Elevator personnel presence detection method and system based on millimeter wave radar

Abstract

The present invention provides a method and system for detecting the presence of personnel in an elevator based on millimeter-wave radar, which relates to the field of detection technology. The method comprises the following steps: orthogonally arranging a positioning radar and a motion detection radar on the top of an elevator car, collecting radar echo signals and performing baseband demodulation; processing the positioning baseband signals and the motion detection baseband signals respectively by a dynamic threshold elimination method and an adaptive harmonic suppression algorithm, and performing compensation enhancement in combination with the geometric structure of the car; performing fast Fourier transform of distance, speed and height dimensions on the enhanced signals respectively, adopting an adaptive spatial spectrum decomposition positioning algorithm to obtain target three-dimensional coordinate data, and obtaining target motion trajectory data by non-uniform sampling Doppler processing; and adopting a recursive state estimation model to perform time series tracking to realize human height judgment, which is used for safety monitoring in the car: early warning of children and pets entering the car alone and early warning of long-term stay in the car.

Description

Technical Field

[0001] The present invention relates to the field of detection technology, and in particular to a method and system for detecting the presence of elevator personnel based on millimeter wave radar. Background Art

[0002] With the continuous improvement of the intelligence level of elevators, elevator energy-saving control technology has become a research hotspot. At present, elevator personnel presence detection mainly uses infrared sensors, cameras and other sensing devices to achieve energy-saving control by detecting the status of people in the car. Among them, the detection accuracy of infrared sensors is low; the visual detection method based on cameras has reduced performance in low-light environments and is prone to cause privacy concerns for passengers. Millimeter-wave radar has the advantages of strong penetration, all-weather operation, and no privacy leakage, and has broad application prospects in the field of elevator personnel detection.

[0003] However, the existing elevator personnel detection method based on millimeter-wave radar still has some problems. First, the elevator car is a closed metal space with serious multipath effects and electromagnetic interference, which affects the quality of the target echo signal; secondly, it is difficult for a single radar to simultaneously meet the performance requirements of spatial positioning and motion detection, resulting in inaccurate detection results; the existing feature extraction method does not fully describe the characteristics of human targets, making it difficult to effectively distinguish human targets from other objects, reducing detection reliability.

[0004] In summary, there is an urgent need for a method for detecting the presence of people in elevators based on millimeter-wave radar. By orthogonally arranging positioning radar and motion detection radar, suppressing multipath interference, extracting target features, and realizing accurate recognition and tracking of human targets, the present invention can effectively improve the accuracy and reliability of elevator personnel presence detection and provide reliable technical support for elevator energy-saving control. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for detecting the presence of people in an elevator based on millimeter-wave radar, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for detecting the presence of a person in an elevator based on a millimeter-wave radar is provided, comprising:

[0008] A millimeter-wave radar array is orthogonally arranged on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; echo signals received by the positioning radar and the motion detection radar are collected, and orthogonal demodulation is performed to obtain a positioning baseband signal and a motion detection baseband signal respectively; multipath echoes in the positioning baseband signal are removed by a dynamic threshold elimination method to obtain a first target echo signal; clutter components in the motion detection baseband signal are eliminated by an adaptive harmonic suppression algorithm to obtain a second target echo signal; and compensation and enhancement processing are performed on the first target echo signal and the second target echo signal respectively according to a pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal;

[0009] Performing a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extracting the spatial position information of the target, and using an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; performing a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extracting the motion information of the target, and performing non-uniform sampling Doppler processing on the motion information to obtain motion trajectory data of the target including a velocity vector; inputting the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, performing feature extraction through a position branch and a motion branch, and obtaining target fusion features;

[0010] According to the target fusion feature, human feature parameters are extracted, and the human feature parameters include scattering cross-sectional area parameters and motion feature parameters; the scattering cross-sectional area parameters are matched with a preset human scattering cross-sectional area template to obtain a first matching coefficient, and the motion feature parameters are matched with a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, the target is determined to be a human target; a recursive state estimation model is used to perform time-series tracking on the human target to obtain the real-time position and motion state of the human target, and the number of people in the elevator car is calculated; when the number of people is zero and continues for a preset time, an unmanned signal is sent to the elevator control system to make the elevator enter an energy-saving standby mode.

[0011] In an optional embodiment,

[0012] Removing multipath echoes from the positioning baseband signal by a dynamic threshold elimination method to obtain a first target echo signal includes:

[0013] The positioning baseband signal is divided into a plurality of adjacent range window signals, envelope detection is performed on each of the range window signals to obtain a signal amplitude envelope, and a mean value of the signal amplitude envelope is calculated by sliding average according to a preset sliding window length to obtain a power estimation value;

[0014] A constant false alarm rate detection method is adopted to calculate the background noise power mean of each of the distance window signals, and the background noise power mean is subjected to exponential smoothing to obtain the noise variance; a constant false alarm rate equation is established according to a preset false alarm probability and the noise variance;

[0015] Normalizing each noise variance to obtain a weight coefficient, and multiplying the weight coefficient by a solution of a constant false alarm rate equation to obtain an adaptive threshold value;

[0016] The signal with an amplitude greater than the corresponding adaptive threshold value in each distance window signal is marked as a valid target echo, and the valid target echo is subjected to neighborhood smoothing processing by adopting Gaussian weighted averaging; the smoothed distance window signals are merged by using the overlap-addition method to obtain the first target echo signal.

[0017] In an optional embodiment,

[0018] The clutter component in the motion detection baseband signal is eliminated by an adaptive harmonic suppression algorithm to obtain the second target echo signal including:

[0019] Performing wavelet decomposition on the motion detection baseband signal to obtain multi-scale wavelet coefficients; constructing an interference feature vector based on the multi-scale wavelet coefficients, including amplitude and phase information of multiple interference components;

[0020] Performing autocorrelation analysis on the motion detection baseband signal to obtain a time domain autocorrelation sequence, extracting a peak point from the time domain autocorrelation sequence, and determining a signal period;

[0021] Calculating the harmonic fundamental frequency according to the signal period, determining a harmonic frequency group according to the harmonic fundamental frequency, combining the harmonic frequency group with the interference feature vector to generate an interference feature template, wherein the interference feature template includes harmonic frequency and amplitude;

[0022] Constructing an adaptive notch filter group based on the interference feature template, the adaptive notch filter group consists of a plurality of cascaded second-order notch filter units, the center frequencies of the second-order notch filter units corresponding one-to-one to the harmonic frequencies in the interference feature template; dynamically adjusting the quality factor of the second-order notch filter unit according to the amplitude in the interference feature vector;

[0023] The motion detection baseband signal is filtered through the adaptive notch filter group to obtain a filtered signal, and the amplitude of the filtered signal is normalized to obtain a second target echo signal.

[0024] In an optional embodiment,

[0025] Adopting the adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information, the three-dimensional coordinate data of the target is obtained, including:

[0026] Based on a preset number of distance spectrum windows, the distance spectrum is divided into a corresponding number of overlapping sliding windows, and in each overlapping sliding window, a preset number of data points are sampled to construct a multidimensional data matrix; singular value decomposition is performed on the multidimensional data matrix to obtain a left singular vector matrix, a singular value diagonal matrix and a right singular vector matrix;

[0027] Constructing an autocorrelation matrix of a distance spectrum using the left singular vector matrix and the singular value diagonal matrix, performing eigenvalue decomposition on the autocorrelation matrix to obtain an eigenvalue sequence and an eigenvector matrix; calculating a cumulative energy ratio based on the eigenvalue sequence, and determining a signal subspace dimension according to the cumulative energy ratio;

[0028] Based on the signal subspace dimension, extract the eigenvectors of the number of columns corresponding to the dimension from the eigenvector matrix to construct a signal subspace matrix; perform a conjugate transpose operation on the signal subspace matrix to obtain a transposed matrix, and calculate the product of the signal subspace matrix and the transposed matrix to obtain a projection matrix;

[0029] Orthogonally projecting the spatial position information through the projection matrix to obtain a projected position matrix, calculating the correlation coefficients between adjacent windows in the position matrix to obtain a spatial correlation matrix; multiplying the amplitude product of the spatial correlation matrix and the position matrix as a peak detection function; searching for a local maximum point in the peak detection function to obtain a target candidate position;

[0030] K-means clustering is performed on the candidate target positions, and the three-dimensional coordinate data of the target is determined based on the cluster center.

[0031] In an optional embodiment,

[0032] The motion information is subjected to non-uniform sampling Doppler processing to obtain the motion trajectory data of the target including the velocity vector, including:

[0033] Calculating the real part and the imaginary part of the velocity spectrum as the orthogonal component and the in-phase component respectively; calculating the envelope and phase information based on the orthogonal component and the in-phase component;

[0034] Calculate the instantaneous frequency of the target according to the phase difference between adjacent sampling points; construct a new sampling interval sequence using the inverse of the instantaneous frequency; interpolate and reconstruct the envelope and phase information according to the sampling interval sequence to obtain a resampled signal;

[0035] Calculating the instantaneous acceleration according to the amplitude change rate of the resampled signal; calculating the mean and standard deviation of the instantaneous acceleration, calculating and setting the acceleration threshold according to the mean and standard deviation; constructing a weighting coefficient of the instantaneous acceleration based on the instantaneous acceleration and the acceleration threshold;

[0036] The resampled signal is weighted by the weighting coefficient to obtain a weighted sampled signal; the frequency component of the weighted sampled signal is calculated by weighted least square method to obtain the radial velocity of the target;

[0037] Using the relative azimuth and elevation angles of the target's current position and the radar, a conversion matrix from radial velocity to a spatial rectangular coordinate system is established; the radial velocity is mapped through the conversion matrix to obtain an initial value of the three-dimensional velocity, and a state prediction equation is established;

[0038] Taking a weighted average of the prediction result of the state prediction equation and the current three-dimensional velocity initial value to obtain an optimized three-dimensional velocity, calculating an intermediate estimate of the optimized three-dimensional velocity and the instantaneous acceleration at the midpoint of the time interval, and performing a prediction correction on the target position based on the intermediate estimate to obtain a target position change;

[0039] The position change is superimposed on the initial position of the target to obtain a motion trajectory, and the motion trajectory is combined with the optimized three-dimensional velocity to obtain motion trajectory data of the target including a velocity vector.

