A Multi-Channel Radar Human Behavior Recognition Method and System Based on Scatter Separation
By employing PCA separation and feature extraction methods for multi-channel radar signals, the problem of strong scattering torso echo masking was solved, thus improving the accuracy of radar human behavior recognition.
Patent Information
- Application Number
- CN202410462147.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-04-17
AI Technical Summary
In existing radar human behavior recognition technology, the strong scattering echo from the torso masks the weak scattering echoes from the limbs and head, thus limiting the behavior recognition performance.
After acquiring multi-channel radar signals, the main scattering component and the secondary scattering component are separated by PCA. After STFT processing, features are extracted using HOG and CNN networks, and classification and recognition are performed in conjunction with an attention mechanism.
It effectively avoids the masking effect of strong scattering echoes from the torso on weak scattering echoes from the limbs and head, thus improving the accuracy of human behavior recognition.
Smart Images

Figure CN118311530B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of behavior recognition technology, and in particular to a multi-channel radar human behavior recognition method and system based on scattering separation. Background Technology
[0002] Human behavior recognition technology is widely used in fields such as intelligent security, smart transportation, healthcare, and human-computer interaction. In general, sensors used for human behavior recognition can be divided into two main categories: wearable and non-wearable. Wearable sensors, such as RFID tags, accelerometers, and gyroscopes used in smart bracelets and watches, can cause user discomfort and inconvenience and cannot be used in non-cooperative human behavior recognition scenarios. Non-wearable sensors mainly include optical, thermal infrared, and radar. Optical sensors are easily affected by lighting conditions and cannot protect privacy, while thermal infrared sensors are easily affected by ambient temperature and are easily blocked. In contrast, radar has the characteristics of all-weather, 24 / 7 operation, is insensitive to ambient light and temperature conditions, poses no privacy risks, and has good penetration through wood, glass, and even walls. These advantages expand the application scenarios of radar-based human behavior recognition technology.
[0003] Because human targets exhibit typical multi-scattering characteristics relative to radar, the strong scattering echoes from the torso can mask the weaker scattering echoes from the limbs and head, thus weakening the comprehensiveness of the expression of human behavioral features and limiting the performance of behavior recognition. Common feature representation methods for human target echoes collected by radar sensors include two-dimensional time-range images, time-frequency spectrograms, and range-Doppler images, as well as three-dimensional time-range-Doppler images, range-azimuth-elevation images, and point clouds. Feature extraction methods are divided into manually constructed algorithms and automatic feature extraction using deep learning. Manual extraction methods rely heavily on expert knowledge and are only suitable for extracting shallow, concrete features, such as the envelope of the time-range image, the histogram of directional gradients in the time-frequency spectrogram, and scale-invariant feature transformations. Deep learning methods, based on neural network models, automatically learn and extract multi-level abstract features from ample training data, providing a more comprehensive and robust description of human behavior, and have become the mainstream feature extraction method. For radar-based human behavior recognition scenarios, statistical learning classifiers such as support vector machines, random forests, and K-nearest neighbors have been used. Neural networks possess powerful nonlinear fitting and learning prediction capabilities, and have been proven to be capable of serving as classifiers for various target recognition tasks. Summary of the Invention
[0004] In view of this, the purpose of the present invention is to provide a multi-channel radar human behavior recognition method and system based on scattering separation, which effectively improves the recognition accuracy.
[0005] To achieve the above objectives, the present invention provides the following technical solution: The multi-channel radar human behavior recognition method based on scattering separation provided by this invention includes the following steps: Step S1: Use radar to collect human behavior data, stack the multi-channel data and perform PCA to separate the main scattering component and the secondary scattering component. Step S2: Substitute the main scattering component and the secondary scattering component into the STFT to obtain the time spectrum; Step S3: Use HOG and CNN networks to extract the feature information of the main scattering component and the secondary scattering component from the time-frequency spectrogram, and then stack them to form human behavior features; Step S4: Apply an attention mechanism to human behavioral characteristics and input them into a classifier to complete classification and recognition.
