional features and Transformer provided by the present invention. For the radar echo signal of a human target, by analyzing the signal, three types of three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features are stacked, and a fall behavior recognition network based on Transformer and cross-attention is constructed to achieve human fall behavior recognition. For the three-dimensional signal, the network divides it into a fixed number of numerical array blocks, and uses a linear projection network to map it into a fixed-dimensional feature vector. The obtained feature matrix is input into the Transformer module to realize the abstraction of signal representation, and then it is fused with other signal representations through cross-attention to obtain fused features for the classifier to realize fall behavior recognition. This method can achieve a more comprehensive and complete three-dimensional behavior feature expression under the condition of limited data, and Transformer has strong feature extraction ability, with a high fall behavior recognition rate.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and particularly to a millimeter-wave radar human fall behavior recognition method based on multi-class three-dimensional features and Transformer. Background Art

[0002] Human fall behavior recognition is an important research direction in the field of radar signal processing and applications, and is widely used in nursing homes, medical institutions, and the homes of the elderly living alone.

[0003] The injuries caused by falls are major public health problems. If the fall behavior recognition system can detect and issue an alarm in time when the elderly fall, medical staff or family members can be notified immediately, so as to obtain timely medical assistance and avoid secondary injuries caused by the inability to obtain assistance for a long time after the fall; therefore, the research on fall behavior recognition can effectively reduce the severe injury rate and mortality rate of the elderly after falling, and effectively protect the physical and mental health of the elderly.

[0004] Currently, the technologies available for human fall behavior recognition include wearable devices and non-wearable devices. Wearable devices monitor the limb behaviors and postural information of the elderly in real time by having the elderly wear devices equipped with detection components such as acceleration sensors and gyroscopes, and then design recognition methods to determine whether the elderly have fallen. It has the characteristics of being unaffected by the environment and high accuracy, but these devices need to be worn by the elderly in real time, which has a certain impact on the lives of the elderly. Non-wearable devices are divided into visible imaging and invisible imaging. Visible imaging devices include infrared cameras, video surveillance, etc. These devices, as optical instruments, have a high accuracy rate in the field of fall behavior recognition, but they are easily affected by the observation environment and there are privacy leakage problems. Therefore, in recent years, the research focus in related fields has gradually shifted to invisible imaging.

[0005] Therefore, a human fall behavior recognition method and system with high recognition efficiency are needed. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a millimeter-wave radar human fall behavior recognition method and system based on multi-class three-dimensional features and Transformer. This method uses radar signals to achieve human fall behavior recognition, and it is a multi-class three-dimensional feature and Transformer-based millimeter-wave radar human fall behavior recognition method with complete feature expression, high efficiency, and high accuracy.

[0007] To achieve the above purpose, the present invention provides the following technical solutions:

[0008] The millimeter-wave radar human fall behavior recognition method based on multi-class three-dimensional features and Transformer provided by the present invention includes the following steps:

[0009] (1) Install a millimeter-wave radar in the detection area, and there are a determined number of targets within the radar coverage area;

[0010] (2) Through the millimeter-wave radar module, monitor the targets in the detection area in real time, collect the radar echo signals returned by the targets, and upload them to the behavior characterization module of the system at the same time. Generate three types of three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features for the same target and the same behavior, and construct a data set according to different targets and different behaviors;

[0011] (3) Design a fall behavior recognition network based on Transformer spatio-temporal feature extraction and cross-attention feature fusion with multi-representation input, then use the data set to train and test the network, and design a feature extraction and behavior recognition module based on this network;

[0012] (4) In the feature extraction and behavior recognition module, classify the target behavior according to the input fusion features, and judge in real time whether the target has fallen; if so, enter step 5; if not, return to step 4;

[0013] (5) Alarm the remote monitoring platform through the communication module to achieve real-time fall behavior recognition.

[0014] Furthermore, the specific installation details of the millimeter-wave radar used in step (1) are:

[0015] Install the millimeter-wave radar within the detection area, at a height of 2.0 m to 3.0 m; the angle with the vertical direction is about 25° to 35°.

