A millimeter-wave radar fall detection method based on AI deep learning

Abstract

This invention discloses a fall detection method for millimeter-wave radar based on AI deep learning, specifically including: Step 1: Acquiring millimeter-wave radar echo signals and performing distance, velocity, and angle analysis to generate a motion feature sequence; Step 2: Dividing the motion feature sequence into time windows to generate a time window feature vector set; Step 3: Inputting the time window feature vector set into a deep learning behavior recognition model to generate a behavior probability sequence; Step 4: Performing hierarchical probability bridging on the behavior probability sequence to generate a behavior probability distribution sequence; Step 5: Performing non-uniform diffusion adjustment on the behavior probability distribution sequence to generate a continuous probability flow sequence; Step 6: Performing behavior recognition on the continuous probability flow sequence to generate a fall probability sequence; Step 7: Performing threshold comparison on the fall probability sequence to generate a fall detection result. This invention performs continuous recognition calculation on millimeter-wave radar human behavior sequences, achieving stable determination and dynamic detection of fall behavior.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar detection technology, and in particular to a millimeter-wave radar fall detection method based on AI deep learning. Background Technology

[0002] As the population ages and the proportion of elderly people living alone increases, falls have become a common and high-risk emergency in home safety monitoring. If falls are not detected promptly, they can easily cause serious personal injury. Therefore, automatic fall detection technology has become an important research direction in the fields of intelligent monitoring, smart elderly care, and smart home monitoring. By constructing a human behavior recognition system to continuously monitor human movement and provide timely detection results when a fall occurs, the safety risks associated with falls can be reduced to some extent.

[0003] Existing fall detection technologies mainly include methods based on video image recognition and methods based on wearable devices. Video image recognition methods typically acquire sequences of human images using cameras, then use image processing or deep learning algorithms to identify changes in posture to determine if a fall has occurred. These methods can obtain relatively clear images in stable, unobstructed lighting conditions, but are prone to instability in environments with significant lighting variations. Furthermore, video capture devices raise privacy concerns, limiting their application in homes and other privacy-sensitive areas. Wearable device-based methods typically collect human motion data using accelerometers or gyroscopes, and then analyze this data to determine fall behavior. These methods require continuous device wear, and their effectiveness is susceptible to changes in wear position and user compliance.

[0004] Millimeter-wave radar sensing technology, with its advantages of non-contact sensing, insensitivity to lighting conditions, and ability to acquire information on human movement distance, speed, and spatial location, is increasingly being applied in the field of human behavior detection. By acquiring human motion echo signals using millimeter-wave radar and extracting human motion features, human behavior states can be identified. However, in the process of millimeter-wave radar human behavior recognition, the human motion state changes continuously over time, and the behavior probability distribution exhibits obvious dynamic evolution characteristics. How to stably model the process of behavior probability change to achieve accurate identification of falls has become a problem that needs to be solved in millimeter-wave radar human behavior detection technology. Summary of the Invention

[0005] One objective of this invention is to propose a fall detection method for millimeter-wave radar based on AI deep learning. This invention introduces a hierarchical probability bridging model of human motion and a non-uniform diffusion bridging adjustment structure with continuous probability flow constraints. It performs hierarchical bridging calculation and diffusion adjustment operation on the probability sequence of human behavior of millimeter-wave radar to construct a continuously evolving human behavior probability flow recognition process. It has the advantages of strong ability to analyze behavior changes, high accuracy of fall recognition, and stability of the detection process.

[0006] A fall detection method based on AI deep learning using millimeter-wave radar according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain the sequence of human motion echo signals from millimeter-wave radar, and perform range analysis, velocity analysis, and angle analysis on the echo signals to generate a radar motion feature sequence;

[0008] Step 2: Perform time window division and feature combination operations on the radar motion feature sequence to generate a set of time window motion feature vectors;

[0009] Step 3: Establish a deep learning behavior recognition model. Input the set of motion feature vectors within the time window into the behavior recognition model to generate a sequence of behavior state probability vectors.

[0010] Step 4: Construct a hierarchical probabilistic bridging model for human motion. Establish a probability layer for posture change, a probability layer for velocity change, and a probability layer for spatial displacement in the sequence of behavioral state probability vectors. Perform probability bridging operations on each probability layer to generate a behavioral probability bridging matrix, and perform cross-layer mapping to generate a sequence of behavioral probability distributions.

[0011] Step 5: Construct a non-uniform diffusion bridging adjustment structure with continuous probability flow constraints, calculate the probability flow change index for the behavior probability distribution sequence, configure the time non-uniform diffusion coefficient, and perform diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence;

[0012] Step 6: Perform behavior state recognition operation on the continuous probability stream sequence to generate a fall behavior probability sequence;

[0013] Step 7: Perform a threshold comparison operation on the fall behavior probability sequence to generate fall detection results.

[0014] Optionally, step 1 specifically includes:

[0015] Start the millimeter-wave radar data acquisition module to perform continuous scanning and acquisition operations on the monitoring area, record the radar echo signals corresponding to each scanning cycle, and establish a radar echo signal sequence according to the scanning time order.

[0016] Read the radar echo signal sequence, perform a range-dimensional fast Fourier transform operation on the radar echo signal sequence to generate a range spectrum matrix, and extract the human target range component sequence in the range spectrum matrix according to the range index position;

[0017] Perform a velocity-dimensional fast Fourier transform operation on the distance spectrum matrix to generate a velocity spectrum matrix, and extract the human motion velocity component sequence from the velocity spectrum matrix according to the velocity index position;

[0018] Angle analysis operation is performed on the radar echo signal sequence, the angle spectrum matrix is ​​calculated through the phase difference of the array channels, and the spatial angle component sequence of the human target is extracted from the angle spectrum matrix.

[0019] Read the human target distance component sequence, human motion velocity component sequence, and human spatial angle component sequence, perform vector combination operation according to the time index position, and generate human motion feature vector at each time index position;

[0020] The motion feature vectors of each human body are collected in chronological order to establish a radar motion feature sequence.

[0021] Optionally, step 2 includes the following steps:

[0022] Read the radar motion feature sequence, perform time window division operation on the radar motion feature sequence according to a unified time index order, generate a set of time windows, and extract human motion feature vectors according to time index position within each time window;

[0023] Within each time window, read the human motion feature vector, perform position difference calculation on the position component of the human motion feature vector to generate a position change component sequence, perform velocity difference calculation on the velocity component to generate a velocity change component sequence, and perform trajectory direction change calculation on the trajectory component to generate a trajectory change component sequence.