[0006] Furthermore, the acquisition and processing process of the multi-channel radar signal described in step S1 is as follows: S1.1 First, a multi-channel step-frequency continuous wave radar is used to collect human behavior data and obtain echo signals. S1.2 The radar echo signals of each channel were subjected to MTI mean cancellation to obtain... indivual dimensional matrix ; S1.3 will Channel matrix Extract the vector data from the same column and arrange them together to form a new matrix. Transpose the new matrix to obtain the matrix. ,matrix The columns represent the number of channels, and the rows represent the number of time periods of the echo data; S1.4 The matrix obtained in step S1.3 After performing PCA, the first two principal component components were extracted. and The main scattering component and the secondary scattering component are obtained.
[0007] Furthermore, the radar echo signal mentioned in step S1.1 is: , in, Here are the scattering coefficients for each part. For the echo delay at a certain moment, This refers to the number of parts of the human body that emit light. The starting frequency for transmission. For frequency stepping, The step frequency is the number of points.
[0008] Furthermore, the feature extraction process in step S3 is as follows: S3.1 Perform HOG artificial feature extraction on the time spectrum of the main scattering component; S3.2 The temporal spectrum of the secondary scattering component is fed into a CNN network for automatic feature extraction; S3.3 Stack the two feature vectors obtained in steps S3.1 and S3.2.
[0009] Furthermore, the human behavior data collection in step S1 includes the collection of data on human behavior actions such as jumping, picking up, squatting, stepping, swinging arms, and walking.
[0010] Furthermore, in step S1.4, the first two principal component components are extracted using PCA. and The specific process for obtaining the main scattering component and the secondary scattering component is as follows: PCA transforms the matrix through orthogonal transformation. The original linearly correlated channel vectors are reorganized into a new matrix according to their contribution magnitude. In the uncorrelated row vectors, uncorrelated components are separated from the multi-channel data; The recombination matrix was obtained using PCA. The specific method is to find covariance matrix And find The eigenvalues are arranged in descending order, and the corresponding unit eigenvectors are... Arranged in rows to form a matrix Then the original signal is mapped to a new feature space. In this process, a new component matrix is obtained. That is, the recombination of the original signal, expressed by the formula: The resulting recombination matrix The row vectors are independent of each other and are the original matrix. The linear weighted rearrangement of row vectors, and their arrangement order is the order of their contribution to the observed signal.
[0011] Furthermore, the specific process of focusing the human behavioral features into the feature information obtained by the attention mechanism and then using a classifier to complete the classification and recognition in step S4 is as follows: S4.1 divides the feature information data into m tokens as input, and then processes them through the network weights. generate , and vector, where Then to , Perform a dot product operation to calculate the similarity between the two, then divide by... Square root of matrix dimension This stabilizes the numerical gradient, then uses softmax to convert the similarity calculation results into a probability distribution of the weights, and finally applies the weights to... The calculation result of self-attention is obtained, and the expression for self-attention is as follows: in, , , ; S4.2 The result obtained in step S4.1 is input into the classification layer consisting of two fully connected layers. ReLU and Softmax are selected as activation functions respectively. The Softmax function outputs a result vector with the same dimension as the human behavior category. Each dimension represents the probability value of the input being identified as a certain behavior. The sum of the probabilities of each dimension is 1. Finally, the human behavior with the highest probability is determined as the recognition result of the behavior action.
[0012] Furthermore, the specific process of feeding the temporal spectrum of the secondary scattering component into the CNN network for automatic feature extraction in step S3.2 is as follows: Automatic feature extraction is performed by using a 2DCNN network consisting of N interleaved and stacked two-dimensional convolutional layers and N interleaved pooling layers. Then, a flattening layer is used to flatten the extracted two-dimensional features into a one-dimensional vector. The two-dimensional convolutional layers are responsible for feature extraction from the input image or input tensor, and the two-dimensional pooling layers complete the filtering and dimensionality reduction of the feature vectors obtained from the previous layer.
[0013] The present invention provides a multi-channel radar human behavior recognition system based on scattering separation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method.