[0040] In an optional embodiment,

[0041] The three-dimensional coordinate data and the motion trajectory data are input into a dual-stream feature fusion network, and feature extraction is performed through the position branch and the motion branch to obtain target fusion features including:

[0042] Convert the three-dimensional coordinate data into a voxel grid according to a point cloud format, perform density clustering on the voxel grid to obtain a spatial point cluster, calculate the covariance matrix of each voxel in the spatial point cluster, extract the eigenvalue of the covariance matrix, construct a geometric descriptor sequence, construct a matching graph for the geometric descriptor sequence according to a spatial neighborhood relationship, and extract spatial domain features based on the matching graph;

[0043] Projecting the motion trajectory data onto the spherical harmonic function space to obtain spherical harmonic coefficients, extracting the time series variation sequence of the spherical harmonic coefficients, constructing a Markov jump model based on the time series variation sequence, calculating the corresponding state transition probability matrix, and extracting time domain motion features based on the steady-state distribution of the state transition probability matrix;

[0044] Constructing an affinity matrix of the spatial domain features and the temporal motion features, decomposing the affinity matrix using a spectral clustering method to obtain a feature subspace, projecting the spatial domain features and the temporal motion features onto the feature subspace to obtain a projection feature, calculating a geodesic distance matrix of the projection feature, constructing a feature manifold graph, performing a random walk on the feature manifold graph to obtain an access frequency sequence, and constructing a feature grouping matrix based on the access frequency sequence;

[0045] Based on the feature grouping matrix, hierarchical grouping is performed on the projection features to obtain feature groups, the local covariance structure of the feature groups is calculated, the main direction vectors of the local covariance structure are extracted, an orthogonal transformation matrix is ​​constructed, and the feature groups are mapped to an orthogonal space through the orthogonal transformation matrix to obtain decoupled features;

[0046] Graph Fourier transform is performed on the decoupled features to obtain frequency domain features, a multi-scale feature pyramid is constructed based on the frequency domain features, the multi-scale feature pyramid is processed by a recursive graph neural network, hierarchical representation features are extracted, and the target fusion features are obtained by combining adaptive aggregation.

[0047] In an optional embodiment,

[0048] Extracting human feature parameters according to the target fusion features, the human feature parameters include scattering cross-sectional area parameters and motion feature parameters; performing matching calculation on the scattering cross-sectional area parameters and a preset human scattering cross-sectional area template to obtain a first matching coefficient, and performing matching calculation on the motion feature parameters and a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold, determining that the target is a human target; using a recursive state estimation model to perform time-series tracking on the human target, obtaining the real-time position and motion state of the human target, and calculating the number of people in the elevator car; when the number of people is zero and continues for a preset time, sending an unmanned signal to the elevator control system to make the elevator enter an energy-saving standby mode, including:

[0049] Extracting spectrum distribution features from the target fusion features to obtain human body scattering cross-sectional area parameters, extracting trajectory continuity features to obtain human body motion feature parameters; performing wavelet decomposition on the human body scattering cross-sectional area parameters to obtain multi-scale spectrum coefficients, performing cross-correlation operation on the multi-scale spectrum coefficients and a preset human body scattering cross-sectional area template to obtain a first matching coefficient; constructing a phase space trajectory for the human body motion feature parameters, performing dynamic time warping calculation on the phase space trajectory and a preset human body motion feature template to obtain a second matching coefficient;

[0050] Based on the first matching coefficient and the second matching coefficient, a two-dimensional feature vector is constructed, and the two-dimensional feature vector is input into a preset support vector classifier to determine the target category; when the support vector classifier outputs the human target category, a state vector including position coordinates, velocity vector, and acceleration vector is constructed to initialize the state distribution of the particle swarm;

[0051] A recursive state estimation model is constructed, the recursive state estimation model comprising: constructing a state transfer equation based on the state vector, discretizing the state transfer equation to obtain a state prediction model; extracting a target scattering intensity sequence from the multi-scale spectrum coefficient, constructing an observation equation using the target scattering intensity sequence, and using the observation equation as a likelihood function of a particle weight; using the state prediction model to perform state prediction on a particle swarm, and calculating the weight of each particle based on the observation equation; performing importance resampling on the particle swarm to obtain an updated particle distribution, and calculating a weighted mean of the particle distribution as a target state estimation value;

[0052] The target state estimation values ​​of the continuous time series are combined into a target state sequence, and the trajectory of the target state sequence is predicted to obtain the target motion trajectory. Based on the target motion trajectory, the real-time number of people in the car is calculated; when the real-time number of people is zero and the duration exceeds the preset vacancy time, an unmanned signal is sent to the elevator control system to trigger the energy-saving standby mode.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is an elevator personnel presence detection system based on millimeter wave radar, comprising:

[0055] The first unit is used to orthogonally arrange a millimeter-wave radar array on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; the echo signals received by the positioning radar and the motion detection radar are collected, and orthogonal demodulation is performed to obtain a positioning baseband signal and a motion detection baseband signal respectively; the multipath echo in the positioning baseband signal is removed by a dynamic threshold elimination method to obtain a first target echo signal; the clutter component in the motion detection baseband signal is eliminated by an adaptive harmonic suppression algorithm to obtain a second target echo signal; the first target echo signal and the second target echo signal are respectively compensated and enhanced according to the pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal;

[0056] The second unit is used to perform a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extract the spatial position information of the target, and use an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; perform a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extract the motion information of the target, perform non-uniform sampling Doppler processing on the motion information, and obtain motion trajectory data of the target including a velocity vector; input the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, perform feature extraction through a position branch and a motion branch, and obtain a target fusion feature;

[0057] The third unit is used to extract human feature parameters according to the target fusion features, wherein the human feature parameters include scattering cross-sectional area parameters and motion feature parameters; perform matching calculation on the scattering cross-sectional area parameters and a preset human scattering cross-sectional area template to obtain a first matching coefficient, and perform matching calculation on the motion feature parameters and a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determine that the target is a human target; use a recursive state estimation model to perform time-series tracking on the human target to obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; when the number of people is zero and continues for a preset time, send an unmanned signal to the elevator control system to put the elevator into energy-saving standby mode.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In the embodiment of the present invention, by orthogonally arranging the positioning radar and the motion detection radar, combined with a series of signal processing technologies such as dynamic threshold elimination method, adaptive harmonic suppression algorithm, compensation enhancement processing and adaptive spatial spectrum decomposition positioning algorithm, multipath echo and clutter interference are effectively eliminated, and accurate extraction of the target spatial position and motion information is achieved. The application of the dual-stream feature fusion network integrates the three-dimensional coordinate data and motion trajectory data of the target, improves the accuracy and robustness of human target recognition, effectively avoids misjudgment, and thus more reliably detects whether there is a person in the elevator car; the use of non-uniform sampling Doppler processing can adapt to the different movement speeds and trajectories of the target, and combines the human scattering cross-sectional area parameters and motion feature parameters for matching, which can effectively identify human targets in different body shapes, postures and motion states. This makes the method more adaptable to various complex scenarios, for example, even when the passenger carries items, has difficulty in movement or has irregular motion trajectories, it can accurately detect; the recursive state estimation model is used to track the human target in time series to monitor the number of people in the car in real time. When it is detected that there is no one in the car for a long time, the system will automatically send an empty signal to the elevator control system, causing the elevator to enter energy-saving standby mode, thereby effectively reducing the energy consumption of the elevator and achieving the goal of energy saving and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flowchart of a method for detecting the presence of personnel in an elevator based on millimeter-wave radar according to an embodiment of the present invention;

[0067] Figure 2 The figure is a schematic diagram of the structure of an elevator personnel presence detection system based on millimeter wave radar according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a method for detecting the presence of personnel in an elevator based on millimeter wave radar according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. Arrange a millimeter-wave radar array orthogonally on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; collect echo signals received by the positioning radar and the motion detection radar, perform orthogonal demodulation, and obtain a positioning baseband signal and a motion detection baseband signal respectively; remove multipath echoes in the positioning baseband signal by a dynamic threshold elimination method to obtain a first target echo signal; eliminate clutter components in the motion detection baseband signal by an adaptive harmonic suppression algorithm to obtain a second target echo signal; perform compensation enhancement processing on the first target echo signal and the second target echo signal respectively according to the pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal;

[0072] S102. Perform a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extract the spatial position information of the target, and use an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; perform a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extract the motion information of the target, perform non-uniform sampling Doppler processing on the motion information, and obtain motion trajectory data of the target including a velocity vector; input the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, perform feature extraction through a position branch and a motion branch, and obtain target fusion features;

[0073] S103. Extract human feature parameters according to the target fusion features, the human feature parameters including scattering cross-sectional area parameters and motion feature parameters; calculate the matching degree between the scattering cross-sectional area parameters and a preset human scattering cross-sectional area template to obtain a first matching coefficient, calculate the matching degree between the motion feature parameters and a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determine that the target is a human target; use a recursive state estimation model to perform time-series tracking on the human target to obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; when the number of people is zero and continues for a preset time, send an unmanned signal to the elevator control system to put the elevator into energy-saving standby mode.

[0074] First, a positioning radar is installed at the center of the top of the elevator car, and a motion detection radar is installed at the edge of the top of the car. These two radars together form an orthogonal millimeter-wave radar array. The two radars work together to obtain more comprehensive target information in the car.

[0075] Next, the positioning radar and the motion detection radar receive the echo signals respectively and perform orthogonal demodulation to obtain the positioning baseband signal and the motion detection baseband signal. In order to improve the signal quality, these two signals need to be denoised. For the positioning baseband signal, the dynamic threshold elimination method is used to remove the multipath echo interference and obtain a clear first target echo signal. For example, a dynamic threshold is set, and the signal below the threshold is considered to be noise or multipath echo and is filtered out. Assuming that the received signal amplitude sequence is [0.1, 0.2, 1.5, 0.3, 0.2, 2.0, 0.1], and the dynamic threshold is set to 0.4, the processed signal sequence is [0, 0, 1.5, 0, 0, 2.0, 0]. For the motion detection baseband signal, an adaptive harmonic suppression algorithm is used to eliminate the clutter component to obtain a purer second target echo signal. For example, by analyzing the spectral characteristics of the signal, the interference components of the harmonic frequency are identified and removed.