[0016] Furthermore, step (2) is specifically as follows:

[0017] (21) Preprocess the original echo signals collected by the radar;

[0018] (22) Perform short-time fractional Fourier transforms of different orders on the preprocessed signals, and stack the fractional-order time-frequency spectrograms of different orders to obtain a fractional-order three-dimensional feature representation;

[0019] (23) Perform S-transform, smoothed pseudo-Wigner-Ville distribution and downsampled short-time Fourier transform on the preprocessed signals, and stack these different types of second-order time-frequency spectrograms to form a second-order three-dimensional feature representation;

[0020] (24) Perform short-time bispectral transform with a time sliding window on the preprocessed signals, and stack these high-order time-frequency spectrograms in chronological order to obtain a high-order three-dimensional feature representation.

[0021] Further, step (21) is specifically as follows:

[0022] Perform static clutter suppression on the original echo signal. Here, the phasor mean cancellation algorithm is adopted. First, average all received pulses to obtain a reference received pulse, and then subtract the reference received pulse from each received pulse beam to obtain the target echo signal.

[0023] Further, step (22) is specifically as follows:

[0024] First, frame and window the preprocessed signal, then perform fractional Fourier transforms of different orders on the windowed data, and finally stack these fractional-order time-frequency spectrograms according to a fixed order to obtain a fractional-order three-dimensional feature representation.

[0025] Further, step (23) is specifically as follows:

[0026] First, perform an S transform on the signal, that is, first add a Gaussian window with a height and width that vary with frequency to the signal, and then perform a Fourier transform; secondly, perform an SPWVD on the signal, that is, window, smooth, and perform a Wigner-Ville distribution on the signal; finally, perform a DS-STFT on the signal, that is, first downsample the signal and then perform a short-time Fourier transform; finally, stack these different types of second-order time-frequency spectrograms to form a second-order three-dimensional feature representation.

[0027] Further, step (24) is specifically as follows:

[0028] First, window the original signal, then perform a bispectrum transform, and let the window slide on the original signal matrix to obtain multiple high-order time-frequency spectrograms. Finally, stack these time-frequency spectrograms in chronological order to obtain a high-order three-dimensional feature representation.

[0029] Further, step (3) is specifically as follows:

[0030] First, divide the three-dimensional signal into a fixed number of numerical array blocks, and use a linear projection network to map it into a feature vector with a fixed dimension. Input the obtained feature matrix into the Transformer structure to realize the abstraction of signal representation, then perform cross-self-attention fusion with other signal representations to obtain a fused feature, and finally use logistic regression to classify the feature.

[0031] The millimeter-wave radar human fall behavior recognition system based on multi-class three-dimensional features and Transformer provided by the present invention includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above method is implemented.

[0032] The beneficial effects of the present invention are as follows:

[0033] A millimeter-wave radar human fall behavior recognition method and system based on multi-class three-dimensional features and Transformer provided by the present invention obtains radar detection targets in the detection area, collects radar echo signals returned by the targets, and uploads them to the behavior characterization module of the system at the same time. Three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features are generated for the same target and the same behavior, and a data set is constructed according to different targets and different behaviors; a fall behavior recognition network is established, and then the data set is used to train and test the network, and a feature extraction and behavior recognition module is designed based on this network; in the feature extraction and behavior recognition module, the target behavior is classified according to the input fusion features, and it is judged in real time whether the target has fallen; an alarm is sent to the remote monitoring platform through the communication module to achieve real-time fall behavior recognition.

[0034] For the radar echo signal of the human target, the present invention forms three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features by performing short-time fractional Fourier transform, second-order time-frequency analysis, and short-time high-order bispectrum transform on the signal, and constructs a fall behavior recognition network based on Transformer spatio-temporal feature extraction and cross-attention feature fusion, so as to realize human fall behavior recognition.

[0035] This method uses an intelligent recognition network with the idea of cyclic recursion in the RNN. This network divides the three-dimensional signal into numerical array blocks with a fixed number, and uses a linear projection network to map it into a feature vector with a fixed dimension. The obtained feature matrix is input into the Transformer structure to realize the abstraction of signal representation, and then it is fused with other signal representations through cross-attention to obtain fusion features for the classifier to realize fall behavior recognition; through the cyclic recursive local cross-attention module and the Transformer module, it has the feature fusion of any type or different types of signal representations, greatly enhancing the network's feature fusion ability and scalability. This method can achieve a more comprehensive and complete three-dimensional behavior feature expression under the condition of limited data, and the Transformer has strong feature extraction ability and a high fall behavior recognition rate.