[0024] Read the position change component sequence, velocity change component sequence, and trajectory change component sequence, and perform component combination operation according to the time index position to generate a time window motion change vector sequence;

[0025] Within each time window, statistical feature calculations are performed on the motion change vector sequence of the time window to generate a statistical feature vector. The statistical feature vector contains statistical components of position change, velocity change, and trajectory change.

[0026] Read the statistical feature vectors, perform feature vector concatenation operations on the statistical feature vectors, and generate a set of motion feature vectors for each time window.

[0027] Optionally, step 3 specifically includes:

[0028] Read the set of motion feature vectors in the time window, perform a sequence permutation operation on the motion feature vectors in the time window according to the time index order to generate a motion feature sequence in the time window, perform vector normalization operation on the motion feature sequence in the time window, and perform numerical standardization processing on each component of the motion feature vector at each time index position to generate a normalized feature sequence.

[0029] The normalized feature sequence is input into the deep learning behavior recognition model. The values ​​of each motion feature vector are written in the time index order in the model input layer to generate the model input feature matrix. In the deep learning behavior recognition model, feature mapping operation is performed on the model input feature matrix to generate a sequence of behavior representation vectors.

[0030] Perform probability mapping operations on the behavior representation vector sequence, perform probability normalization calculations on the behavior feature components at each time index position to generate behavior state probability vectors, and aggregate the behavior state probability vectors in time index order to generate a behavior state probability vector sequence.

[0031] Optionally, the generation of the behavior probability bridging matrix in step 4 specifically includes:

[0032] Read the sequence of behavior state probability vectors, extract the values ​​of each probability component in the behavior state probability vector at each time index position, perform a hierarchical partitioning operation according to the probability component category to generate an attitude change probability layer, a velocity change probability layer and a spatial displacement probability layer, record the attitude change probability component sequence in the attitude change probability layer, record the velocity change probability component sequence in the velocity change probability layer, and record the spatial displacement probability component sequence in the spatial displacement probability layer.

[0033] In the attitude change probability layer, attitude change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the attitude change probability components to generate attitude probability bridging components. In the velocity change probability layer, velocity change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the velocity change probability components to generate velocity probability bridging components. In the spatial displacement probability layer, spatial displacement probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the spatial displacement probability components to generate displacement probability bridging components.

[0034] The attitude probability bridging components are subjected to a sequence permutation operation to generate an attitude probability bridging sequence in time index order; the velocity probability bridging components are subjected to a sequence permutation operation to generate a velocity probability bridging sequence; and the displacement probability bridging components are subjected to a sequence permutation operation to generate a displacement probability bridging sequence.

[0035] Read the attitude probability bridging sequence, velocity probability bridging sequence and displacement probability bridging sequence, perform matrix writing operation on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at a unified time index position, and write the values ​​of the three types of probability bridging components at the corresponding row and column positions of the matrix to generate the behavior probability bridging matrix.

[0036] Record the attitude bridging probability value, velocity bridging probability value and displacement bridging probability value corresponding to each time index position in the behavior probability bridging matrix, and establish a sequence of behavior probability bridging matrices according to the time index order.

[0037] Optionally, the generation of the behavioral probability distribution sequence in step 4 includes the following steps:

[0038] Read the behavior probability bridging matrix, extract the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at the unified time index position, and perform cross-layer mapping operation on the three types of probability bridging components to generate cross-layer probability mapping vector.

[0039] In the process of generating cross-layer probability mapping vectors, component alignment operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a probability component alignment sequence. In the probability component alignment sequence, weight combination operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a unified behavior probability component sequence.

[0040] Generate a behavior probability distribution sequence by assembling the unified behavior probability component sequences in time index order, and record the set of behavior probability values ​​corresponding to each time index position in the behavior probability distribution sequence.

[0041] Optionally, step 5, which involves calculating the probability flow change index of the behavioral probability distribution sequence and configuring the time non-uniform diffusion coefficient, specifically includes:

[0042] Read the behavior probability distribution sequence, extract the behavior probability distribution vector at a unified time index position, perform probability difference calculation on the behavior probability distribution vector at adjacent time index positions to generate a probability change component sequence, and record the value of each probability component in the probability change component sequence;

[0043] Perform absolute value calculation on the probability change component sequence to generate probability change amplitude component sequence, perform component accumulation operation on the probability change amplitude component sequence to generate probability change amplitude value corresponding to time index position, and record probability change amplitude value sequence at each time index position;

[0044] Read the probability change magnitude numerical sequence, perform sequence normalization operation on the probability change magnitude numerical sequence to generate a probability flow change index sequence, and record the probability flow change index value corresponding to each time index position in the probability flow change index sequence.

[0045] Read the probability flow change index sequence, perform diffusion coefficient mapping operation on the probability flow change index values ​​at a unified time index position to generate a time non-uniform diffusion coefficient sequence, and record the diffusion coefficient values ​​corresponding to each time index position in the time non-uniform diffusion coefficient sequence.

[0046] Optionally, step 5, performing diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence, includes the following steps:

[0047] Read the behavior probability bridging matrix sequence, extract the attitude bridging probability component, velocity bridging probability component and displacement bridging probability component at a unified time index position, and establish a probability bridging component vector sequence in time index order;

[0048] Read the time-uniform diffusion coefficient sequence, write the diffusion coefficient values ​​at each time index position into the probability bridging component vector sequence, and perform a diffusion weight mapping operation on the probability bridging component vector sequence to generate a diffusion weight vector sequence.

[0049] In the diffusion weight vector sequence, the attitude bridging probability component, velocity bridging probability component, and displacement bridging probability component are multiplied by diffusion weights to generate a diffusion probability component sequence.

[0050] The diffusion probability component sequences are collected in time index order, and a probability normalization operation is performed on the diffusion probability components at each time index position to generate a continuous probability flow vector.

[0051] A continuous probability stream sequence is generated by assembling continuous probability stream vectors in time index order.

[0052] Optionally, step 6, which involves performing behavior state recognition operations to generate a fall behavior probability sequence, specifically includes:

[0053] Read the continuous probability flow sequence, extract the continuous probability flow vector at a unified time index position, and extract the attitude probability component, velocity probability component and displacement probability component from the continuous probability flow vector to generate a set of behavior probability components.