[0076] In order to further enhance the signal quality, the first target echo signal and the second target echo signal need to be compensated and enhanced according to the car geometry measured in advance. For example, the signal amplitude and phase can be corrected according to the shape and size of the car to obtain the first enhanced signal and the second enhanced signal.

[0077] The first enhanced signal is subjected to a fast Fourier transform in the distance dimension to obtain a distance spectrum, from which the spatial position information of the target is extracted. Then, an adaptive spatial spectrum decomposition positioning algorithm is used to accurately position the spatial position information, and finally the three-dimensional coordinate data of the target is obtained. For example, by analyzing the peak position and amplitude of the distance spectrum, the three-dimensional coordinates (x, y, z) of the target in the car can be determined.

[0078] The second enhanced signal is subjected to a fast Fourier transform in the velocity dimension to obtain a velocity spectrum, from which the target's motion information is extracted. Then, the motion information is subjected to non-uniform sampling Doppler processing to obtain the target's motion trajectory data including the velocity vector. For example, by analyzing the changes in the velocity spectrum, the speed and direction of the target at different times can be obtained, thereby drawing the target's motion trajectory.

[0079] The obtained 3D coordinate data and motion trajectory data are input into a two-stream feature fusion network. The network contains a position branch and a motion branch, which extract features from the position information and motion information respectively, and finally obtain the target fusion features.

[0080] According to the target fusion features, human body feature parameters are extracted, including scattering cross-sectional area parameters and motion feature parameters. The extracted scattering cross-sectional area parameters are matched with the preset human body scattering cross-sectional area template to obtain a first matching coefficient. For example, assuming that the scattering cross-sectional area of ​​the target is 0.5 square meters and the human body scattering cross-sectional area template is 0.4 to 0.6 square meters, the first matching coefficient is relatively high. At the same time, the extracted motion feature parameters are matched with the preset human body motion feature template to obtain a second matching coefficient. For example, assuming that the target's movement speed is between 0.1 and 1.5 meters per second, and the movement direction changes relatively slowly, which is consistent with the human body motion feature template, the second matching coefficient is relatively high.

[0081] If both the first matching coefficient and the second matching coefficient exceed the preset threshold, the target is determined to be a human target. For example, the threshold is set to 0.8, and when both the first matching coefficient and the second matching coefficient are greater than 0.8, it is considered that a human body is detected.

[0082] The recursive state estimation model is used to track the identified human targets in time series to obtain the real-time position and motion state of the human targets, thereby calculating the number of people in the elevator car. For example, the position and speed of the target are predicted and updated through methods such as Kalman filtering to achieve real-time tracking of the target.

[0083] When it is detected that the number of people in the car is zero and continues for a preset period of time (for example, 1 minute), an empty signal is sent to the elevator control system, causing the elevator to enter energy-saving standby mode.

[0084] In this embodiment, by fusing the data of the positioning radar and the motion detection radar, and combining the dual-stream feature fusion network and human feature parameter matching, human targets and non-human targets can be effectively distinguished, the detection accuracy can be improved, and misjudgment can be reduced; the recursive state estimation model is used to perform time-series tracking of human targets, the position and motion state of the human targets can be obtained in real time, and the number of people in the car can be accurately calculated; by detecting whether there is anyone in the car and putting the elevator into energy-saving standby mode when there is no one, the energy consumption of the elevator can be effectively reduced, energy can be saved, and green environmental protection can be achieved.

[0085] In an optional implementation, removing the multipath echo in the positioning baseband signal by a dynamic threshold elimination method to obtain the first target echo signal includes:

[0086] The positioning baseband signal is divided into a plurality of adjacent range window signals, envelope detection is performed on each of the range window signals to obtain a signal amplitude envelope, and a mean value of the signal amplitude envelope is calculated by sliding average according to a preset sliding window length to obtain a power estimation value;

[0087] A constant false alarm rate detection method is adopted to calculate the background noise power mean of each of the distance window signals, and the background noise power mean is subjected to exponential smoothing to obtain the noise variance; a constant false alarm rate equation is established according to a preset false alarm probability and the noise variance;

[0088] Normalizing each noise variance to obtain a weight coefficient, and multiplying the weight coefficient by a solution of a constant false alarm rate equation to obtain an adaptive threshold value;

[0089] The signal with an amplitude greater than the corresponding adaptive threshold value in each distance window signal is marked as a valid target echo, and the valid target echo is subjected to neighborhood smoothing processing by adopting Gaussian weighted averaging; the smoothed distance window signals are merged by using the overlap-addition method to obtain the first target echo signal.

[0090] The envelope detection specifically refers to a technique in signal processing that is used to extract the amplitude information of a signal, and is usually used for the analysis of modulated signals or echo signals. In radar signal processing, envelope detection can obtain the amplitude envelope of the reflected signal, that is, the intensity variation of the signal over time. The envelope detection process is usually implemented by calculating the amplitude (modulus) of the complex signal (including the real part and the imaginary part) at each sampling point. This step can help with subsequent target detection and feature extraction.

[0091] The constant false alarm rate detection specifically refers to an adaptive threshold detection technology for target detection, which is widely used in radar, sonar and other systems. The goal of constant false alarm rate detection is to maintain a fixed false alarm rate (false alarm rate), that is, in the absence of target signals, the probability of the system falsely reporting a target is constant. To achieve this goal, the constant false alarm rate detection dynamically adjusts the detection threshold value based on the noise background of the signal to keep the false alarm rate unchanged in different noise environments.

[0092] The threshold specifically refers to a threshold used to determine whether a signal is a valid target. In radar echo signal processing, the threshold is used to compare the signal amplitude with the size of the threshold. If the signal amplitude exceeds the threshold, it is considered that there may be a target at the corresponding position of the signal. The choice of threshold has a great impact on the detection performance of the system; if the threshold value is too low, false alarms are likely to occur; if the threshold value is too high, the target may be missed. By adaptively adjusting the threshold through the constant false alarm rate detection method, the threshold can be dynamically adjusted under different noise backgrounds to ensure the reliability of detection.

[0093] First, the received positioning baseband signal is divided into multiple continuous range window signals according to the distance dimension. The length of each range window signal can be set according to the actual situation, for example, set to 10 sampling points. Assuming that the length of the received baseband signal is 1000 sampling points, it can be divided into 100 range window signals.

[0094] Next, envelope detection is performed on each range window signal. The envelope detection method may be a Hilbert transform or an absolute value method. Taking the absolute value method as an example, the absolute value of each sampling point in each range window signal is taken to obtain the amplitude envelope of the range window signal.

[0095] Then, the amplitude envelope of each range window signal is calculated by sliding average to obtain a power estimate. Assuming that the sliding window length is set to 5 sampling points, the average of the first 5 amplitude envelope values ​​is used as the first power estimate, the average of the 2nd to 6th amplitude envelope values ​​is used as the second power estimate, and so on, to obtain multiple power estimates of the range window signal. For example, if the amplitude envelope of a range window signal is [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], the power estimate obtained after sliding average is [3, 4, 5, 6, 7, 8].

[0096] In order to calculate the adaptive threshold value, it is necessary to estimate the background noise power first. The constant false alarm rate detection method is used to calculate the background noise power mean of each distance window signal. The part with a smaller power estimate in the distance window signal can be selected as the background noise, for example, the first 10% of the power estimate is averaged to obtain the background noise power mean. Then, the calculated background noise power mean is used to obtain the noise variance by the exponential smoothing method. The exponential smoothing formula can be adjusted according to the actual situation. For example, the noise power mean at the current moment can be multiplied by a smaller weight coefficient, plus the noise variance at the previous moment can be multiplied by a larger weight coefficient.

[0097] According to the preset false alarm probability and the calculated noise variance, a constant false alarm rate equation is established. The false alarm probability can be set according to actual needs, for example, set to 0.01. The solution of the constant false alarm rate equation can be obtained by table lookup or numerical calculation.

[0098] Each noise variance is normalized to obtain a weight coefficient. The normalization method may be to divide each noise variance by the sum of all noise variances. The weight coefficient is multiplied by the solution of the constant false alarm rate equation to obtain an adaptive threshold value.

[0099] The signal with an amplitude greater than the corresponding adaptive threshold value in each range window signal is marked as a valid target echo. Gaussian weighted averaging is used to perform neighborhood smoothing on the valid target echo. The window size of Gaussian weighted averaging can be set according to actual conditions, for example, set to 3 sampling points.

[0100] Finally, the smoothed range window signals are combined using the overlap-add method to obtain the first target echo signal. The length of the overlap portion of the overlap-add method can be set according to actual conditions, for example, to half the length of the sliding window.

[0101] In this embodiment, by dynamically adjusting the threshold value, the interference of background noise and multipath echo can be effectively suppressed, thereby improving the accuracy of target echo detection; Gaussian weighted averaging is used for neighborhood smoothing to effectively remove noise and burrs, thereby enhancing the robustness of the target echo; the implementation process is relatively simple, the amount of calculation is small, and it is easy to implement in practical applications.

[0102] In an optional implementation, eliminating the clutter component in the motion detection baseband signal by an adaptive harmonic suppression algorithm to obtain the second target echo signal includes:

[0103] Performing wavelet decomposition on the motion detection baseband signal to obtain multi-scale wavelet coefficients; constructing an interference feature vector based on the multi-scale wavelet coefficients, including amplitude and phase information of multiple interference components;

[0104] Performing autocorrelation analysis on the motion detection baseband signal to obtain a time domain autocorrelation sequence, extracting a peak point from the time domain autocorrelation sequence, and determining a signal period;

[0105] Calculating the harmonic fundamental frequency according to the signal period, determining a harmonic frequency group according to the harmonic fundamental frequency, combining the harmonic frequency group with the interference feature vector to generate an interference feature template, wherein the interference feature template includes harmonic frequency and amplitude;

[0106] Constructing an adaptive notch filter group based on the interference feature template, the adaptive notch filter group consists of a plurality of cascaded second-order notch filter units, the center frequencies of the second-order notch filter units corresponding one-to-one to the harmonic frequencies in the interference feature template; dynamically adjusting the quality factor of the second-order notch filter unit according to the amplitude in the interference feature vector;

[0107] The motion detection baseband signal is filtered through the adaptive notch filter group to obtain a filtered signal, and the amplitude of the filtered signal is normalized to obtain a second target echo signal.