[0036] The present invention uses radar for fall behavior recognition, which has the advantages of protecting user privacy and being unaffected by the environment; the three-dimensional feature expression generated by multi-class time-frequency spectrograms can achieve a more comprehensive and complete behavior feature expression under the condition of limited data.

[0037] Other advantages, objectives and features of the present invention will be described to some extent in the subsequent specification, and to some extent, they will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. Brief Description of the Drawings

[0038] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0039] Figure 1 It is a flowchart of a millimeter-wave radar human fall behavior recognition method based on multi-class three-dimensional features and Transformer;

[0040] Figure 2 It is a schematic diagram of the installation of a millimeter-wave radar;

[0041] Figure 3 It is a schematic diagram of multi-class three-dimensional characterization technology;

[0042] Figure 4 It is a fall behavior recognition network structure based on the fusion of Transformer and cross-attention features;

[0043] Figure 5 It is a confusion matrix obtained after testing the network;

[0044] Figure 6 It is a block diagram of a non-line-of-sight human behavior recognition system. Detailed Embodiments

[0045] The following further describes the present invention in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited do not limit the present invention.

[0046] Embodiment 1

[0047] As an environmental sensor, radar can not only effectively avoid the problem of privacy infringement, but also has unique advantages such as strong penetration ability and high range resolution. Especially, it has broad application prospects in detecting and tracking human targets in different motion states. The radar-based fall behavior recognition method mainly includes three steps: feature expression, feature extraction, and classification recognition. First, the radar echo is processed to form an expression image containing behavior feature information, then the feature information contained in the expression image is manually extracted or automatically extracted by an intelligent network, and finally a classifier is designed to classify the behavior according to the feature information, so as to recognize the fall behavior.

[0048] Radar fall behavior recognition mostly forms a time-frequency spectrogram through time-frequency analysis to complete feature expression. The time-frequency spectrogram can be regarded as a power spectrum sequence that changes with time, reflecting the Doppler modulation effect of the movement of multiple scattering parts of the human target on the radar signal. Using different time-frequency analysis methods will generate time-frequency spectrograms with different resolution and aggregation characteristics, and there are differences in the spatial texture features and temporal transformation features of the expression of the same behavior, that is, different time-frequency spectrograms show complementarity in the feature expression of the same behavior. In existing research, one or several two-dimensional time-frequency spectrograms are mostly used for feature expression. In order to fully explore and utilize the complementary spatio-temporal features contained in multiple feature expression image sequences, the present invention generates three types of three-dimensional feature expressions using two-dimensional feature images to more completely express behavior features in a higher-dimensional space and improve the accuracy of human fall behavior recognition.

[0049] The feature extraction methods in radar fall behavior recognition mainly include manual extraction and automatic extraction by intelligent networks. Manual feature extraction obtains useful information from data through artificially designed feature extraction methods and utilizes it. However, in actual operation, the method of manual feature extraction has high requirements for professional knowledge, is difficult to extract high-level discriminant information from the original image, and highly depends on a specific classification environment, with the defect of low efficiency. To improve the efficiency and effectiveness of feature extraction, the present invention proposes a fall behavior recognition network based on Transformer spatio-temporal feature extraction and cross-attention feature fusion, and uses deep learning methods to achieve fall behavior recognition.

[0050] This embodiment uses radar signal processing and deep learning technologies to achieve human fall behavior recognition. Methods such as STFrFT, ST, SPWVD, DS-STFT, and short-time bispectrum transformation are used to extract the time-frequency features of different behaviors of the target, generate multiple types of three-dimensional feature spectrograms, and perform feature extraction, feature fusion, and behavior classification based on the fall behavior recognition network of Transformer spatio-temporal feature extraction and cross-attention feature fusion, and finally achieve fall behavior recognition. This method has the characteristics of protecting user privacy, being unaffected by the environment, and having complete feature expression, and can achieve a better recognition rate.

[0051] As Figure 1 shown, this embodiment provides a millimeter-wave radar non-line-of-sight human behavior recognition method based on multi-class feature fusion, including the following steps:

[0052] (1) Install a millimeter-wave radar in the detection area, and there are a determined number of targets within the radar coverage area;

[0053] (2) The millimeter-wave radar module is used to monitor the targets in the detection area in real time, collect the radar echo signals returned by the targets, and upload them to the behavior characterization module of the system at the same time. Three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features are generated for the same target and the same behavior, and a dataset is constructed according to different targets and different behaviors.