[0054] In the set of behavior probability components, attitude change calculation operations are performed on attitude probability components to generate attitude change component sequences, velocity change calculation operations are performed on velocity probability components to generate velocity change component sequences, displacement change calculation operations are performed on displacement probability components to generate displacement change component sequences, and at a unified time index position, probability combination operations are performed on attitude change components, velocity change components and displacement change components to generate behavior state probability vectors.

[0055] Extract the fall behavior probability component from the behavior state probability vector, record the fall behavior probability value at each time index position, and collect the fall behavior probability values ​​in time index order to generate a fall behavior probability sequence.

[0056] The beneficial effects of this invention are:

[0057] (1) This invention acquires human motion echo signal sequences through millimeter-wave radar, performs distance analysis, velocity analysis and angle analysis operations on the echo signals to generate radar motion feature sequences, and then constructs human motion features through a time window feature vector construction process to express human motion features in a structured manner, so that changes in human position, speed and spatial trajectory are described in a unified time index structure, thereby forming a motion feature data structure with time continuity, providing a stable data input basis for subsequent behavior state probability calculation.

[0058] (2) The present invention constructs a hierarchical probability bridging model of human motion, establishes a posture change probability layer, a velocity change probability layer and a spatial displacement probability layer in the sequence of behavioral state probability vectors, generates a behavioral probability bridging matrix through probability bridging operation, and generates a behavioral probability distribution sequence through cross-layer mapping, so that the multidimensional motion features of the human body are fused and expressed in a unified probability structure, thereby forming a behavioral probability description method with structural correlation.

[0059] (3) The present invention constructs a non-uniform diffusion bridging adjustment structure with continuous probability flow constraint. By calculating the probability flow change index of the behavior probability distribution sequence, a time non-uniform diffusion coefficient is generated. A diffusion adjustment operation is performed on the behavior probability bridging matrix to generate a continuous probability flow sequence. Then, a fall behavior probability sequence is generated through behavior state recognition operation, so that the human behavior probability change process forms a continuous evolution structure in the time series, thereby forming a stable human fall detection process. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a millimeter-wave radar fall detection method based on AI deep learning proposed in this invention.

[0062] Figure 2 This is a schematic diagram illustrating the implementation process of the human motion hierarchical probability bridging model for a millimeter-wave radar fall detection method based on AI deep learning proposed in this invention.

[0063] Figure 3This is a schematic diagram of the non-uniform diffusion bridging adjustment structure of the millimeter-wave radar fall detection method based on AI deep learning proposed in this invention, which is constrained by continuous probability flow. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-3 A fall detection method based on millimeter-wave radar using AI deep learning includes the following steps:

[0066] Step 1: Obtain the sequence of human motion echo signals from millimeter-wave radar, and perform range analysis, velocity analysis, and angle analysis on the echo signals to generate a radar motion feature sequence;

[0067] Step 2: Perform time window division and feature combination operations on the radar motion feature sequence to generate a set of time window motion feature vectors;

[0068] Step 3: Establish a deep learning behavior recognition model. Input the set of motion feature vectors within the time window into the behavior recognition model to generate a sequence of behavior state probability vectors.

[0069] Step 4: Construct a hierarchical probabilistic bridging model for human motion. Establish a probability layer for posture change, a probability layer for velocity change, and a probability layer for spatial displacement in the sequence of behavioral state probability vectors. Perform probability bridging operations on each probability layer to generate a behavioral probability bridging matrix, and perform cross-layer mapping to generate a sequence of behavioral probability distributions.

[0070] Step 5: Construct a non-uniform diffusion bridging adjustment structure with continuous probability flow constraints, calculate the probability flow change index for the behavior probability distribution sequence, configure the time non-uniform diffusion coefficient, and perform diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence.

[0071] Step 6: Perform behavior state recognition operation on the continuous probability stream sequence to generate a fall behavior probability sequence;

[0072] Step 7: Perform a threshold comparison operation on the fall behavior probability sequence to generate fall detection results.

[0073] In this embodiment, step 1 specifically includes:

[0074] Start the millimeter-wave radar data acquisition module to perform continuous scanning acquisition operations on the monitoring area, record the radar echo signals corresponding to each scanning cycle, and establish a radar echo signal sequence according to the scanning time order.

[0075] Read the radar echo signal sequence, perform a range-dimensional fast Fourier transform operation on the radar echo signal sequence to generate a range spectrum matrix, and extract the human target range component sequence in the range spectrum matrix according to the range index position;

[0076] Perform a velocity-dimensional fast Fourier transform operation on the distance spectrum matrix to generate a velocity spectrum matrix, and extract the human motion velocity component sequence from the velocity spectrum matrix according to the velocity index position;

[0077] Angle analysis operation is performed on the radar echo signal sequence, the angle spectrum matrix is ​​calculated through the phase difference of the array channels, and the spatial angle component sequence of the human target is extracted from the angle spectrum matrix.

[0078] Read the human target distance component sequence, human motion velocity component sequence, and human spatial angle component sequence, perform vector combination operation according to the time index position, and generate human motion feature vector at each time index position;

[0079] The motion feature vectors of each human body are collected in chronological order to establish a radar motion feature sequence.

[0080] In this embodiment, step 1, performing the distance-dimensional fast Fourier transform operation, specifically includes:

[0081] The radar echo signal sequence is segmented, with each time period corresponding to one echo signal sequence. Within each signal segment, a fast Fourier transform is performed to calculate each frequency component in the frequency domain.

[0082] Based on the amplitude and phase information of each frequency component in the Fourier transform result, a distance spectrum matrix is ​​generated. Each row of the distance spectrum matrix represents a frequency component, each column represents the signal change at different time points, and the element values ​​in the matrix represent the echo intensity at different frequencies.

[0083] After the range spectrum matrix is ​​generated, the range component sequence of the human target is extracted according to the range index position. By analyzing each frequency component in the frequency domain and combining the physical characteristics of radar signal propagation, the actual range value corresponding to each frequency component can be determined.

[0084] The distance component sequences related to the human target are extracted to form the final distance component data sequence.

[0085] In this embodiment, step 1, which involves performing a speed-dimensional Fast Fourier Transform, specifically includes:

[0086] The radar echo signal sequence is read, and these signals contain echo information at different time points. In order to perform velocity analysis on these echo signals, the signals need to be processed in chronological order, and velocity information needs to be extracted within each time period of the signal.