[0108] The harmonic fundamental frequency specifically refers to the most basic periodic component frequency in the signal, that is, the inverse of the signal period. For a periodic signal, the harmonic fundamental frequency is the frequency multiple basis of all harmonic frequencies in the signal. The fundamental frequency can be determined by the peak point of the autocorrelation sequence. Many periodic signals can be decomposed into harmonic components (fundamental frequency and its multiples), and the fundamental frequency determines the basic repetition rate of the signal.

[0109] The harmonic frequency group specifically refers to a set of integer multiple frequencies generated on the basis of the fundamental frequency. If the fundamental frequency of the signal is f0, its harmonic frequency group includes f0, 2f0, 3f0, ... and other frequencies. The frequency components in the harmonic frequency group reflect the spectral characteristics of the signal, which plays an important role in signal analysis and filter design. In interference suppression, the interference within a specific frequency range can be effectively reduced by processing the harmonic frequencies.

[0110] The second-order notch filter unit is specifically a signal processing unit used to remove or suppress interference from specific frequency components. The second-order notch filter can efficiently filter out interference signals in the target frequency range by designing its center frequency (or notch frequency) and bandwidth. The "second order" of the notch filter means that the order of the filter transfer function is two, that is, its mathematical model is a second-order equation. By setting different center frequencies, multiple second-order notch filter units can be cascaded into a notch filter group to filter out multiple harmonic frequency interference components.

[0111] The quality factor specifically refers to a parameter that measures the selectivity of a filter or resonant circuit, and reflects the sharpness of the filter's response to the target frequency. For a notch filter, the quality factor is defined as the ratio of the center frequency f0 to the bandwidth, that is, Q=f0 / bandwidth. Notch filters with high Q values ​​have a narrower response to the target frequency and can filter out the target frequency more accurately; while notch filters with low Q values ​​have a wider bandwidth and can filter out frequency components in a larger range. In an adaptive notch filter, the quality factor can be dynamically adjusted according to the interference amplitude, thereby flexibly controlling the filtering effect.

[0112] First, the motion detection baseband signal is decomposed by wavelet. The db4 wavelet is selected, and the decomposition layer is 5. For example, a baseband signal with a sampling rate of 1000Hz, after 5 layers of wavelet decomposition, obtains 5 detail coefficient subbands and 1 approximate coefficient subband. The frequency range of each subband decreases successively, representing different frequency components of the signal.

[0113] Then, the interference feature vector is constructed based on the multi-scale wavelet coefficients. The energy and peak frequency of each detail coefficient subband are extracted as feature vector elements. For example, after 5-layer wavelet decomposition, 5 detail coefficient subbands are obtained. Each subband extracts two features, energy and peak frequency. Then the interference feature vector contains 10 elements, covering the amplitude and phase information of interference components in different frequency ranges.

[0114] Next, the motion detection baseband signal is subjected to autocorrelation analysis. The original baseband signal is subjected to autocorrelation operation to obtain a time domain autocorrelation sequence. For example, a baseband signal with a length of 1000 is subjected to autocorrelation analysis to obtain an autocorrelation sequence with a length of 1999.

[0115] Extract the peak points from the time domain autocorrelation sequence to determine the signal period. In the autocorrelation sequence, except for the maximum peak at the zero point, the other peak points correspond to the period information of the signal. For example, in the autocorrelation sequence, the first peak point appears at a lag time of 100, and the signal period is 100 sampling points.

[0116] Calculate the harmonic fundamental frequency based on the signal period. The harmonic fundamental frequency is the inverse of the frequency corresponding to the signal period. For example, if the signal period is 100 sampling points and the sampling rate is 1000 Hz, the harmonic fundamental frequency is 10 Hz.

[0117] The harmonic frequency group is determined according to the harmonic fundamental frequency. The harmonic frequency group contains frequencies that are integer multiples of the harmonic fundamental frequency. For example, if the fundamental frequency is 10 Hz, the harmonic frequency group may contain frequency components such as 20 Hz, 30 Hz, and 40 Hz.

[0118] The harmonic frequency group is combined with the interference feature vector to generate an interference feature template. The interference feature template contains the harmonic frequency and amplitude information. For example, the interference feature template can be represented as a table, where one column is the harmonic frequency and the other column is the corresponding amplitude, and the amplitude can be extracted from the interference feature vector.

[0119] An adaptive notch filter group is constructed based on the interference feature template. The notch filter group consists of multiple cascaded second-order notch filter units, and the center frequency of each second-order notch filter unit corresponds to the harmonic frequency in the interference feature template. For example, if the interference feature template contains 3 harmonic frequencies, the notch filter group consists of 3 cascaded second-order notch filter units.

[0120] The quality factor of the second-order notch filter unit is dynamically adjusted according to the amplitude in the interference characteristic vector. For example, if the amplitude corresponding to a harmonic frequency is large, the quality factor of the corresponding second-order notch filter unit is set higher to more effectively suppress the frequency component.

[0121] The motion detection baseband signal is filtered through an adaptive notch filter group to obtain a filtered signal.

[0122] The amplitude of the filtered signal is normalized to obtain a second target echo signal. The amplitude of the filtered signal is adjusted to a preset range, such as [-1, 1], to facilitate subsequent processing and analysis.

[0123] In this embodiment, by effectively suppressing harmonic clutter interference, the signal-to-noise ratio can be improved, thereby making it easier to detect weak target echo signals; eliminating clutter components can reduce false alarms and improve the accuracy of target recognition; the adaptive notch filter can dynamically adjust the filtering parameters according to the interference characteristics, thereby improving the adaptability and robustness of the system in different environments.

[0124] In an optional implementation, the spatial position information is accurately positioned using an adaptive spatial spectrum decomposition positioning algorithm to obtain the three-dimensional coordinate data of the target, including:

[0125] Based on a preset number of distance spectrum windows, the distance spectrum is divided into a corresponding number of overlapping sliding windows, and in each overlapping sliding window, a preset number of data points are sampled to construct a multidimensional data matrix; singular value decomposition is performed on the multidimensional data matrix to obtain a left singular vector matrix, a singular value diagonal matrix and a right singular vector matrix;

[0126] Constructing an autocorrelation matrix of a distance spectrum using the left singular vector matrix and the singular value diagonal matrix, performing eigenvalue decomposition on the autocorrelation matrix to obtain an eigenvalue sequence and an eigenvector matrix; calculating a cumulative energy ratio based on the eigenvalue sequence, and determining a signal subspace dimension according to the cumulative energy ratio;

[0127] Based on the signal subspace dimension, extract the eigenvectors of the number of columns corresponding to the dimension from the eigenvector matrix to construct a signal subspace matrix; perform a conjugate transpose operation on the signal subspace matrix to obtain a transposed matrix, and calculate the product of the signal subspace matrix and the transposed matrix to obtain a projection matrix;

[0128] Orthogonally projecting the spatial position information through the projection matrix to obtain a projected position matrix, calculating the correlation coefficients between adjacent windows in the position matrix to obtain a spatial correlation matrix; multiplying the amplitude product of the spatial correlation matrix and the position matrix as a peak detection function; searching for a local maximum point in the peak detection function to obtain a target candidate position;

[0129] K-means clustering is performed on the candidate target positions, and the three-dimensional coordinate data of the target is determined based on the cluster center.

[0130] First, set the number of distance spectrum windows and the number of data points sampled in each window. For example, divide the distance spectrum into 10 overlapping sliding windows, and sample 100 data points in each window. Then, according to the preset number of windows and sampling points, extract data from the received distance spectrum data in sequence to construct a multidimensional data matrix with 100 rows and 10 columns. Each column represents the data of a window, and each row represents a sampling moment.

[0131] Next, perform singular value decomposition on the constructed multidimensional data matrix. Singular value decomposition can decompose the original matrix into three matrices: a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. For example, through singular value decomposition, a 100×100 left singular vector matrix, a 10×10 singular value diagonal matrix (singular values ​​on the diagonal and the rest of the elements are 0), and a 10×10 right singular vector matrix are obtained.

[0132] The autocorrelation matrix of the distance spectrum is constructed using the left singular vector matrix and the singular value diagonal matrix. The specific method is to multiply the left singular vector matrix with the singular value diagonal matrix, and then multiply the result with the transpose of the left singular vector matrix. A 100×100 autocorrelation matrix is ​​obtained.

[0133] Perform eigenvalue decomposition on the autocorrelation matrix to obtain an eigenvalue sequence and an eigenvector matrix. Eigenvalue decomposition can decompose the autocorrelation matrix into eigenvalues ​​and eigenvectors. For example, 100 eigenvalues ​​are obtained to form a 100×1 eigenvalue sequence and a corresponding 100×100 eigenvector matrix.

[0134] The cumulative energy ratio is calculated based on the eigenvalue sequence. The cumulative energy ratio refers to the ratio of the sum of the first k eigenvalues ​​to the sum of all eigenvalues ​​after the eigenvalues ​​are sorted. By setting a threshold, such as 95%, the minimum k value that satisfies the cumulative energy ratio greater than or equal to the threshold is found, and the k value is determined as the signal subspace dimension. Assuming that the sum of the first 10 eigenvalues ​​accounts for 98% of the total sum of eigenvalues, the signal subspace dimension is 10.

[0135] According to the determined signal subspace dimension, the corresponding number of eigenvectors are extracted from the eigenvector matrix to construct the signal subspace matrix. For example, if the signal subspace dimension is 10, the first 10 columns are extracted from the 100×100 eigenvector matrix to form a 100×10 signal subspace matrix.

[0136] Perform a conjugate transpose operation on the signal subspace matrix to obtain a transposed matrix. Then calculate the product of the signal subspace matrix and its transposed matrix to obtain a projection matrix. For example, transpose the 100×10 signal subspace matrix to a 10×100 matrix, and then multiply it with the 100×10 signal subspace matrix to obtain a 100×100 projection matrix.

[0137] The spatial position information is orthogonally projected through the projection matrix to obtain the projected position matrix. Assuming that the spatial position information is a 100×1 vector, multiply it with the 100×100 projection matrix to obtain a 100×1 projected position matrix.

[0138] Calculate the correlation coefficients between adjacent windows in the position matrix to obtain a spatial correlation matrix. For example, calculate the correlation coefficients between the data of the first window and the second window, the second window and the third window, and so on, in the projected position matrix to obtain a 9×1 spatial correlation matrix.

