Target detection method and system based on millimeter wave radar

Abstract

The invention discloses a target detection method and system based on a millimeter wave radar. The target detection method and system are used for realizing accurate detection, positioning and dynamic and static recognition of multiple targets under a complex background. According to the method, a distance-Doppler spectrogram is generated through the technical means of sliding window construction, spectral analysis, clutter suppression and the like, and candidate target points are detected by adopting an SO-CFAR algorithm. Then, determining a target position through high-resolution direction estimation and coordinate transformation, performing spatial clustering in combination with a density-based DBSCAN algorithm, and extracting a target geometric center and a bounding box; in the aspect of target tracking, Kalman filtering is used for predicting and updating the position and speed of the target, and a beam forming technology is used for enhancing a target signal, so that the target recognition stability is improved. And finally, the system performs robust dynamic and static state recognition on the target through a dynamic and static judgment module, so that high precision and robustness of the target detection process are ensured. The method can effectively cope with static background interference and dynamic target changes, and is suitable for target detection and tracking in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and intelligent sensing, specifically a target detection method and system based on millimeter-wave radar. Background Technology

[0002] With the widespread application of millimeter-wave radar in target detection, its advantages in stability and penetration under all-weather and low-light conditions have become increasingly prominent, making it an important sensing means for continuous target monitoring and behavior analysis. Currently, target detection technologies mainly include visual recognition, lidar point cloud clustering, and millimeter-wave radar reflection feature processing.

[0003] While visual methods offer good resolution, their reliability is insufficient in complex lighting, occlusion, or privacy-protected scenarios. LiDAR, despite its high-precision spatial positioning capabilities, suffers from high cost, high power consumption, and susceptibility to environmental interference. In contrast, millimeter-wave radar, with its low power consumption, high robustness, and ability to penetrate harsh conditions such as rain and fog, has become an ideal choice for indoor-outdoor fusion sensing.

[0004] However, existing multi-target detection methods based on millimeter-wave radar still face several technical challenges: (1) Strong DC components caused by static background and ground reflection can easily mask weak moving target signals, affecting the accuracy of Doppler estimation; (2) Multi-channel phased array information is not fully utilized, and spatial resolution is limited, resulting in a decrease in target clustering and positioning accuracy; (3) It is difficult to stably judge the state of dynamic and static targets, and there are problems such as short-term misjudgment and frequent label switching.

[0005] Therefore, there is an urgent need to propose a millimeter-wave radar sensing method that combines frequency domain clutter suppression, high-resolution direction estimation, and motion state discrimination mechanism to achieve robust target detection and accurate identification in complex environments. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a target detection method and system based on millimeter-wave radar, which can be used to achieve accurate detection, localization and dynamic and static identification of multiple targets in complex backgrounds.

[0007] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0008] A target detection method based on millimeter-wave radar includes the following steps:

[0009] S1. Raw signal acquisition and sliding window construction: By controlling the millimeter-wave radar to transmit linear frequency modulated continuous wave signals and receiving echo signals from multiple channels, a four-dimensional complex data matrix is ​​generated after mixing, filtering and analog-to-digital conversion. The system introduces a sliding window mechanism in the time dimension, selecting specified chirp pulses from each frame at fixed steps to construct a three-dimensional complex data cube.

[0010] S2. Range-Doppler Spectrum Generation and Static Clutter Suppression: For the data cube within the sliding window, firstly, window function weighting is performed along the range dimension and a fast Fourier transform is executed in the range dimension; then, mean removal is performed on the transformed data in the Doppler dimension to suppress DC components and static clutter; then, window function weighting is performed on the data along the Doppler dimension and a fast Fourier transform is executed, finally generating a two-dimensional range-Doppler spectrum containing motion information.

[0011] S3. Candidate Target Point Detection: In the generated distance-Doppler spectrum, the Smallest-Of-CFAR algorithm is used to construct a symmetrical one-dimensional sliding detection window centered on each unit to be detected in the Doppler direction. The average power values ​​of the training units on the left and right sides of the window are calculated respectively, and the smaller value is taken as the background noise power estimate. The detection threshold is calculated according to the set false alarm rate, and candidate target points exceeding the threshold are extracted.

[0012] S4. Direction Estimation and Coordinate Transformation: For each candidate target point, the complex echo signal in each receiving channel is extracted from the original data tensor. A phase compensation factor is constructed based on the Doppler characteristics between channels for coherent correction. Then, a high-resolution direction estimation algorithm is used to search for the angle, obtain the target direction angle, and combine it with its corresponding radial distance to convert it into a planar coordinate representation.