[0087] The radar echo signal sequence is segmented in the time dimension, with each time segment corresponding to a sub-sequence. For the sub-signal sequence within each time segment, a fast Fourier transform operation is performed to convert the signal from the time domain to the frequency domain. The operation maps the signal in the time domain to the frequency domain by calculating the amplitude and phase of each frequency component, thereby revealing the velocity information contained in the echo signal.

[0088] In the velocity spectrum matrix, each column represents the signal intensity distribution at a point in time, while each row represents a specific frequency component.

[0089] The velocity components in the velocity spectrum matrix are extracted, and these velocity components reflect the movement speed of the human target.

[0090] In this embodiment, step 2 includes the following steps:

[0091] Read the radar motion feature sequence, perform time window division operation on the radar motion feature sequence according to a unified time index order, generate a set of time windows, and extract human motion feature vectors according to time index position within each time window;

[0092] Within each time window, read the human motion feature vector, perform position difference calculation on the position component of the human motion feature vector to generate a position change component sequence, perform velocity difference calculation on the velocity component to generate a velocity change component sequence, and perform trajectory direction change calculation on the trajectory component to generate a trajectory change component sequence.

[0093] Read the position change component sequence, velocity change component sequence, and trajectory change component sequence, and perform component combination operation according to the time index position to generate a time window motion change vector sequence;

[0094] Within each time window, statistical feature calculations are performed on the motion change vector sequence of the time window to generate a statistical feature vector. The statistical feature vector contains statistical components of position change, velocity change, and trajectory change.

[0095] Read the statistical feature vectors, perform feature vector concatenation operations on the statistical feature vectors, and generate a set of motion feature vectors for each time window.

[0096] In this embodiment, the calculation of the trajectory direction change in step 2 specifically includes:

[0097] Read the sequence of human spatial position vectors arranged in time index order within the same time window, extract the spatial position vectors of the previous time position and the spatial position vectors of the next time position at adjacent time index positions, and perform displacement vector calculation operations on the two position vectors to generate a motion displacement vector sequence.

[0098] The first displacement vector and the second displacement vector are read at three consecutive time index positions. The direction angle calculation operation is performed on the first displacement vector and the second displacement vector. The direction change value corresponding to each time index position is calculated according to the vector dot product and the vector magnitude. The first displacement vector represents the direction of movement and displacement change of the human body between the first time position and the second time position. The second displacement vector represents the direction of movement and displacement change of the human body between the second time position and the third time position.

[0099] The calculated direction change values ​​are written into the direction change record sequence, and the direction change values ​​are collected in time index order within a unified time window to form a trajectory direction change value sequence.

[0100] A vector writing operation is performed on the numerical sequence of trajectory direction changes, generating corresponding trajectory change component values ​​at each time index position. The numerical values ​​of each trajectory change component are then collected in time index order to form a trajectory change component sequence.

[0101] In this embodiment, step 2 specifically includes the following steps: reading the position change component sequence, velocity change component sequence and trajectory change component sequence, establishing a time index position set within a unified time window, and extracting the corresponding position change component value, velocity change component value and trajectory change component value at each time index position.

[0102] Establish the component writing order at the current time index position, write the position change component value to the first position of the vector, write the velocity change component value to the second position of the vector, and write the trajectory change component value to the third position of the vector to form the motion change vector corresponding to the current time index position.

[0103] The component write operation is repeatedly performed at each time index position within the time window to generate a corresponding motion change vector at each time index position;

[0104] Perform a sequence arrangement operation on each motion change vector according to the time index order, and write the motion change vector record set sequentially within the time window to form a time window motion change vector sequence.

[0105] In this embodiment, the statistical feature vector in step 2 specifically includes:

[0106] Read the motion change vector sequence within each time window, and extract the position change component sequence, velocity change component sequence, and trajectory change component sequence from each motion change vector.

[0107] The location change component sequence is subjected to statistical calculation operations to calculate the average, maximum, minimum and variance values ​​of the sequence within the time window, forming a set of location change statistical components;

[0108] Perform statistical calculations on the velocity change component sequence to calculate the mean, maximum, minimum and variance values ​​of the velocity change sequence, forming a set of velocity change statistical components;

[0109] Perform statistical calculations on the trajectory change component sequence to calculate the mean, maximum, minimum and variance values ​​of the trajectory change sequence, forming a set of trajectory change statistical components;

[0110] The statistical feature vector is generated by writing the set of position change statistical components, the set of velocity change statistical components, and the set of trajectory change statistical components into the same feature vector in vector order.

[0111] In this embodiment, step 3 specifically includes:

[0112] Read the set of motion feature vectors in the time window, perform a sequence permutation operation on the motion feature vectors in the time window according to the time index order to generate a motion feature sequence in the time window, perform vector normalization operation on the motion feature sequence in the time window, and perform numerical standardization processing on each component of the motion feature vector at each time index position to generate a normalized feature sequence.

[0113] The normalized feature sequence is input into the deep learning behavior recognition model. The values ​​of each motion feature vector are written in the time index order in the model input layer to generate the model input feature matrix. In the deep learning behavior recognition model, feature mapping operation is performed on the model input feature matrix to generate a sequence of behavior representation vectors.

[0114] Perform probability mapping operations on the behavior representation vector sequence, perform probability normalization calculations on the behavior feature components at each time index position to generate behavior state probability vectors, and aggregate the behavior state probability vectors in time index order to generate a behavior state probability vector sequence.

[0115] In this embodiment, the deep learning behavior recognition model in step 3 specifically includes:

[0116] The model structure includes a feature input layer, a temporal feature extraction layer, a temporal state modeling layer, and a probability mapping layer; the feature input layer is used to receive motion feature vectors arranged in time index order to form the model input feature matrix;

[0117] The temporal feature extraction layer uses a one-dimensional convolutional operation structure to perform feature mapping calculation on the model input feature matrix, extracting human motion change features in the time dimension and generating a temporal feature representation matrix;

[0118] The temporal state modeling layer uses a long short-term memory network structure to perform sequence state update calculations on the temporal feature representation matrix, generating a sequence of behavioral state hidden vectors at each time index position;

[0119] The probability mapping layer performs fully connected mapping operations and probability normalization calculations on the behavior state hidden vector, generating behavior state probability vectors at each time index position;

[0120] During model operation, the normalized feature sequence is written into the model input feature matrix in time index order. Then, feature calculation is performed sequentially through the time-series feature extraction layer, the time-series state modeling layer, and the probability mapping layer. At each time index position, a sequence of behavioral state probability vectors is generated to represent the probability distribution of human behavior in the time series.