[0139] The amplitude product of the spatial correlation matrix and the position matrix is ​​used as the peak detection function. For example, the amplitude of the corresponding window data in the 9×1 spatial correlation matrix and the 100×1 position matrix is ​​multiplied to obtain a 9×1 peak detection function vector.

[0140] Search for the local maximum point in the peak detection function to obtain the target candidate position. For example, find the index corresponding to the maximum value in the 9×1 peak detection function vector and the local maximum points around it as the target candidate position.

[0141] Perform K-means clustering on the target candidate locations. For example, use the multiple local maximum points found as the initial cluster centers and perform K-means clustering to cluster these points into one category.

[0142] The three-dimensional coordinate data of the target is determined based on the cluster center. For example, the coordinates of the cluster center obtained after K-means clustering are calculated as the final three-dimensional coordinate data of the target.

[0143] In this embodiment, adaptive spatial spectrum decomposition and orthogonal projection are used to effectively suppress noise interference and improve target positioning accuracy; sliding window and multidimensional data matrix processing are used to enhance the adaptability and robustness of the algorithm to complex environments; singular value decomposition and eigenvalue decomposition are used to reduce computational complexity and improve real-time performance.

[0144] In an optional implementation manner, performing non-uniform sampling Doppler processing on the motion information to obtain motion trajectory data of the target including a velocity vector includes:

[0145] Calculating the real part and the imaginary part of the velocity spectrum as the orthogonal component and the in-phase component respectively; calculating the envelope and phase information based on the orthogonal component and the in-phase component;

[0146] Calculate the instantaneous frequency of the target according to the phase difference between adjacent sampling points; construct a new sampling interval sequence using the inverse of the instantaneous frequency; interpolate and reconstruct the envelope and phase information according to the sampling interval sequence to obtain a resampled signal;

[0147] Calculating the instantaneous acceleration according to the amplitude change rate of the resampled signal; calculating the mean and standard deviation of the instantaneous acceleration, calculating and setting the acceleration threshold according to the mean and standard deviation; constructing a weighting coefficient of the instantaneous acceleration based on the instantaneous acceleration and the acceleration threshold;

[0148] The resampled signal is weighted by the weighting coefficient to obtain a weighted sampled signal; the frequency component of the weighted sampled signal is calculated by weighted least square method to obtain the radial velocity of the target;

[0149] Using the relative azimuth and elevation angles of the target's current position and the radar, a conversion matrix from radial velocity to a spatial rectangular coordinate system is established; the radial velocity is mapped through the conversion matrix to obtain an initial value of the three-dimensional velocity, and a state prediction equation is established;

[0150] Taking a weighted average of the prediction result of the state prediction equation and the current three-dimensional velocity initial value to obtain an optimized three-dimensional velocity, calculating an intermediate estimate of the optimized three-dimensional velocity and the instantaneous acceleration at the midpoint of the time interval, and performing a prediction correction on the target position based on the intermediate estimate to obtain a target position change;

[0151] The position change is superimposed on the initial position of the target to obtain a motion trajectory, and the motion trajectory is combined with the optimized three-dimensional velocity to obtain motion trajectory data of the target including a velocity vector.

[0152] First, the radar system receives the echo signal and performs Doppler processing on it. The real part of the calculated velocity spectrum is taken as the in-phase component, and the imaginary part is taken as the orthogonal component. For example, suppose the in-phase component at a certain moment is 0.8 and the orthogonal component is 0.6.

[0153] Next, the in-phase component and the quadrature component are used to calculate the envelope and phase information of the signal. The envelope can be understood as the change in the signal amplitude, and the phase reflects the relative position of the signal at a specific moment. Taking the above case as an example, the envelope calculation result is (0.8²+0.6²) ½ =1, the phase is arctan(0.6 / 0.8)=36.87°.

[0154] Then, the instantaneous frequency of the target is calculated based on the phase difference between adjacent sampling points. The instantaneous frequency reflects the rate of change of the target speed. Assuming that the phase of the next sampling point is 45.87°, the phase difference is 9°, corresponding to the instantaneous frequency change of the target.

[0155] The reciprocal of the instantaneous frequency is used to construct a new sampling interval sequence. This means that when the target motion state changes drastically and the instantaneous frequency is higher, the sampling interval is shorter, and vice versa. For example, if the instantaneous frequency is 10Hz, the sampling interval is 0.1 seconds; if the instantaneous frequency is 20Hz, the sampling interval is 0.05 seconds.

[0156] The envelope and phase information are interpolated and reconstructed according to the new sampling interval sequence to obtain the resampled signal. This is equivalent to non-uniform sampling of the original signal, and more dense sampling is performed in areas where the target motion changes violently.

[0157] The instantaneous acceleration is calculated based on the amplitude change rate of the resampled signal. The instantaneous acceleration reflects how fast the target speed changes. For example, if the amplitude of the resampled signal changes from 1 to 1.2 in 0.1 seconds, the instantaneous acceleration can be approximated as (1.2-1) / 0.1=2.

[0158] Calculate the mean and standard deviation of all instantaneous accelerations. Assume that the calculated mean is 1 and the standard deviation is 0.5.

[0159] Calculate and set the acceleration threshold based on the mean and standard deviation. The threshold can be set to the mean plus a number of standard deviations, for example, threshold = 1 + 2*0.5 = 2.

[0160] The weighting coefficient of instantaneous acceleration is constructed based on the instantaneous acceleration and the acceleration threshold. The weighting coefficient is used to distinguish the impact of different degrees of acceleration on the motion trajectory. For example, if the instantaneous acceleration is greater than the threshold, the weighting coefficient is larger, and vice versa. Assuming that a certain instantaneous acceleration is 3, its weighting coefficient can be set to 3 / 2=1.5.

[0161] The resampled signal is weighted by the weighting coefficient to obtain a weighted sampled signal, which means that a larger weight is given to the part with larger acceleration.

[0162] The frequency component of the weighted sampling signal is calculated by weighted least square method to obtain the radial velocity of the target.

[0163] The relative azimuth and elevation angles between the target's current position and the radar are used, for example, the azimuth angle is 30° and the elevation angle is 15°, to establish a conversion matrix from radial velocity to a spatial rectangular coordinate system.

[0164] The radial velocity is mapped through the transformation matrix to obtain the initial value of the three-dimensional velocity.

[0165] A state prediction equation is established to predict the three-dimensional velocity of the target at the next moment.

[0166] The prediction result of the state prediction equation and the current three-dimensional velocity initial value are weighted averaged to obtain the optimized three-dimensional velocity.

[0167] Intermediate estimates of the optimized three-dimensional velocity and instantaneous acceleration are calculated at the midpoint of the time interval.

[0168] The target position is predicted and corrected based on the intermediate estimated value to obtain the target position change.

[0169] The position change is superimposed on the initial position of the target to obtain the motion trajectory, and the motion trajectory is combined with the optimized three-dimensional velocity to obtain the motion trajectory data of the target including the velocity vector.

[0170] In this embodiment, the accuracy of the target motion trajectory is improved, especially when the target moves at a variable speed, its motion details can be captured more accurately; the anti-interference ability of the radar system is improved, and the non-uniform sampling method can effectively suppress the influence of noise and clutter; the utilization of radar resources is optimized, and the non-uniform sampling can reduce the amount of data processing and improve the real-time performance of the radar system.

[0171] In an optional implementation, the three-dimensional coordinate data and the motion trajectory data are input into a dual-stream feature fusion network, and feature extraction is performed through a position branch and a motion branch to obtain target fusion features, including:

[0172] Convert the three-dimensional coordinate data into a voxel grid according to a point cloud format, perform density clustering on the voxel grid to obtain a spatial point cluster, calculate the covariance matrix of each voxel in the spatial point cluster, extract the eigenvalue of the covariance matrix, construct a geometric descriptor sequence, construct a matching graph for the geometric descriptor sequence according to a spatial neighborhood relationship, and extract spatial domain features based on the matching graph;

[0173] Projecting the motion trajectory data onto the spherical harmonic function space to obtain spherical harmonic coefficients, extracting the time series variation sequence of the spherical harmonic coefficients, constructing a Markov jump model based on the time series variation sequence, calculating the corresponding state transition probability matrix, and extracting time domain motion features based on the steady-state distribution of the state transition probability matrix;

[0174] Constructing an affinity matrix of the spatial domain features and the temporal motion features, decomposing the affinity matrix using a spectral clustering method to obtain a feature subspace, projecting the spatial domain features and the temporal motion features onto the feature subspace to obtain a projection feature, calculating a geodesic distance matrix of the projection feature, constructing a feature manifold graph, performing a random walk on the feature manifold graph to obtain an access frequency sequence, and constructing a feature grouping matrix based on the access frequency sequence;

[0175] Based on the feature grouping matrix, hierarchical grouping is performed on the projection features to obtain feature groups, the local covariance structure of the feature groups is calculated, the main direction vectors of the local covariance structure are extracted, an orthogonal transformation matrix is ​​constructed, and the feature groups are mapped to an orthogonal space through the orthogonal transformation matrix to obtain decoupled features;

[0176] Graph Fourier transform is performed on the decoupled features to obtain frequency domain features, a multi-scale feature pyramid is constructed based on the frequency domain features, the multi-scale feature pyramid is processed by a recursive graph neural network, hierarchical representation features are extracted, and the target fusion features are obtained by combining adaptive aggregation.

[0177] The spherical harmonic coefficients specifically refer to a set of coefficients obtained by expanding a function in three-dimensional space (such as motion trajectory data) through a spherical harmonic function. Spherical harmonics are a set of orthogonal basis functions defined on a spherical coordinate system, which are usually used to describe the direction and frequency components of three-dimensional data. When processing three-dimensional trajectories, after projecting the data into the spherical harmonic function space, the spherical harmonic coefficients become the characteristic expression of the data in different directions and frequencies. The time series changes of the spherical harmonic coefficients can reveal the characteristics of the trajectory changing over time.

[0178] The Markov jump model specifically refers to a random process model used to describe the process of a system jumping between different states. The Markov jump model assumes that the current state depends only on the previous state (Markov property) and uses a state transition probability matrix to describe the probability of transitioning from one state to another. The model can generate a state transition probability matrix by tracking the transition of the system between various states and analyze its steady-state distribution. In motion analysis, the Markov jump model can be used to analyze the state transition characteristics of the trajectory, thereby extracting motion features in the time domain.