[0014] The beneficial effects of this invention are as follows: This invention provides a multi-channel radar human behavior recognition method and system based on scattering separation, belonging to the field of behavior recognition technology. The method includes the following steps: First, the acquired multi-channel radar signals are stacked, and the main scattering component and secondary scattering component are separated using PCA. STFT is then performed on each component to obtain a time-spectrum image. Next, HOG features are extracted from the time-spectrum image of the main scattering component, and CNN features are extracted from the time-spectrum image of the secondary scattering component. The features are then stacked and applied to an attention mechanism, and a classifier is used to complete the classification and recognition. This method effectively avoids the masking effect of strong scattering echoes from the torso on weak scattering echoes from the limbs and head, thus improving the performance of human behavior recognition.
[0015] The multi-channel radar human behavior recognition method based on scattering separation provided by this invention separates the strong scattering echoes from the torso from the weak scattering echoes from the limbs and head through PCA, effectively avoiding the masking effect of strong scattering and significantly improving the recognition accuracy.
[0016] The multi-channel radar human behavior recognition method based on scattering separation provided by this invention uses manual feature extraction for simple strong scattering torso echoes in images, and automatic feature extraction for complex weak scattering echoes in images. It also combines an attention mechanism to focus on more important features, so that the motion information carried by the echo signal can be fully utilized and the recognition accuracy is high.
[0017] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0018] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a schematic diagram of the overall process of this method.
[0019] Figure 2 This is a schematic diagram of the echo signal separation process.
[0020] Figure 3 The spectrum diagrams are of the original signal and the two separated components after the "jump".
[0021] Figure 4 The spectrum diagram for "picking up" the original signal and the two separated components.
[0022] Figure 5 The spectrum diagrams are of the original signal and the two separated components after the "squatting" motion.
[0023] Figure 6 The spectrum diagrams of the original signal and the two separated components are shown.
[0024] Figure 7 The spectrum diagrams are of the original signal and the two separated components of the "swing arm".
[0025] Figure 8 The spectrum diagrams are of the original signal and the two separated components. Detailed Implementation
[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0027] Please see Figure 1-2 The multi-channel radar human behavior recognition method based on scattering separation provided in this embodiment of the invention includes the following steps: Step S1: Use radar to collect human behavior data. After preprocessing, stack the multi-channel data and perform PCA to separate the main scattering component and the secondary scattering component. These correspond to the strong scattering echo of the torso and the weak scattering echo of the limbs, head and other parts of the human body, respectively. The strong scattering echo of the torso is the main scattering component, and the echo of the weak scattering echo of the limbs, head and other parts of the human body is the secondary scattering component. S1.1 First, a multi-channel step-frequency continuous wave radar is used to collect data on six types of human behavioral movements: jumping, picking up, squatting, stepping, swinging arms, and walking. Its transmission start frequency is Frequency step The step frequency is The echo signal is represented as: , in, Here are the scattering coefficients for each part. The echo delay at a certain moment for each part. This represents the number of parts of the human body that emit light.
[0028] S1.2 is designed to suppress environmental clutter while preserving signals near zero frequency from subtle human body movements. The radar echo data from each channel were subjected to MTI mean cancellation to obtain... indivual dimensional matrix , Each row represents the step frequency echo signal of a single sweep frequency cycle, and each column represents the echo data vector of a single frequency in each cycle. S1.3 will convert the matrix of each channel Extract and arrange the vector data from the same column together, that is... Vectors at the same frequency point in each channel are stacked together to form a new matrix, and then transposed to obtain the matrix. ,matrix The columns represent the number of channels, and the rows represent the number of time periods of the echo data; S1.4 The matrix obtained in step S1.3 PCA is performed to obtain the matrix To achieve reconstruction of multi-channel echo signals, the matrix... The first two rows, i.e., the first two principal component components, are extracted to obtain the principal scattering components. and secondary scattering components This achieves the separation of trunk signals from signals from other parts of the body during human behavior, thus avoiding the masking effect of trunk signals on signals from other parts of the body.