[0121] In this embodiment, the execution probability mapping operation in step 3 specifically includes:

[0122] Read the sequence of behavior representation vectors, extract the values ​​of each feature component in the behavior representation vector at a unified time index position, and establish a sequence of probability component positions in the behavior representation vector;

[0123] Perform a linear mapping calculation on the behavior representation vector at the current time index position, perform a weighted product calculation on each component position in the probability component position sequence to generate a behavior score component sequence, and record the corresponding score value at each component position;

[0124] Perform an exponential mapping operation on the behavior score component sequence to generate an exponential component sequence, and perform a component accumulation operation on the exponential component sequence to generate a normalized denominator value;

[0125] Normalize and divide the exponential component values ​​at each component position to generate behavioral probability component values, and generate a behavioral state probability vector at the current time index position.

[0126] The probability vectors of each behavior state are collected in time index order to form a sequence of behavior state probability vectors.

[0127] In this embodiment, step 4, generating the behavior probability bridging matrix, specifically includes:

[0128] Read the sequence of behavior state probability vectors, extract the values ​​of each probability component in the behavior state probability vector at each time index position, perform a hierarchical partitioning operation according to the probability component category to generate an attitude change probability layer, a velocity change probability layer and a spatial displacement probability layer, record the attitude change probability component sequence in the attitude change probability layer, record the velocity change probability component sequence in the velocity change probability layer, and record the spatial displacement probability component sequence in the spatial displacement probability layer.

[0129] In the attitude change probability layer, attitude change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the attitude change probability components to generate attitude probability bridging components. In the velocity change probability layer, velocity change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the velocity change probability components to generate velocity probability bridging components. In the spatial displacement probability layer, spatial displacement probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the spatial displacement probability components to generate displacement probability bridging components.

[0130] The attitude probability bridging components are subjected to a sequence permutation operation to generate an attitude probability bridging sequence in time index order; the velocity probability bridging components are subjected to a sequence permutation operation to generate a velocity probability bridging sequence; and the displacement probability bridging components are subjected to a sequence permutation operation to generate a displacement probability bridging sequence.

[0131] Read the attitude probability bridging sequence, velocity probability bridging sequence and displacement probability bridging sequence, perform matrix writing operation on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at a unified time index position, and write the values ​​of the three types of probability bridging components at the corresponding row and column positions of the matrix to generate the behavior probability bridging matrix.

[0132] Record the attitude bridging probability value, velocity bridging probability value and displacement bridging probability value corresponding to each time index position in the behavior probability bridging matrix, and establish a sequence of behavior probability bridging matrices according to the time index order.

[0133] In this embodiment, step 4, the execution probability join operation, specifically includes:

[0134] Read the attitude change probability components at the current time index position and the adjacent time index position. The attitude change probability components at these two time positions will be used to calculate the connection relationship between them.

[0135] Calculate the probabilistic connection weights for these two time index positions;

[0136] Based on the calculated weights, the attitude change probability component at the current time index position is weighted and combined with the attitude change probability components at adjacent time index positions to generate a new attitude probability bridging component.

[0137] In this embodiment, step 4, the execution sequence permutation operation, specifically includes:

[0138] Read the sequence of attitude probability bridging components in the behavior probability bridging matrix, extract the attitude probability bridging components of each time position in time index order, and the attitude probability bridging components of each time position represent the probability relationship between the current time position and the adjacent time positions.

[0139] Arrange these attitude probability bridging components in chronological order to ensure that all attitude probability bridging components are arranged sequentially in the sequence from the beginning to the end of the time series.

[0140] The arranged attitude probability bridging components are written into the attitude probability bridging sequence in chronological order to generate a complete attitude probability bridging sequence.

[0141] In this embodiment, step 4, generating the probability distribution sequence of behaviors, includes the following steps:

[0142] Read the behavior probability bridging matrix, extract the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at the unified time index position, and perform cross-layer mapping operation on the three types of probability bridging components to generate cross-layer probability mapping vector.

[0143] In the process of generating cross-layer probability mapping vectors, component alignment operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a probability component alignment sequence. In the probability component alignment sequence, weight combination operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a unified behavior probability component sequence.

[0144] Generate a behavior probability distribution sequence by assembling the unified behavior probability component sequences in time index order, and record the set of behavior probability values ​​corresponding to each time index position in the behavior probability distribution sequence.

[0145] In this embodiment, step 4, performing the component alignment operation, specifically includes:

[0146] Read the posture probability bridging component, velocity probability bridging component and displacement probability bridging component. These components correspond to the changes in human posture, velocity and spatial displacement, respectively. They may be generated from different probability layers and represent the changes in human motion in different aspects.

[0147] When performing component alignment operations, it is necessary to ensure that all components at each time index position can be correctly matched;

[0148] The three probability components (attitude, velocity, and displacement) at each time point are combined into a unified alignment sequence. During the alignment process, if the values ​​of some components are missing or incomplete, interpolation or padding can be used to fill in the missing values ​​to ensure that all components have corresponding values ​​at each time point.

[0149] In this embodiment, step 5, calculating the probability flow change index for the behavioral probability distribution sequence and configuring the time non-uniform diffusion coefficient, specifically includes:

[0150] Read the behavior probability distribution sequence, extract the behavior probability distribution vector at a unified time index position, perform probability difference calculation on the behavior probability distribution vector at adjacent time index positions to generate a probability change component sequence, record the value of each probability component in the probability change component sequence, and the probability difference is the difference between the probability component values ​​at adjacent time positions.

[0151] Perform absolute value calculation on the probability change component sequence to generate probability change amplitude component sequence, perform component accumulation operation on the probability change amplitude component sequence to generate probability change amplitude value corresponding to time index position, and record probability change amplitude value sequence at each time index position;

[0152] Read the probability change magnitude numerical sequence, perform sequence normalization on the probability change magnitude numerical sequence to generate a probability flow change index sequence, and record the probability flow change index value corresponding to each time index position in the probability flow change index sequence; sequence normalization is to convert the probability change magnitude numerical sequence into a standardized numerical range, so that all change magnitude numerical sequence is compared within the same standard range, thereby avoiding the impact of numerical scale differences on subsequent calculations.