[0179] The affinity matrix specifically refers to a matrix that represents the similarity between different features or data points. In the graph structure, each element represents the "affinity" or similarity between two features or data points, and the matrix is ​​usually a symmetric matrix. The construction method of the affinity matrix depends on the specific data features, such as similarity calculation based on the spatial domain and the time domain. The affinity matrix is ​​used in algorithms such as spectral clustering to identify the intrinsic relationship between features. By decomposing the matrix, the implicit structural information of the data is extracted and converted into the feature subspace for analysis.

[0180] First, the acquired three-dimensional coordinate data is converted into a voxel grid in point cloud format. For example, a point cloud data containing 10,000 points is converted into a 10×10×10 voxel grid. Then, density clustering is performed on the voxel grid to divide adjacent and high-density voxels into a spatial point cluster. Assume that 5 spatial point clusters are obtained through density clustering. Next, the covariance matrix of the points within each voxel in each spatial point cluster is calculated, and the eigenvalues ​​of the covariance matrix are extracted. For example, there are 50 points in a voxel, and a 3×3 covariance matrix is ​​calculated, and its three eigenvalues ​​are extracted. These eigenvalues ​​are constructed into a sequence of geometric descriptors. According to the neighborhood relationship between voxels in the spatial point cluster (for example, shared faces, edges, or vertices), a matching graph is constructed. Based on the matching graph, spatial domain features are extracted using methods such as graph convolutional networks.

[0181] At the same time, the target's motion trajectory data is projected into the spherical harmonic function space to obtain the spherical harmonic coefficients. For example, a motion trajectory containing 100 time steps is projected into the 5th-order spherical harmonic function space to obtain a 5×5 spherical harmonic coefficient matrix. The time series of spherical harmonic coefficient changes is extracted. Based on the time series of changes, a Markov jump model is constructed to calculate the corresponding state transition probability matrix. For example, the transition probability from one spherical harmonic coefficient state to another spherical harmonic coefficient state can be calculated. Based on the steady-state distribution of the state transition probability matrix, the time domain motion features are extracted.

[0182] Next, construct an affinity matrix of spatial domain features and temporal motion features. For example, you can use a Gaussian kernel function to calculate the similarity between spatial domain features and temporal motion features to construct an affinity matrix. Use the spectral clustering method to decompose the affinity matrix to obtain a feature subspace. Project the spatial domain features and temporal motion features into the feature subspace to obtain the projected features. Calculate the geodesic distance matrix of the projected features and construct a feature manifold graph. Perform a random walk on the feature manifold graph to obtain an access frequency sequence. For example, you can start a random walk 1,000 times from each node and count the number of visits to each node. Construct a feature grouping matrix based on the access frequency sequence.

[0183] Based on the feature grouping matrix, hierarchical grouping is performed on the projected features to obtain feature groups. The local covariance structure of the feature group is calculated, the main direction vector of the local covariance structure is extracted, and the orthogonal transformation matrix is ​​constructed. The feature group is mapped to the orthogonal space through the orthogonal transformation matrix to obtain the decoupled features.

[0184] Perform graph Fourier transform on the decoupled features to obtain frequency domain features. Construct a multi-scale feature pyramid based on the frequency domain features. For example, you can construct a feature pyramid of three scales. Process the multi-scale feature pyramid through a recursive graph neural network to extract hierarchical representation features. Combined with adaptive aggregation, the final target fusion feature is obtained.

[0185] In this embodiment, by fusing spatial geometric features and temporal motion features and using multi-scale feature pyramids and recursive graph neural networks to extract hierarchical representation features, the target characteristics can be described more comprehensively, thereby improving the recognition accuracy; it has strong robustness to interference factors such as noise and occlusion because it utilizes highly stable features such as the covariance matrix of spatial point clusters and the spherical harmonic coefficients of motion trajectories; through feature subspace projection and hierarchical grouping, the feature dimension and computational complexity are reduced, thereby improving the computational efficiency.

[0186] In an optional implementation, human feature parameters are extracted according to the target fusion features, and the human feature parameters include scattering cross-sectional area parameters and motion feature parameters; the scattering cross-sectional area parameters are matched with a preset human scattering cross-sectional area template to obtain a first matching coefficient, and the motion feature parameters are matched with a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold, the target is determined to be a human target; a recursive state estimation model is used to perform time-series tracking on the human target to obtain a real-time position and motion state of the human target, and the number of people in the elevator car is calculated; when the number of people is zero and continues for a preset time, an unmanned signal is sent to the elevator control system to make the elevator enter an energy-saving standby mode, including:

[0187] Extracting spectrum distribution features from the target fusion features to obtain human body scattering cross-sectional area parameters, extracting trajectory continuity features to obtain human body motion feature parameters; performing wavelet decomposition on the human body scattering cross-sectional area parameters to obtain multi-scale spectrum coefficients, performing cross-correlation operation on the multi-scale spectrum coefficients and a preset human body scattering cross-sectional area template to obtain a first matching coefficient; constructing a phase space trajectory for the human body motion feature parameters, performing dynamic time warping calculation on the phase space trajectory and a preset human body motion feature template to obtain a second matching coefficient;

[0188] Based on the first matching coefficient and the second matching coefficient, a two-dimensional feature vector is constructed, and the two-dimensional feature vector is input into a preset support vector classifier to determine the target category; when the support vector classifier outputs the human target category, a state vector including position coordinates, velocity vector, and acceleration vector is constructed to initialize the state distribution of the particle swarm;

[0189] A recursive state estimation model is constructed, the recursive state estimation model comprising: constructing a state transfer equation based on the state vector, discretizing the state transfer equation to obtain a state prediction model; extracting a target scattering intensity sequence from the multi-scale spectrum coefficient, constructing an observation equation using the target scattering intensity sequence, and using the observation equation as a likelihood function of a particle weight; using the state prediction model to perform state prediction on a particle swarm, and calculating the weight of each particle based on the observation equation; performing importance resampling on the particle swarm to obtain an updated particle distribution, and calculating a weighted mean of the particle distribution as a target state estimation value;

[0190] The target state estimation values ​​of the continuous time series are combined into a target state sequence, and the trajectory of the target state sequence is predicted to obtain the target motion trajectory. Based on the target motion trajectory, the real-time number of people in the car is calculated; when the real-time number of people is zero and the duration exceeds the preset vacancy time, an unmanned signal is sent to the elevator control system to trigger the energy-saving standby mode.

[0191] First, the fusion features of the target are extracted from the sensor data such as radar echo. These fusion features include spectrum distribution features and trajectory continuity features. The spectrum distribution features reflect the scattering characteristics of the target, such as information such as the size and shape of the target. The trajectory continuity features describe the movement law of the target in the time dimension, such as information such as the speed and acceleration of the target.

[0192] Extract spectrum distribution features: Perform spectrum analysis on the radar echo signal to obtain the spectrum distribution of the target. For example, perform fast Fourier transform on a radar echo signal with a sampling frequency of 10kHz to obtain the energy distribution of the target at different frequencies. Use the spectrum distribution as the human body scattering cross-sectional area parameter.

[0193] Extract trajectory continuity features: Analyze the target's motion trajectory and calculate the target's position, speed, acceleration and other information at different times. For example, assuming that the target's position coordinates at three times are (1, 1), (2, 2), and (3, 3), the target's speed can be calculated to be (1, 1) units / second and its acceleration to be (0, 0) units / second². The trajectory continuity feature is used as a human motion feature parameter.

[0194] Then, the extracted human body scattering cross-sectional area parameters are matched with the preset human body scattering cross-sectional area template. The preset template can be obtained by statistically analyzing the scattering cross-sectional area parameters of a large number of human targets. The matching degree calculation can be performed by wavelet decomposition and cross-correlation operation. For example, the extracted spectrum distribution is subjected to wavelet decomposition to obtain multi-scale spectrum coefficients. The multi-scale spectrum coefficients are cross-correlated with the preset template to obtain a first matching coefficient. Assume that the calculated first matching coefficient is 0.9.

[0195] At the same time, the extracted human motion feature parameters are matched with the preset human motion feature template. The preset template can be obtained by statistically analyzing the motion feature parameters of a large number of human targets. The matching degree calculation can be performed using methods such as phase space trajectory and dynamic time warping. For example, the extracted trajectory continuity features are used to construct a phase space trajectory, and the phase space trajectory is dynamically time warped with the preset template to obtain a second matching coefficient. Assume that the calculated second matching coefficient is 0.85.

[0196] Next, the first matching coefficient and the second matching coefficient are constructed into a two-dimensional feature vector (0.9, 0.85). The two-dimensional feature vector is input into a preset support vector classifier for target category determination. Assuming the preset threshold is 0.8, since both the first matching coefficient and the second matching coefficient exceed the preset threshold, the support vector classifier outputs the human target category.

[0197] Then, a state vector containing position coordinates, velocity vector, and acceleration vector is constructed. For example, assuming that the initial position coordinates of the target are (1, 1), the velocity vector is (1, 1), and the acceleration vector is (0, 0), then the state vector is (1, 1, 1, 1, 0, 0). And initialize the state distribution of the particle swarm, for example, randomly generate 1000 particles, and the state vector of each particle is randomly distributed around the initial state vector.

[0198] A recursive state estimation model is constructed to perform time-series tracking of human targets. The model includes a state prediction model and an observation equation. The state prediction model constructs a state transfer equation based on the state vector and discretizes it. The observation equation is constructed using the target scattering intensity sequence extracted from the multi-scale spectral coefficients and used as the likelihood function of the particle weight. For example, assuming that the scattering intensity of the target at time t is 10, the observation equation can be expressed as the square of the difference between the scattering intensity of the particle at time t and 10.

[0199] The state prediction model is used to predict the state of the particle swarm, and the weight of each particle is calculated based on the observation equation. Importance resampling is performed on the particle swarm to obtain the updated particle distribution, and the weighted mean of the particle distribution is calculated as the target state estimate.

[0200] The target state estimation values ​​of the continuous time series are combined into a target state sequence, and the trajectory of the target state sequence is predicted to obtain the target motion trajectory. Based on the target motion trajectory, the real-time number of people in the car is calculated. For example, if at time t, the target motion trajectory shows that there are 2 people in the car, then the real-time number of people is 2.