[0029] S1.4.1 PCA transforms the matrix through orthogonal transformation. The original linearly correlated channel vectors are reorganized into a new matrix according to their contribution magnitude. The uncorrelated row vectors are used to separate uncorrelated components from the multi-channel data; PCA is then used to obtain the recombination matrix. The specific method is to find covariance matrix And find The eigenvalues are arranged in descending order, and the corresponding unit eigenvectors are... Arranged in rows to form a matrix Then the original signal is mapped to a new feature space. In this process, a new component matrix is obtained. That is, the recombination of the original signal, expressed by the formula: The resulting recombination matrix The row vectors are independent of each other and are the original matrix. The linear weighted rearrangement of row vectors, and their arrangement order is the order of their contribution to the observed signal.
[0030] S1.4.2 In practice, the correlation between the human torso and other branches such as limbs is extremely weak, and separation can be achieved through PCA. Furthermore, radar-acquired data still contains weak noise signals after MTI, and these noise signals are independent of the human echo. Since the human echo signal has high energy and the noise energy is low, the components in the principal components obtained by PCA that contribute little to the original signal are motion clutter components and should be filtered out. Therefore, the reconstruction matrix... The first two rows of main scattering components and secondary scattering components correspond to the signal echoes of the torso and other branch parts, respectively. By directly extracting these two vectors, the signals of the torso and other parts in human behavior are separated, so as to perform subsequent processing operations and avoid the masking effect of the torso signals on the signals of other parts.
[0031] Step S2: Convert the main scattering component and secondary scattering components Substitute the values into the STFT to obtain the time-frequency spectrum; perform STFT calculation on the main scattering component to obtain the time-frequency spectrum of the main scattering component; perform STFT calculation on the secondary scattering component to obtain the time-frequency spectrum of the secondary scattering component; Step S3: Use HOG and CNN networks to extract the feature information of the main scattering component and the secondary scattering component from the time spectrum obtained in step S2, and then stack them to form human behavior features; use HOG to extract the feature information of the main scattering component from the time spectrum of the main scattering component, and use CNN network to extract the feature information of the secondary scattering component from the time spectrum of the secondary scattering component.
[0032] S3.1 For the time spectrum of the relatively simple main scattering component of the image, spectral features are manually extracted using the HOG feature algorithm.
[0033] S3.1.1 performs preprocessing on the time spectrum of the main scattering component, including grayscale conversion and Gamma correction, and then calculates the gradient at each pixel according to the following formula: In the formula, Representing image pixels The horizontal gradient, vertical gradient, and pixel value are considered. The magnitude and direction of the pixel gradient are calculated using the following formula: S3.1.2 Divide the image into several containing The gradient of each pixel within a given unit is calculated to form a 9-channel histogram; this histogram is then slid across the pixel with a step size to obtain several histograms containing... The unit is divided into blocks, and the unit gradient histogram vector in each block is normalized using the L2 norm. The vectors are then concatenated to obtain the feature vector of the image.
[0034] S3.2 For the temporal spectrum of the relatively complex subscattering component of the image, an automatic feature extraction is performed by using a 2DCNN network consisting of three 2D convolutional layers and three 2D pooling layers stacked together. Then, a flattening layer is used to flatten the extracted 2D features into a 1D vector, which achieves the effect of effectively obtaining feature information while reducing redundant features. The 2D convolutional layer is responsible for feature extraction from the input image or input tensor, and the 2D pooling layer completes the filtering and dimensionality reduction of the feature vector obtained from the previous layer, reducing the computational load of the network while retaining key features.
[0035] S3.3 The two feature vectors obtained in steps S3.1 and S3.2 are directly concatenated and spliced to complete feature stacking.
[0036] Step S4: The human behavior features obtained in step S3 are input into the attention mechanism to focus on more important feature information, and then input into the classifier to complete the classification and recognition.
[0037] S4.1 divides the feature information data into m tokens as input, and then processes them through the network weights. generate , and vector, where Then to , Perform a dot product operation to calculate the similarity between the two, then divide by... Square root of matrix dimension This stabilizes the numerical gradient, then uses softmax to convert the similarity calculation results into a probability distribution of the weights, and finally applies the weights to... The calculation result of self-attention is obtained, and the expression for self-attention is as follows: in, , , .