[0054] (3) Design a fall behavior recognition network based on Transformer spatio-temporal feature extraction and cross-attention feature fusion with multi-representation input, then use the dataset to train and test the network, and design a feature extraction and behavior recognition module based on this network.

[0055] (4) In the feature extraction and behavior recognition module, classify the target behavior according to the input fusion features, and judge in real time whether the target has fallen; if so, go to step 5; if not, return to step 4.

[0056] (5) Alarm the remote monitoring platform through the communication module to achieve real-time fall behavior recognition.

[0057] As Figure 2 shown, in the above step (1), to enable the radar to achieve the best measurement effect, the millimeter-wave radar is installed on the wall, with a height of 2.0 m to 3.0 m; the angle with the vertical direction is about 25° to 35°; in this embodiment, it is preferably 2.5 m from the ground and tilted downward by 30°. This millimeter-wave radar is the FMCW 6843ISK radar of Texas Instruments.

[0058] As Figure 3 shown, the above step (2) has the following specific steps:

[0059] (21) Preprocess the original echo signals collected by the radar.

[0060] (22) Perform STFrFT of different orders on the preprocessed signals, and stack the fractional-order time-frequency spectrograms of different orders to obtain a fractional-order three-dimensional feature representation.

[0061] (23) Perform ST, SPWVD, and DS-STFT on the preprocessed signals, and stack these different types of second-order time-frequency spectrograms to form a second-order three-dimensional feature representation.

[0062] (24) Perform short-time bispectral transform with a time sliding window on the preprocessed signals, and stack these high-order time-frequency spectrograms in chronological order to obtain a high-order three-dimensional feature representation.

[0063] The specific content of the above step (21) is as follows:

[0064] Static clutter suppression is performed on the original echo signal. Here, the phasor mean cancellation algorithm is adopted. First, the average of all received pulses is calculated to obtain the reference received pulse, and its calculation formula is:

[0065]

[0066] where m is the sampling point in the fast time dimension, i is the sampling point in the slow time dimension, C[m] represents the reference received pulse; R[m,i] represents the data in the m-th row and i-th column; N represents the number of received pulses;

[0067] Then, subtracting the reference received pulse from each received pulse beam can obtain the target echo signal, and its calculation method is:

[0068] S(m,n) = R(m,n) - C(m)

[0069] where S(m,n) represents the target echo signal; R(m,n) represents the pulse data in the m-th row and n-th column; C(m) represents the reference received pulse;

[0070] The specific content of step (22) is:

[0071] First, frame and window the preprocessed signal, then perform fractional Fourier transforms of different orders on the windowed data, and finally stack these fractional-order time-frequency spectrograms according to a fixed order to obtain the fractional-order three-dimensional feature expression, which can be obtained through the following calculation formula:

[0072]

[0073] where s(τ) is the target echo signal, is the window function, α is the transformation angle, K α (t,u) is the transformation kernel, STFrFT α (t,u) is the obtained fractional Fourier transform result.

[0074] K α The specific content of K

[0075]

[0076] where n is an integer;

[0077] The specific content of step (23) is:

[0078] First, perform an S transform on the signal, that is, first add a Gaussian window with a height and width varying with frequency to the signal, and then perform a Fourier transform, and its calculation method is:

[0079]

[0080] Among them, STs(m, f) is the result of the S transform; s(m, n) is the data of the m-th row and n-th column of the target echo; f represents the frequency;

[0081] Secondly, perform SPWVD on the signal, that is, window the signal, smooth it, and perform the Wigner-Ville distribution (WVD), namely:

[0082]

[0083] Finally, perform DS-STFT on the signal, that is, first downsample the signal and then perform the short-time Fourier transform. The short-time Fourier transform of the m-th column after extraction is:

[0084]

[0085] Among them, STFTs(m, f) represents the result of the short-time Fourier transform; s(k, n) represents the data of the k-th row and n-th column of the target echo; w(m - k) is the window function.

[0086] Finally, stack these second-order time-frequency spectrograms of different categories to form a second-order three-dimensional feature representation.