[0201] When the real-time number of people is zero and the duration exceeds the preset vacancy time (for example, 3 minutes), an unoccupied signal is sent to the elevator control system to trigger the energy-saving standby mode.

[0202] In this embodiment, by fusing the scattering cross-sectional area parameters and the motion feature parameters and combining them with a support vector classifier, it is possible to effectively distinguish between human targets and non-human targets and improve the accuracy of human target detection. By adopting a recursive state estimation model to perform time-series tracking of human targets, it is possible to effectively cope with complex scenarios such as target occlusion and lighting changes and improve the robustness of human target tracking. By real-time monitoring of the number of people in the car and automatically triggering the energy-saving standby mode when no one is in the car, it is possible to effectively reduce the energy consumption of the elevator and achieve the goal of energy conservation and emission reduction.

[0203] In another aspect of the embodiment of the present application,

[0204] First, a 60GHz millimeter-wave radar sensor is installed on the top of the elevator car. Its transmitting antenna array consists of 16 transmitting units and the receiving antenna array consists of 32 receiving units. The multi-antenna array configuration achieves high-precision angular resolution, enabling the radar to have high resolution capabilities in the three spatial dimensions of distance, azimuth, and pitch angle. The radar uses frequency-modulated continuous wave technology, with a signal bandwidth of 4GHz and a distance resolution of up to 4 cm.

[0205] During the signal acquisition phase, the radar scans every 20 milliseconds to obtain a point cloud data frame. Considering the vibration of the elevator during operation and the interference of metal wall reflection, the system uses an adaptive threshold CFAR algorithm to suppress clutter. In view of the multipath effect of the elevator metal wall, space-time adaptive processing technology is introduced to effectively filter out false targets generated by fixed reflectors by analyzing the space-time characteristics of the echo signal.

[0206] In the point cloud data preprocessing stage, spatial clustering segmentation is first performed to classify point cloud clusters belonging to the same human target. The density clustering algorithm is used to classify spatially adjacent point clouds into one category by setting the minimum number of points and the maximum distance threshold. At the same time, the Kalman filter is combined for target tracking to establish the temporal association between point cloud clusters and form a continuous trajectory sequence.

[0207] In the human posture analysis stage, the system uses a deep learning network to extract the time-series point cloud features. The network contains a spatiotemporal convolution module and an attention mechanism, which can process the features of the spatial and temporal dimensions at the same time. The input is a 20-frame point cloud sequence, and the joint point position and motion trajectory features are extracted. The network outputs the three-dimensional coordinates of 25 key points of the human body to realize the recognition of basic postures such as standing, walking, and squatting.

[0208] For special application scenarios of elevators, the system has established a posture anomaly detection model. By analyzing the speed and acceleration of key point position changes and combining human kinematic constraints, it can identify abnormal conditions such as falls and fainting. The model uses a sliding time window for real-time monitoring, with a window length of 2 seconds and a step length of 200 milliseconds. When an abnormality is detected, the system can immediately trigger an alarm.

[0209] In actual applications, taking a shopping mall elevator as an example, the system collected more than 100,000 frames of point cloud data for model training within a month. The training set was constructed by annotating 3,000 typical posture samples, including normal and abnormal behavior scenarios. In the test data, the system's basic posture recognition accuracy reached 95%, the abnormal behavior detection accuracy reached 90%, and the average processing delay was less than 100 milliseconds.

[0210] The system also optimizes the operating conditions of the elevator. It dynamically adjusts the signal processing parameters at different stages, such as elevator start-up, braking, and constant speed operation. It compensates for the Doppler effect caused by elevator movement by analyzing the acceleration sensor data. It also builds a structural model of the elevator car, uses ray tracing technology to analyze the multipath propagation path, and optimizes the signal processing algorithm.

[0211] In terms of anti-interference, the system uses adaptive beamforming technology to suppress electromagnetic interference. By estimating the direction of the interference source in real time, the beam zero point is formed in the spatial domain to improve the signal-to-interference ratio. For the reflection of the metal wall of the car, polarization diversity reception technology is used to suppress clutter by using the difference in polarization characteristics of the reflected signal.

[0212] The system integrates edge computing units and uses heterogeneous computing architecture to accelerate data processing. Signal processing and target detection tasks are completed on the FPGA, and deep learning reasoning is performed on the GPU to achieve millisecond-level response. All processing results are uploaded to the management platform through the elevator IoT interface, supporting remote monitoring and event backtracking analysis.

[0213] Through the above technical solutions, the system has overcome multiple technical challenges in elevator scenarios, achieved accurate perception and analysis of passenger postures, and provided strong support for elevator safety management. This solution has strong practicality and promotion value and can be widely used in various elevator scenarios.

[0214] Figure 2 FIG. 1 is a schematic diagram of the structure of an elevator personnel presence detection system based on millimeter wave radar according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0215] The first unit is used to orthogonally arrange a millimeter-wave radar array on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; the echo signals received by the positioning radar and the motion detection radar are collected, and orthogonal demodulation is performed to obtain a positioning baseband signal and a motion detection baseband signal respectively; the multipath echo in the positioning baseband signal is removed by a dynamic threshold elimination method to obtain a first target echo signal; the clutter component in the motion detection baseband signal is eliminated by an adaptive harmonic suppression algorithm to obtain a second target echo signal; the first target echo signal and the second target echo signal are respectively compensated and enhanced according to the pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal;

[0216] The second unit is used to perform a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extract the spatial position information of the target, and use an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; perform a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extract the motion information of the target, perform non-uniform sampling Doppler processing on the motion information, and obtain motion trajectory data of the target including a velocity vector; input the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, perform feature extraction through a position branch and a motion branch, and obtain a target fusion feature;

[0217] The third unit is used to extract human feature parameters according to the target fusion features, wherein the human feature parameters include scattering cross-sectional area parameters and motion feature parameters; perform matching calculation on the scattering cross-sectional area parameters and a preset human scattering cross-sectional area template to obtain a first matching coefficient, and perform matching calculation on the motion feature parameters and a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determine that the target is a human target; use a recursive state estimation model to perform time-series tracking on the human target to obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; when the number of people is zero and continues for a preset time, send an unmanned signal to the elevator control system to put the elevator into energy-saving standby mode.

[0218] According to a third aspect of the embodiments of the present invention,

[0219] An electronic device is provided, comprising:

[0220] processor;

[0221] a memory for storing processor-executable instructions;

[0222] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0223] According to a fourth aspect of the embodiments of the present invention,

[0224] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0225] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0226] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the presence of people in an elevator based on millimeter wave radar, characterized in that: include: A millimeter-wave radar array is orthogonally arranged on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; echo signals received by the positioning radar and the motion detection radar are collected, and orthogonal demodulation is performed to obtain a positioning baseband signal and a motion detection baseband signal respectively; multipath echoes in the positioning baseband signal are removed by a dynamic threshold elimination method to obtain a first target echo signal; clutter components in the motion detection baseband signal are eliminated by an adaptive harmonic suppression algorithm to obtain a second target echo signal; and compensation and enhancement processing are performed on the first target echo signal and the second target echo signal respectively according to a pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal; Performing a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extracting spatial position information of the target, and using an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; Performing a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extracting motion information of the target, performing non-uniform sampling Doppler processing on the motion information to obtain motion trajectory data of the target including a velocity vector; inputting the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, performing feature extraction through a position branch and a motion branch, and obtaining a target fusion feature; Extracting human body characteristic parameters according to the target fusion characteristics, wherein the human body characteristic parameters include scattering cross-sectional area parameters and motion characteristic parameters; performing matching calculation on the scattering cross-sectional area parameters and a preset human body scattering cross-sectional area template to obtain a first matching coefficient, and performing matching calculation on the motion characteristic parameters and a preset human body motion characteristic template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determining that the target is a human body target; Using a recursive state estimation model to track the human target in time series, obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; When the number of people is zero and continues for a preset period of time, an unmanned signal is sent to the elevator control system, causing the elevator to enter energy-saving standby mode.

2. The method according to claim 1, characterized in that: Removing multipath echoes from the positioning baseband signal by a dynamic threshold elimination method to obtain a first target echo signal includes: The positioning baseband signal is divided into a plurality of adjacent range window signals, envelope detection is performed on each of the range window signals to obtain a signal amplitude envelope, and a mean value of the signal amplitude envelope is calculated by sliding average according to a preset sliding window length to obtain a power estimation value; A constant false alarm rate detection method is adopted to calculate the background noise power mean of each of the distance window signals, and the background noise power mean is subjected to exponential smoothing to obtain the noise variance; a constant false alarm rate equation is established according to a preset false alarm probability and the noise variance; Normalizing each noise variance to obtain a weight coefficient, and multiplying the weight coefficient by a solution of a constant false alarm rate equation to obtain an adaptive threshold value; The signal with an amplitude greater than the corresponding adaptive threshold value in each distance window signal is marked as a valid target echo, and the valid target echo is subjected to neighborhood smoothing processing by adopting Gaussian weighted averaging; the smoothed distance window signals are merged by using the overlap-addition method to obtain the first target echo signal.

3. The method according to claim 1, characterized in that The clutter component in the motion detection baseband signal is eliminated by an adaptive harmonic suppression algorithm to obtain the second target echo signal including: Performing wavelet decomposition on the motion detection baseband signal to obtain multi-scale wavelet coefficients; constructing an interference feature vector based on the multi-scale wavelet coefficients, including amplitude and phase information of multiple interference components; Performing autocorrelation analysis on the motion detection baseband signal to obtain a time domain autocorrelation sequence, extracting a peak point from the time domain autocorrelation sequence, and determining a signal period; Calculating the harmonic fundamental frequency according to the signal period, determining a harmonic frequency group according to the harmonic fundamental frequency, combining the harmonic frequency group with the interference feature vector to generate an interference feature template, wherein the interference feature template includes harmonic frequency and amplitude; Constructing an adaptive notch filter group based on the interference feature template, the adaptive notch filter group consists of a plurality of cascaded second-order notch filter units, the center frequencies of the second-order notch filter units corresponding one-to-one to the harmonic frequencies in the interference feature template; dynamically adjusting the quality factor of the second-order notch filter unit according to the amplitude in the interference feature vector; The motion detection baseband signal is filtered through the adaptive notch filter group to obtain a filtered signal, and the amplitude of the filtered signal is normalized to obtain a second target echo signal.