[0153] Read the probability flow change index sequence, perform diffusion coefficient mapping operation on the probability flow change index values ​​at a unified time index position to generate a time non-uniform diffusion coefficient sequence, and record the diffusion coefficient values ​​corresponding to each time index position in the time non-uniform diffusion coefficient sequence.

[0154] In this embodiment, step 5, performing diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence, includes the following steps:

[0155] Read the behavior probability bridging matrix sequence, extract the attitude bridging probability component, velocity bridging probability component and displacement bridging probability component at a unified time index position, and establish a probability bridging component vector sequence in time index order;

[0156] Read the time-uniform diffusion coefficient sequence, write the diffusion coefficient values ​​at each time index position into the probability bridging component vector sequence, and perform a diffusion weight mapping operation on the probability bridging component vector sequence to generate a diffusion weight vector sequence.

[0157] In the diffusion weight vector sequence, the attitude bridging probability component, velocity bridging probability component, and displacement bridging probability component are multiplied by diffusion weights to generate a diffusion probability component sequence.

[0158] The diffusion probability component sequences are collected in time index order, and a probability normalization operation is performed on the diffusion probability components at each time index position to generate a continuous probability flow vector.

[0159] A continuous probability stream sequence is generated by assembling continuous probability stream vectors in time index order.

[0160] In this embodiment, step 6, the execution behavior state recognition operation to generate the fall behavior probability sequence, specifically includes:

[0161] Read the continuous probability flow sequence, extract the continuous probability flow vector at a unified time index position, and extract the attitude probability component, velocity probability component and displacement probability component from the continuous probability flow vector to generate a set of behavior probability components.

[0162] In the set of behavior probability components, attitude change calculation operations are performed on attitude probability components to generate attitude change component sequences, velocity change calculation operations are performed on velocity probability components to generate velocity change component sequences, displacement change calculation operations are performed on displacement probability components to generate displacement change component sequences, and at a unified time index position, probability combination operations are performed on attitude change components, velocity change components and displacement change components to generate behavior state probability vectors.

[0163] Extract the fall behavior probability component from the behavior state probability vector, record the fall behavior probability value at each time index position, and collect the fall behavior probability values ​​in time index order to generate a fall behavior probability sequence.

[0164] In this embodiment, step 6, performing the attitude change calculation, specifically includes:

[0165] Read the attitude probability component sequence from the behavior probability component set, and extract the attitude probability component values ​​for each time position according to the time index order;

[0166] The attitude probability components at adjacent time index positions are compared and calculated. By calculating the difference between the attitude probability components at two time points, the magnitude of the attitude change is obtained.

[0167] The difference calculated at each time index position is stored in the attitude change component sequence, representing the intensity and direction of the attitude change within that time period.

[0168] In this embodiment, step 6, the calculation of the change in execution speed, specifically includes:

[0169] Read the velocity change component sequence and extract the velocity change component values ​​at each time index position according to the time index order;

[0170] The velocity change components of adjacent time index positions are compared and calculated. By calculating the difference between the velocity change components of the current time index position and the previous time index position, the magnitude of the velocity change is obtained.

[0171] The calculated velocity change amplitude is stored in the velocity change component sequence.

[0172] In this embodiment, step 6, performing the displacement change calculation, specifically includes:

[0173] Read the displacement change component sequence and extract the displacement change component values ​​for each time position according to the time index order;

[0174] The displacement change components of adjacent time index positions are compared and calculated. By calculating the difference between the displacement change components of the current time position and the previous time position, the magnitude of the displacement change is obtained.

[0175] The calculated displacement change amplitude is stored in the displacement change component sequence, representing the displacement change amplitude within that time period.

[0176] Example 1:

[0177] To verify the feasibility of this invention in a real-world environment, it was applied to an indoor personnel safety monitoring scenario. In this scenario, millimeter-wave radar equipment is deployed in the monitoring area to continuously collect radar echo signals generated by human activity. In practical applications, traditional video monitoring methods often struggle to reliably identify falls in environments with insufficient lighting, significant obstructions, or high privacy requirements. Furthermore, ordinary motion recognition algorithms are prone to misinterpreting actions such as quickly sitting down, bending over, or picking up objects as falls, leading to false alarms. This invention acquires human motion echo signals using millimeter-wave radar, performs distance, velocity, and angle analysis on the echo signals to generate a radar motion feature sequence, and then constructs a time-window motion feature vector set through time-window feature combination. Subsequently, a deep learning behavior recognition model is used to generate a sequence of behavior state probability vectors. A hierarchical probability bridging model for human motion is established at the probability level. Furthermore, by combining continuous probability flow constraints and non-uniform diffusion adjustment mechanisms, continuous constraints are applied to the probability change process of human behavior, thereby enabling stable identification of falls even in complex motion environments. In practical applications, millimeter-wave radar continuously scans and acquires data from the monitoring area. The radar echo signal undergoes a Fast Fourier Transform (FFT) in the range dimension to obtain the human target's range component, followed by a Fast Fourier Transform in the velocity dimension to obtain the human motion velocity component. Then, the spatial angle component of the human body is calculated using array phase difference. The range, velocity, and angle components are combined at a unified time index position to generate a radar motion feature sequence. Subsequently, a time window division operation is performed on the continuous time sequence. Within each time window, position change, velocity change, and trajectory change are calculated on the human motion feature vector to generate a time window motion change vector. A set of time window motion feature vectors is then generated through statistical feature calculation. A deep learning behavior recognition model reads the time window motion feature vector set and outputs a sequence of behavior state probability vectors. An attitude change probability layer, a velocity change probability layer, and a spatial displacement probability layer are established within the behavior probability sequence. A behavior probability bridging matrix is ​​constructed through probability bridging operations, and a behavior probability distribution sequence is generated through cross-layer mapping. After the behavior probability distribution sequence is generated, the difference between the probability changes of adjacent time positions is calculated to obtain a probability change amplitude sequence. This is then normalized to generate a probability flow change index sequence. The probability flow change index is used to determine the time-uniform diffusion coefficient. This diffusion coefficient is written into the probability bridging matrix, and diffusion adjustment operations are performed on each probability component to form a continuous probability flow sequence. This continuous probability flow sequence reflects the probabilistic evolution of human movement over time. The probabilities of posture change, velocity change, and displacement change are extracted from the continuous probability flow sequence. A behavioral state probability vector is generated through probability combination operations, and the fall behavior probability component is extracted to form a fall behavior probability sequence. When the fall behavior probability exceeds a set threshold, the fall detection result is output.Under continuous operating conditions, long-term tests were conducted on different types of human movements, including normal walking, bending over to pick up objects, sitting down, standing and turning, and actual falls. The results were statistically analyzed to form experimental data, which was used to verify the stability and accuracy of the invention in fall detection. Test results show that in complex human movement environments, the invention maintains high recognition accuracy while significantly reducing the probability of false alarms, especially under conditions of drastic posture changes that are not falls, where it still maintains stable recognition performance.