[0087] The specific content of step (24) is as follows:

[0088] First, window the original signal, then divide the observed data sequence of known length into K segments, perform zero-mean preprocessing on the segmented data, and then calculate the third-order cumulant estimate c k (i, j), namely:

[0089]

[0090] Among them, k = 1, 2,..., K, s k (n + i) is the k-th segment of data, M is the number of observed samples in each segment, M1 = max(0, -i, -j), M2 = min(M - 1, M - 1 - i, M - 1 - j),

[0091] Next, find the mean value of the third-order cumulants of all segments Namely:

[0092]

[0093] Finally, calculate the bispectrum estimate of this frame of data The calculation method is:

[0094]

[0095] Among them, l < M - 1, w(i, l) is the two-dimensional lag window function, ω1, ω2 are frequencies.

[0096] Slide the time window over the original data to obtain the bispectrum estimation of multiple frames of data, and then stack the spectrograms of each frame of data to obtain a high-order three-dimensional feature representation.

[0097] Specifically, step (3) is as follows:

[0098] The feature extraction and fusion steps based on the Transformer and cross-attention mechanism are as Figure 4 shown. For three-dimensional signals, divide them into a fixed number of N numerical array blocks, and use a linear projection network to map them into feature vectors with a fixed dimension of C. Input the obtained N×C feature matrix into the Transformer structure to realize the abstraction of signal representation, and then perform cross-attention fusion with other signal representations to obtain fused features. Then input the fused features into a logistic regression model for classification, and finally realize fall behavior recognition.

[0099] In Figure 4 's network structure, the three feature abstraction modules respectively complete the dimensionality reduction of three high-dimensional features, the feature fusion module is responsible for fusing the three feature representations, and the classifier realizes the classification of the fused features, thereby realizing fall behavior recognition.

[0100] In Figure 4 's network structure, the fractional-order three-dimensional feature abstraction module obtains the feature vector Q1 through linear transformation of the fractional-order three-dimensional features; the second-order three-dimensional feature abstraction module obtains the feature vectors K1 and V1 through linear transformation of the second-order three-dimensional features; the high-order three-dimensional feature abstraction module obtains the feature vectors K2 and V2 through linear transformation of the high-order three-dimensional features. Input the feature vectors Q1, K1, and V1 into the feature fusion module 1, and fuse the input features through the cross-attention mechanism and the Transformer encoder to obtain the feature vector Q2. Then input the feature vector Q2, K2, and V2 into the feature fusion module 2, and also use the cross-attention mechanism and the Transformer encoder to fuse the input features to obtain the final fused features. Finally, input the fused features into the classifier for classification, thereby realizing fall behavior recognition.

[0101] Use 200 groups of actual fall data and 600 groups of non-fall (sitting down while walking, squatting down while walking, stopping while walking) actual data to train the network, and then use 100 groups of fall data and 100 groups of non-fall data to test the network. The confusion matrix obtained is as Figure 5 shown, where Actual class is the actual action type, Predict class is the recognition result of this network, non-fall is non-fall, fall is fall. It can be seen from the confusion matrix that the recognition rate of non-fall actions is 90%, the recognition rate of fall actions is 96%, and the total recognition rate can reach 93%.

[0102] Embodiment 2

[0103] The millimeter-wave radar human fall behavior recognition system based on multi-class three-dimensional features and Transformer provided in this embodiment includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above method is implemented.

[0104] As Figure 6 shown, Figure 6 It is a non-line-of-sight human behavior recognition system block diagram, including a millimeter-wave radar module, a behavior characterization module, a feature extraction and behavior recognition module, a communication module, and a remote monitoring platform;

[0105] The millimeter-wave radar module is used to collect the original echo data of the target;

[0106] The behavior characterization module is used to form a multi-class three-dimensional feature expression of the target behavior;

[0107] The feature extraction and behavior recognition module is used to extract and fuse the features in the target three-dimensional representation and classify the target behavior to achieve fall behavior recognition;

[0108] The communication module is used to transmit data and instructions between the radar module and the remote monitoring platform for communication;

[0109] The remote monitoring platform is used to display the monitoring results and issue an alarm when a fall occurs, facilitating the guardians to handle it in time and provide help to the fallen person.