[0038] S4.2 The result obtained in step S4.1 is input into a classification layer consisting of two fully connected layers, with ReLU and Softmax selected as activation functions respectively. The Softmax function outputs a result vector with the same dimension as the human behavior category. Each dimension represents the probability value of the input being identified as a certain behavior, and the sum of the probabilities of each dimension is 1. Finally, the human behavior with the highest probability is determined as the recognition result of the behavior action.
[0039] Figure 2 The diagram illustrates the echo signal separation process, showing the data matrix processing flow in steps S1.3 to S1.4. After mean cancellation, the multi-channel radar data is processed by channel stacking and PCA to obtain the separated main scattering component and secondary scattering component, which correspond to the strong scattering echo of the human body's torso and the weak scattering echo of the limbs, head and other parts, respectively.
[0040] Figure 3 The time spectrum diagrams of the original signal and the two separated components of the "jump" signal are shown. The time spectrum diagram of the signal collected by radar when the human body is performing the "jump" behavior is shown, as well as the time spectrum diagrams of the main scattering component and the secondary scattering component signal after a series of processing. It can be seen from the comparison that the echoes of the weak scattering parts of the human body such as the limbs and head are revealed after separation, avoiding the masking effect of the strong scattering echoes of the torso.
[0041] Figure 4The time spectrum diagrams of the original signal and the two separated components are shown. The time spectrum diagram of the signal collected by the radar when the human body is performing the "picking" behavior is shown, as well as the time spectrum diagrams of the main scattering component and the secondary scattering component signal after a series of processing. It can be seen from the comparison that the echoes of the weak scattering parts such as the human limbs and head are revealed after separation, avoiding the masking effect of the strong scattering torso echo.
[0042] Figure 5 The time-spectrum diagrams of the original signal and the two separated components of the "squatting" signal are shown. This illustrates the time-spectrum diagram of the signal collected by radar when the human body is performing the "squatting" action, as well as the time-spectrum diagrams of the main scattering component and the secondary scattering component signal after a series of processing steps. The comparison shows that the echoes from the weakly scattering parts of the human body, such as the limbs and head, are visible after separation, avoiding the masking effect of the strong scattering echoes from the torso.
[0043] Figure 6 The time-spectrum diagrams of the original "stepping" signal and the two separated components are shown. This illustrates the time-spectrum diagram of the signal collected by radar when the human body is performing the "stepping" action, as well as the time-spectrum diagrams of the main scattering component and the secondary scattering component signal after a series of processing steps. The comparison shows that the echoes from the weakly scattering parts of the human body, such as the limbs and head, are visible after separation, avoiding the masking effect of the strong scattering echoes from the torso.
[0044] Figure 7 The image shows the time spectrum of the original signal and the two separated components of the "arm swing" signal. It shows the time spectrum of the signal collected by radar when the human body is performing the "arm swing" action, as well as the time spectrum of the main scattering component and the secondary scattering component signal after a series of processing. The comparison shows that the echoes of the weak scattering parts of the human body such as the limbs and head are revealed after separation, avoiding the masking effect of the strong scattering echoes of the torso.
[0045] Figure 8 The time-spectrum diagrams of the original "walking" signal and the two separated components are shown. This illustrates the time-spectrum diagram of the signal collected by radar when the human body is performing the "walking" behavior, as well as the time-spectrum diagrams of the main scattering component and the secondary scattering component signal after a series of processing steps. The comparison shows that the echoes from the weakly scattering parts of the human body, such as the limbs and head, are visible after separation, avoiding the masking effect of the strong scattering echoes from the torso.