[0178] Table 1: Comparison of Human Behavior Detection Results

[0179] Normal walking 320 316 4 98.75 42 Bending down to pick up items 210 205 5 97.62 45 Sitting down 180 174 6 96.67 47 Standing and turning 150 146 4 97.33 43 Slow fall 120 118 2 98.33 41 Falling quickly 130 128 2 98.46 40 Side fall 110 108 2 98.18 41 Falling forward 105 103 2 98.10 42

[0180] As shown in Table 1, out of a total of 1325 human motion samples, the system correctly identified normal walking movements 316 times with only 4 false positives, achieving an accuracy rate of 98.75%. The bending-over-to-pick-up movement was tested 210 times, with 205 correct identifications and 5 false positives, resulting in an accuracy rate of 97.62%. The sitting-down movement was tested 180 times, with 174 correct identifications, achieving an accuracy rate of 96.67%. These movements are easily misidentified as falls in traditional fall detection algorithms, but the number of false positives is significantly reduced in the method of this invention. For actual falls, slow falls were correctly identified 118 times out of 120 samples, fast falls 128 times out of 130 samples, lateral falls 108 times out of 110 samples, and forward falls 103 times out of 105 samples, maintaining an overall fall recognition accuracy rate above 98%. Meanwhile, the average recognition time for various actions remained within the range of 40 to 47 milliseconds, indicating that the algorithm maintains both recognition accuracy and fast computation speed. Through a probability bridging structure and a non-uniform diffusion probability flow adjustment mechanism, the behavioral probability sequence maintains continuous change in the time dimension, thereby avoiding the impact of instantaneous action fluctuations on the detection results. Therefore, it can maintain stable recognition performance even under bending, sitting, and turning action conditions. Comprehensive statistical results show that the method of this invention still has stable detection capabilities in complex human action environments. In 1325 test samples, the overall recognition accuracy reached 97.96%, and the false positive rate was controlled within 2.04%, verifying the feasibility of this invention in real-world human fall detection scenarios.

[0181] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fall detection method using millimeter-wave radar based on AI deep learning, characterized in that, Includes the following steps: Step 1: Obtain the sequence of human motion echo signals from millimeter-wave radar, and perform range analysis, velocity analysis, and angle analysis on the echo signals to generate a radar motion feature sequence; Step 2: Perform time window division and feature combination operations on the radar motion feature sequence to generate a set of time window motion feature vectors; Step 3: Establish a deep learning behavior recognition model. Input the set of motion feature vectors within the time window into the behavior recognition model to generate a sequence of behavior state probability vectors. Step 4: Construct a hierarchical probabilistic bridging model for human motion. Establish a probability layer for posture change, a probability layer for velocity change, and a probability layer for spatial displacement in the sequence of behavioral state probability vectors. Perform probability bridging operations on each probability layer to generate a behavioral probability bridging matrix, and perform cross-layer mapping to generate a sequence of behavioral probability distributions. Step 5: Construct a non-uniform diffusion bridging adjustment structure with continuous probability flow constraints, calculate the probability flow change index for the behavior probability distribution sequence, configure the time non-uniform diffusion coefficient, and perform diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence. Step 6: Perform behavior state recognition operation on the continuous probability stream sequence to generate a fall behavior probability sequence; Step 7: Perform a threshold comparison operation on the fall behavior probability sequence to generate fall detection results.

2. The millimeter-wave radar fall detection method based on AI deep learning according to claim 1, characterized in that, Step 1 specifically includes: Start the millimeter-wave radar data acquisition module to perform continuous scanning acquisition operations on the monitoring area, record the radar echo signals corresponding to each scanning cycle, and establish a radar echo signal sequence according to the scanning time order. Read the radar echo signal sequence, perform a range-dimensional fast Fourier transform operation on the radar echo signal sequence to generate a range spectrum matrix, and extract the human target range component sequence in the range spectrum matrix according to the range index position; Perform a velocity-dimensional fast Fourier transform operation on the distance spectrum matrix to generate a velocity spectrum matrix, and extract the human motion velocity component sequence from the velocity spectrum matrix according to the velocity index position; Angle analysis operation is performed on the radar echo signal sequence, the angle spectrum matrix is ​​calculated through the phase difference of the array channels, and the spatial angle component sequence of the human target is extracted from the angle spectrum matrix. Read the human target distance component sequence, human motion velocity component sequence, and human spatial angle component sequence, perform vector combination operation according to the time index position, and generate human motion feature vector at each time index position; The motion feature vectors of each human body are collected in chronological order to establish a radar motion feature sequence.

3. The fall detection method based on AI deep learning using millimeter-wave radar according to claim 2, characterized in that, Step 2 includes the following steps: Read the radar motion feature sequence, perform time window division operation on the radar motion feature sequence according to a unified time index order, generate a set of time windows, and extract human motion feature vectors according to time index position within each time window; Within each time window, read the human motion feature vector, perform position difference calculation on the position component of the human motion feature vector to generate a sequence of position change components, perform velocity difference calculation on the velocity component to generate a sequence of velocity change components, and perform trajectory direction change calculation on the trajectory component to generate a sequence of trajectory change components. Read the position change component sequence, velocity change component sequence, and trajectory change component sequence, and perform component combination operation according to the time index position to generate a time window motion change vector sequence; Within each time window, statistical feature calculations are performed on the motion change vector sequence of the time window to generate a statistical feature vector. The statistical feature vector contains statistical components of position change, velocity change, and trajectory change. Read the statistical feature vectors, perform feature vector concatenation operations on the statistical feature vectors, and generate a set of motion feature vectors for each time window.