[0110] The feature extraction and behavior recognition module is implemented through feature extraction and fusion based on Transformer and cross-attention mechanism. The obtained three-dimensional signal is divided into a fixed number of N numerical array blocks, and a linear projection network is used to map it into a feature vector with a fixed dimension of C. The obtained N×C feature matrix is input into the Transformer structure to realize the abstraction of signal representation, and then it is fused with other signal representations through cross-attention to obtain the fused feature. Then, the fused feature is input into a logistic regression model for classification, and finally, fall behavior recognition is achieved.

[0111] The above embodiments are only preferred embodiments cited to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer, characterized in that: It includes the following steps: (1) Install a millimeter-wave radar in the detection area, and there are a determined number of targets within the radar coverage area; (2) Use the millimeter-wave radar module to monitor the targets in the detection area in real time, collect the radar echo signals returned by the targets, and upload them to the behavior characterization module of the system at the same time. Generate three types of three-dimensional spectrograms containing different time-frequency aggregation features and three-dimensional correlation features for the same target and the same behavior, and construct a dataset according to different targets and different behaviors; (2) specifically is: (21) Preprocess the original echo signals collected by the radar; (22) Perform short-time fractional Fourier transforms of different orders on the preprocessed signals, and stack the fractional-order time-frequency spectrograms of different orders to obtain a fractional-order three-dimensional feature representation; (23) Perform S-transform, smoothed pseudo-Wigner-Ville distribution, and downsampled short-time Fourier transform on the preprocessed signals, and stack these different types of second-order time-frequency spectrograms to form a second-order three-dimensional feature representation; (24) Perform short-time bispectral transform with a time sliding window on the preprocessed signals, and stack these high-order time-frequency spectrograms in chronological order to obtain a high-order three-dimensional feature representation; (3) Design a fall behavior recognition network based on Transformer spatio-temporal feature extraction and cross-attention feature fusion with multi-representation inputs, then use the dataset to train and test the network, and design a feature extraction and behavior recognition module based on this network; (3) specifically is: First, divide the three-dimensional signal into a fixed number of numerical array blocks, and use a linear projection network to map it into a feature vector with a fixed dimension. Input the obtained feature matrix into the Transformer structure to realize the abstraction of signal representation, then perform cross-self-attention fusion with other signal representations to obtain the fused features, and finally use logistic regression to classify the features; (4) In the feature extraction and behavior recognition module, classify the target behavior according to the input fused features, and judge in real time whether the target has fallen; if so, go to step (5); if not, return to step (4); (5) Alarm to the remote monitoring platform through the communication module to realize real-time fall behavior recognition.

2. The method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer according to claim 1, characterized in that: (1) The specific installation details of the millimeter-wave radar used are: Install the millimeter-wave radar within the detection area, with a height of 2.0 m to 3.0 m; the angle with the vertical direction is 25° to 35°.

3. The method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer according to claim 1, characterized in that: (21) specifically is: Perform static clutter suppression on the original echo signals. Here, the phasor mean cancellation algorithm is used. First, average all received pulses to obtain a reference received pulse, and then subtract the reference received pulse from each received pulse beam to obtain the target echo signal.

4. The method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer according to claim 1, characterized in that: (22) specifically is: First, frame and window the preprocessed signals, then perform short-time fractional Fourier transforms of different orders on the windowed data, and finally stack these fractional-order time-frequency spectrograms of different orders according to a fixed order to obtain a fractional-order three-dimensional feature representation.

5. The method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer according to claim 1, characterized in that: (23) specifically is: First, perform the S transform on the signal, that is, first add a Gaussian window with height and width varying with frequency to the signal, and then perform the Fourier transform; secondly, perform the SPWVD on the signal, that is, perform windowing, smoothing, and Wigner-Ville distribution on the signal; finally, perform the DS-STFT on the signal, that is, first downsample the signal and then perform the short-time Fourier transform; finally, stack these different types of second-order time-frequency spectrograms to form a second-order three-dimensional feature representation.

6. The method for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer according to claim 1, characterized in that: Step (24) is specifically as follows: First, window the original signal, then perform the bispectrum transform, and let the window slide on the original signal matrix to obtain multiple high-order time-frequency spectrograms. Finally, stack these time-frequency spectrograms in chronological order to obtain a high-order three-dimensional feature representation.

7. A system for identifying human fall behavior using a millimeter-wave radar based on multi-class three-dimensional features and Transformer, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that,When the processor executes the program, it implements the method described in any one of claims 1 to 6 above.

Images