4. The method according to claim 1, characterized in that: Adopting the adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information, the three-dimensional coordinate data of the target is obtained, including: Based on a preset number of distance spectrum windows, the distance spectrum is divided into a corresponding number of overlapping sliding windows, and in each overlapping sliding window, a preset number of data points are sampled to construct a multidimensional data matrix; singular value decomposition is performed on the multidimensional data matrix to obtain a left singular vector matrix, a singular value diagonal matrix and a right singular vector matrix; Constructing an autocorrelation matrix of a distance spectrum using the left singular vector matrix and the singular value diagonal matrix, performing eigenvalue decomposition on the autocorrelation matrix to obtain an eigenvalue sequence and an eigenvector matrix; calculating a cumulative energy ratio based on the eigenvalue sequence, and determining a signal subspace dimension according to the cumulative energy ratio; Based on the signal subspace dimension, extract the eigenvectors of the number of columns corresponding to the dimension from the eigenvector matrix to construct a signal subspace matrix; perform a conjugate transpose operation on the signal subspace matrix to obtain a transposed matrix, and calculate the product of the signal subspace matrix and the transposed matrix to obtain a projection matrix; Orthogonally projecting the spatial position information through the projection matrix to obtain a projected position matrix, calculating the correlation coefficients between adjacent windows in the position matrix to obtain a spatial correlation matrix; multiplying the amplitude product of the spatial correlation matrix and the position matrix as a peak detection function; searching for a local maximum point in the peak detection function to obtain a target candidate position; K-means clustering is performed on the target candidate positions, and the three-dimensional coordinate data of the target is determined based on the cluster center.

5. The method according to claim 1, characterized in that The motion information is subjected to non-uniform sampling Doppler processing to obtain the motion trajectory data of the target including the velocity vector, including: Calculating the real part and the imaginary part of the velocity spectrum as the orthogonal component and the in-phase component respectively; calculating the envelope and phase information based on the orthogonal component and the in-phase component; Calculate the instantaneous frequency of the target according to the phase difference between adjacent sampling points; construct a new sampling interval sequence using the inverse of the instantaneous frequency; interpolate and reconstruct the envelope and phase information according to the sampling interval sequence to obtain a resampled signal; Calculating the instantaneous acceleration according to the amplitude change rate of the resampled signal; calculating the mean and standard deviation of the instantaneous acceleration, calculating and setting the acceleration threshold according to the mean and standard deviation; constructing a weighting coefficient of the instantaneous acceleration based on the instantaneous acceleration and the acceleration threshold; The resampled signal is weighted by the weighting coefficient to obtain a weighted sampled signal; the frequency component of the weighted sampled signal is calculated by weighted least square method to obtain the radial velocity of the target; Using the relative azimuth and elevation angles of the target's current position and the radar, a conversion matrix from radial velocity to a spatial rectangular coordinate system is established; the radial velocity is mapped through the conversion matrix to obtain an initial value of the three-dimensional velocity, and a state prediction equation is established; Taking a weighted average of the prediction result of the state prediction equation and the current three-dimensional velocity initial value to obtain an optimized three-dimensional velocity, calculating an intermediate estimate of the optimized three-dimensional velocity and the instantaneous acceleration at the midpoint of the time interval, and performing a prediction correction on the target position based on the intermediate estimate to obtain a target position change; The position change is superimposed on the initial position of the target to obtain a motion trajectory, and the motion trajectory is combined with the optimized three-dimensional velocity to obtain motion trajectory data of the target including a velocity vector.

6. The method according to claim 1, characterized in that The three-dimensional coordinate data and the motion trajectory data are input into a dual-stream feature fusion network, and feature extraction is performed through the position branch and the motion branch to obtain target fusion features including: Convert the three-dimensional coordinate data into a voxel grid according to a point cloud format, perform density clustering on the voxel grid to obtain a spatial point cluster, calculate the covariance matrix of each voxel in the spatial point cluster, extract the eigenvalue of the covariance matrix, construct a geometric descriptor sequence, construct a matching graph for the geometric descriptor sequence according to a spatial neighborhood relationship, and extract spatial domain features based on the matching graph; Projecting the motion trajectory data onto the spherical harmonic function space to obtain spherical harmonic coefficients, extracting the time series variation sequence of the spherical harmonic coefficients, constructing a Markov jump model based on the time series variation sequence, calculating the corresponding state transition probability matrix, and extracting time domain motion features based on the steady-state distribution of the state transition probability matrix; Constructing an affinity matrix of the spatial domain features and the temporal motion features, decomposing the affinity matrix using a spectral clustering method to obtain a feature subspace, projecting the spatial domain features and the temporal motion features onto the feature subspace to obtain a projection feature, calculating a geodesic distance matrix of the projection feature, constructing a feature manifold graph, performing a random walk on the feature manifold graph to obtain an access frequency sequence, and constructing a feature grouping matrix based on the access frequency sequence; Based on the feature grouping matrix, hierarchical grouping is performed on the projection features to obtain feature groups, the local covariance structure of the feature groups is calculated, the main direction vectors of the local covariance structure are extracted, an orthogonal transformation matrix is ​​constructed, and the feature groups are mapped to an orthogonal space through the orthogonal transformation matrix to obtain decoupled features; Graph Fourier transform is performed on the decoupled features to obtain frequency domain features, a multi-scale feature pyramid is constructed based on the frequency domain features, the multi-scale feature pyramid is processed by a recursive graph neural network, hierarchical representation features are extracted, and the target fusion features are obtained by combining adaptive aggregation.

7. The method according to claim 1, characterized in that Extracting human body characteristic parameters according to the target fusion characteristics, wherein the human body characteristic parameters include scattering cross-sectional area parameters and motion characteristic parameters; performing matching calculation on the scattering cross-sectional area parameters and a preset human body scattering cross-sectional area template to obtain a first matching coefficient, and performing matching calculation on the motion characteristic parameters and a preset human body motion characteristic template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determining that the target is a human body target; Using a recursive state estimation model to track the human target in time series, obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; When the number of people is zero and continues for a preset time, a no-person signal is sent to the elevator control system, and the elevator enters energy-saving standby mode, including: Extracting spectrum distribution features from the target fusion features to obtain human body scattering cross-sectional area parameters, extracting trajectory continuity features to obtain human body motion feature parameters; performing wavelet decomposition on the human body scattering cross-sectional area parameters to obtain multi-scale spectrum coefficients, performing cross-correlation operation on the multi-scale spectrum coefficients and a preset human body scattering cross-sectional area template to obtain a first matching coefficient; constructing a phase space trajectory for the human body motion feature parameters, performing dynamic time warping calculation on the phase space trajectory and a preset human body motion feature template to obtain a second matching coefficient; Based on the first matching coefficient and the second matching coefficient, a two-dimensional feature vector is constructed, and the two-dimensional feature vector is input into a preset support vector classifier to determine the target category; when the support vector classifier outputs the human target category, a state vector including position coordinates, velocity vector, and acceleration vector is constructed to initialize the state distribution of the particle swarm; A recursive state estimation model is constructed, the recursive state estimation model comprising: constructing a state transfer equation based on the state vector, discretizing the state transfer equation to obtain a state prediction model; extracting a target scattering intensity sequence from the multi-scale spectrum coefficient, constructing an observation equation using the target scattering intensity sequence, and using the observation equation as a likelihood function of a particle weight; using the state prediction model to perform state prediction on a particle swarm, and calculating the weight of each particle based on the observation equation; performing importance resampling on the particle swarm to obtain an updated particle distribution, and calculating a weighted mean of the particle distribution as a target state estimation value; The target state estimation values ​​of the continuous time series are combined into a target state sequence, and the trajectory of the target state sequence is predicted to obtain the target motion trajectory. Based on the target motion trajectory, the real-time number of people in the car is calculated; when the real-time number of people is zero and the duration exceeds the preset vacancy time, an unmanned signal is sent to the elevator control system to trigger the energy-saving standby mode.

8. An elevator personnel presence detection system based on millimeter wave radar, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to orthogonally arrange a millimeter-wave radar array on the top of the elevator car, wherein the millimeter-wave radar array includes a positioning radar and a motion detection radar, wherein the positioning radar is arranged at the center of the top of the elevator car, and the motion detection radar is arranged at the edge of the top of the elevator car; the echo signals received by the positioning radar and the motion detection radar are collected, and orthogonal demodulation is performed to obtain a positioning baseband signal and a motion detection baseband signal respectively; the multipath echo in the positioning baseband signal is removed by a dynamic threshold elimination method to obtain a first target echo signal; the clutter component in the motion detection baseband signal is eliminated by an adaptive harmonic suppression algorithm to obtain a second target echo signal; the first target echo signal and the second target echo signal are respectively compensated and enhanced according to the pre-acquired car geometry structure to obtain a first enhanced signal and a second enhanced signal; The second unit is used to perform a fast Fourier transform in the distance dimension on the first enhanced signal to obtain a distance spectrum, extract the spatial position information of the target, and use an adaptive spatial spectrum decomposition positioning algorithm to accurately position the spatial position information to obtain three-dimensional coordinate data of the target; Performing a fast Fourier transform in the velocity dimension on the second enhanced signal to obtain a velocity spectrum, extracting motion information of the target, performing non-uniform sampling Doppler processing on the motion information to obtain motion trajectory data of the target including a velocity vector; inputting the three-dimensional coordinate data and the motion trajectory data into a dual-stream feature fusion network, performing feature extraction through a position branch and a motion branch, and obtaining a target fusion feature; The third unit is used to extract human feature parameters according to the target fusion feature, wherein the human feature parameters include a scattering cross-sectional area parameter and a motion feature parameter; perform matching calculation on the scattering cross-sectional area parameter and a preset human scattering cross-sectional area template to obtain a first matching coefficient, and perform matching calculation on the motion feature parameter and a preset human motion feature template to obtain a second matching coefficient; when both the first matching coefficient and the second matching coefficient exceed a preset threshold value, determine that the target is a human target; Using a recursive state estimation model to track the human target in time series, obtain the real-time position and motion state of the human target, and calculate the number of people in the elevator car; When the number of people is zero and continues for a preset period of time, an unmanned signal is sent to the elevator control system, causing the elevator to enter energy-saving standby mode.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images