4. The millimeter-wave radar fall detection method based on AI deep learning according to claim 3, characterized in that, Step 3 specifically includes: Read the set of motion feature vectors in the time window, perform a sequence permutation operation on the motion feature vectors in the time window according to the time index order to generate a motion feature sequence in the time window, perform vector normalization operation on the motion feature sequence in the time window, and perform numerical standardization processing on each component of the motion feature vector at each time index position to generate a normalized feature sequence. The normalized feature sequence is input into the deep learning behavior recognition model. The values ​​of each motion feature vector are written in the time index order in the model input layer to generate the model input feature matrix. In the deep learning behavior recognition model, feature mapping operation is performed on the model input feature matrix to generate a sequence of behavior representation vectors. Perform probability mapping operations on the behavior representation vector sequence, perform probability normalization calculations on the behavior feature components at each time index position to generate behavior state probability vectors, and aggregate the behavior state probability vectors in time index order to generate a behavior state probability vector sequence.

5. The millimeter-wave radar fall detection method based on AI deep learning according to claim 4, characterized in that, Step 4, generating the behavior probability bridging matrix, specifically includes: Read the sequence of behavior state probability vectors, extract the values ​​of each probability component in the behavior state probability vector at each time index position, perform a hierarchical partitioning operation according to the probability component category to generate an attitude change probability layer, a velocity change probability layer and a spatial displacement probability layer, record the attitude change probability component sequence in the attitude change probability layer, record the velocity change probability component sequence in the velocity change probability layer, and record the spatial displacement probability component sequence in the spatial displacement probability layer. In the attitude change probability layer, attitude change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the attitude change probability components to generate attitude probability bridging components. In the velocity change probability layer, velocity change probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the velocity change probability components to generate velocity probability bridging components. In the spatial displacement probability layer, spatial displacement probability components are read in time index order. At adjacent time index positions, a probability connection operation is performed on the spatial displacement probability components to generate displacement probability bridging components. The attitude probability bridging components are subjected to a sequence permutation operation to generate an attitude probability bridging sequence in time index order; the velocity probability bridging components are subjected to a sequence permutation operation to generate a velocity probability bridging sequence; and the displacement probability bridging components are subjected to a sequence permutation operation to generate a displacement probability bridging sequence. Read the attitude probability bridging sequence, velocity probability bridging sequence and displacement probability bridging sequence, perform matrix writing operation on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at a unified time index position, and write the values ​​of the three types of probability bridging components at the corresponding row and column positions of the matrix to generate the behavior probability bridging matrix. Record the attitude bridging probability value, velocity bridging probability value and displacement bridging probability value corresponding to each time index position in the behavior probability bridging matrix, and establish a sequence of behavior probability bridging matrices according to the time index order.

6. The millimeter-wave radar fall detection method based on AI deep learning according to claim 5, characterized in that, Step 4, generating the probability distribution sequence of behaviors, includes the following steps: Read the behavior probability bridging matrix, extract the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component at the unified time index position, and perform cross-layer mapping operation on the three types of probability bridging components to generate cross-layer probability mapping vector. In the process of generating cross-layer probability mapping vectors, component alignment operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a probability component alignment sequence. In the probability component alignment sequence, weight combination operations are performed on the attitude probability bridging component, velocity probability bridging component and displacement probability bridging component to generate a unified behavior probability component sequence. Generate a behavior probability distribution sequence by assembling the unified behavior probability component sequences in time index order, and record the set of behavior probability values ​​corresponding to each time index position in the behavior probability distribution sequence.

7. The millimeter-wave radar fall detection method based on AI deep learning according to claim 6, characterized in that, Step 5, which involves calculating the probability flow change index of the behavioral probability distribution sequence and configuring the time non-uniform diffusion coefficient, specifically includes: Read the behavior probability distribution sequence, extract the behavior probability distribution vector at a unified time index position, perform probability difference calculation on the behavior probability distribution vector at adjacent time index positions to generate a probability change component sequence, and record the value of each probability component in the probability change component sequence; Perform absolute value calculation on the probability change component sequence to generate probability change amplitude component sequence, perform component accumulation operation on the probability change amplitude component sequence to generate probability change amplitude value corresponding to time index position, and record probability change amplitude value sequence at each time index position; Read the probability change magnitude numerical sequence, perform sequence normalization operation on the probability change magnitude numerical sequence to generate a probability flow change index sequence, and record the probability flow change index value corresponding to each time index position in the probability flow change index sequence. Read the probability flow change index sequence, perform diffusion coefficient mapping operation on the probability flow change index values ​​at a unified time index position to generate a time non-uniform diffusion coefficient sequence, and record the diffusion coefficient values ​​corresponding to each time index position in the time non-uniform diffusion coefficient sequence.

8. The millimeter-wave radar fall detection method based on AI deep learning according to claim 7, characterized in that, Step 5, which involves performing a diffusion adjustment operation on the behavior probability bridging matrix to generate a continuous probability flow sequence, includes the following steps: Read the behavior probability bridging matrix sequence, extract the attitude bridging probability component, velocity bridging probability component and displacement bridging probability component at a unified time index position, and establish a probability bridging component vector sequence in time index order; Read the time-uniform diffusion coefficient sequence, write the diffusion coefficient values ​​at each time index position into the probability bridging component vector sequence, and perform a diffusion weight mapping operation on the probability bridging component vector sequence to generate a diffusion weight vector sequence. In the diffusion weight vector sequence, the attitude bridging probability component, velocity bridging probability component, and displacement bridging probability component are multiplied by diffusion weights to generate a diffusion probability component sequence. The diffusion probability component sequences are collected in time index order, and a probability normalization operation is performed on the diffusion probability components at each time index position to generate a continuous probability flow vector. A continuous probability stream sequence is generated by assembling continuous probability stream vectors in time index order.

9. The millimeter-wave radar fall detection method based on AI deep learning according to claim 8, characterized in that, Step 6, which involves performing behavior state recognition operations to generate a fall behavior probability sequence, specifically includes: Read the continuous probability flow sequence, extract the continuous probability flow vector at a unified time index position, and extract the attitude probability component, velocity probability component and displacement probability component from the continuous probability flow vector to generate a set of behavior probability components. In the set of behavior probability components, attitude change calculation operations are performed on attitude probability components to generate attitude change component sequences, velocity change calculation operations are performed on velocity probability components to generate velocity change component sequences, displacement change calculation operations are performed on displacement probability components to generate displacement change component sequences, and at a unified time index position, probability combination operations are performed on attitude change components, velocity change components and displacement change components to generate behavior state probability vectors. Extract the fall behavior probability component from the behavior state probability vector, record the fall behavior probability value at each time index position, and collect the fall behavior probability values ​​in time index order to generate a fall behavior probability sequence.

Images