[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A multi-channel radar human behavior recognition method based on scattering separation, characterized in that: Includes the following steps: Step S1: Use radar to collect human behavior data, stack the multi-channel data and perform PCA to separate the main scattering component and the secondary scattering component. Step S2: Substitute the main scattering component and the secondary scattering component into the STFT to obtain the time spectrum; Step S3: Use HOG and CNN networks to extract the feature information of the main scattering component and the secondary scattering component from the time-frequency spectrogram, and then stack them to form human behavior features; Step S4: Apply an attention mechanism to human behavioral characteristics and input them into a classifier to complete classification and recognition; The acquisition and processing of multi-channel radar signals in step S1 is as follows: S1.1 First, a multi-channel step-frequency continuous wave radar is used to collect human behavior data and obtain echo signals. S1.2 The radar echo signals of each channel were subjected to MTI mean cancellation to obtain... indivual dimensional matrix ; S1.3 will Channel matrix Extract the vector data from the same column and arrange them together to form a new matrix. Transpose the new matrix to obtain the matrix. ,matrix The columns represent the number of channels, and the rows represent the number of time periods of the echo data; S1.4 The matrix obtained in step S1.3 After PCA, the first two principal component components were extracted. and The main scattering component and the secondary scattering component are obtained. The feature extraction process in step S3 is as follows: S3.1 Perform HOG artificial feature extraction on the time spectrum of the main scattering component; S3.2 The temporal spectrum of the secondary scattering component is fed into a CNN network for automatic feature extraction; S3.3 Stack the two feature vectors obtained in steps S3.1 and S3.2; The specific process of focusing human behavioral features on the obtained feature information through the attention mechanism and then classifying and recognizing it using a classifier in step S4 is as follows: S4.1 divides the feature information data into m tokens as input, and then processes them through the network weights. generate , and vector, where Then to , Perform a dot product operation to calculate the similarity between the two, then divide by... Square root of matrix dimension This stabilizes the numerical gradient, then uses softmax to convert the similarity calculation results into a probability distribution of the weights, and finally applies the weights to... The calculation result of self-attention is obtained, and the expression for self-attention is as follows: in, , , ; S4.2 The result obtained in step S4.1 is input into the classification layer consisting of two fully connected layers. ReLU and Softmax are selected as activation functions respectively. The Softmax function outputs a result vector with the same dimension as the human behavior category. Each dimension represents the probability value of the input being identified as a certain behavior. The sum of the probabilities of each dimension is 1. Finally, the human behavior with the highest probability is determined as the recognition result of the behavior action.
2. The multi-channel radar human behavior recognition method based on scattering separation according to claim 1, characterized in that, The radar echo signal mentioned in step S1.1 is: , in, Here are the scattering coefficients for each part. For the echo delay at a certain moment, This refers to the number of parts of the human body that emit light. The starting frequency for transmission. For frequency stepping, The step frequency is the number of points.
3. The multi-channel radar human behavior recognition method based on scattering separation according to claim 1, characterized in that, The human behavior data collection in step S1 includes the collection of data on human behavior actions such as jumping, picking up, squatting, stepping, swinging arms, and walking.
4. The multi-channel radar human behavior recognition method based on scattering separation according to claim 1, characterized in that, In step S1.4, the first two principal component components are extracted using PCA. and The specific process for obtaining the main scattering component and the secondary scattering component is as follows: PCA transforms the matrix through orthogonal transformation. The original linearly correlated channel vectors are reorganized into a new matrix according to their contribution magnitude. In the uncorrelated row vectors, uncorrelated components are separated from the multi-channel data; The recombination matrix is obtained using PCA. The specific method is to find covariance matrix And find The eigenvalues are arranged in descending order, and the corresponding unit eigenvectors are... Arranged in rows to form a matrix Then the original signal is mapped to a new feature space. In this process, a new component matrix is obtained. That is, the recombination of the original signal, expressed by the formula: The resulting recombination matrix The row vectors are independent of each other and are the original matrix. The linear weighted rearrangement of row vectors, and their arrangement order is the order of their contribution to the observed signal.
5. The multi-channel radar human behavior recognition method based on scattering separation according to claim 1, characterized in that, The specific process of feeding the temporal spectrum of the secondary scattering component into the CNN network for automatic feature extraction in step S3.2 is as follows: Automatic feature extraction is performed by using a 2DCNN network consisting of N interleaved and stacked two-dimensional convolutional layers and N interleaved pooling layers. Then, a flattening layer is used to flatten the extracted two-dimensional features into a one-dimensional vector. The two-dimensional convolutional layers are responsible for feature extraction from the input image or input tensor, and the two-dimensional pooling layers complete the filtering and dimensionality reduction of the feature vectors obtained from the previous layer.
6. A multi-channel radar human behavior recognition system based on scattering separation, comprising a memory, a processor, and a computer program stored in 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 5.