[0013] S5. Spatial Clustering and Boundary Extraction: Input the two-dimensional planar coordinates of all candidate target points into the density-based clustering algorithm DBSCAN to obtain target clusters. Identify target clusters by setting the neighborhood radius and the minimum number of neighbor points and remove isolated noise points. Further extract the geometric center and two-dimensional bounding box of each target cluster to determine the true target.

[0014] S6. Target State Estimation and Tracking Update: Establish a state vector containing two-dimensional position and velocity for each target cluster, and use the Kalman filter method to predict and update the position and velocity of the target; maintain the target trajectory in a continuous sliding window, and stop tracking if it is not associated after the loss threshold is exceeded.

[0015] S7. Beamforming and Signal Enhancement: Based on the target's state prediction results and direction estimation angle, select array observation data for the corresponding region, construct a steering vector, and perform beamforming operation using the minimum variance distortion-free response criterion to enhance the target signal response and suppress sidelobe interference.

[0016] S8. Dynamic and static state discrimination and label update: Based on beam-enhanced images, the area of ​​the main connected region of the target is extracted and compared with a set threshold to determine the dynamic and static state of the target; combined with the discrimination results of multiple consecutive frames, the target state label is smoothly updated using a majority voting strategy.

[0017] The present invention also provides a millimeter-wave radar target detection system, comprising:

[0018] Signal acquisition module, sliding window processing module, spectrum analysis and clutter suppression module, target point detection module, two-dimensional positioning module, spatial clustering module, state tracking module, beamforming module, and state discrimination module;

[0019] The signal acquisition module generates a four-dimensional complex data matrix;

[0020] The sliding window processing module constructs a three-dimensional data cube;

[0021] The spectrum analysis and clutter suppression module outputs a distance-Doppler spectrum.

[0022] The target point detection module performs SO-CFAR detection;

[0023] The two-dimensional positioning module realizes target orientation angle estimation and coordinate transformation;

[0024] The spatial clustering module outputs the target geometric center and bounding box;

[0025] The state tracking module updates the target trajectory using a Kalman filter;

[0026] The beamforming module improves the target signal-to-noise ratio;

[0027] The state discrimination module combines the area of ​​the connected region with the majority voting mechanism to output dynamic and static state labels.

[0028] This invention has the following characteristics and beneficial effects:

[0029] High robust background suppression capability: This invention effectively suppresses static background echoes and DC clutter by means removal within a sliding window and windowing function processing, thereby improving the response capability of weak moving targets.

[0030] Fusion of 2D spectrum construction and spatial resolution: Motion features are extracted using distance-Doppler spectra and combined with high-resolution orientation estimation methods to complete 2D planar localization, solving the problem of insufficient spatial resolution of traditional methods.

[0031] Adaptive target detection and clustering capabilities: The SO-CFAR detection strategy has a higher false alarm suppression capability in multi-target clutter environments, and the density-based DBSCAN algorithm automatically identifies real target areas and eliminates outliers.

[0032] Target state dynamic tracking and enhanced representation: The target state trajectory is predicted by Kalman filtering and enhanced by beamforming to improve the main lobe of the target signal, thereby improving the stability and continuity of target recognition.

[0033] Dynamic and static stability recognition mechanism: By enhancing the analysis of the area of ​​connected regions in the image and introducing a multi-frame majority voting strategy, robust discrimination of the target's dynamic and static state is achieved, effectively avoiding frequent label switching and misjudgment.

[0034] The structure is clear and easy to implement in engineering: the method has a complete structure and distinct modules, which is easy to deploy on embedded platforms and radar front-end processing units, and has good engineering application prospects and promotion value. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a target detection method based on millimeter-wave radar according to the present invention.

[0036] Figure 2 This is a schematic diagram illustrating the generation of the range-Doppler spectrum and the incoherent accumulation results in this invention.

[0037] Figure 3 This is a schematic diagram of the SO-CFAR detection results in this invention.

[0038] Figure 4 This is a schematic diagram of the two-dimensional target localization and DBSCAN clustering results in this invention.

[0039] Figure 5 This is a schematic diagram illustrating the target position estimation achieved by Kalman filtering in this invention.

[0040] Figure 6 This is a schematic diagram illustrating the extraction of different target information to achieve beamforming in this invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0042] Example 1

[0043] A target detection method based on millimeter-wave radar, such as Figure 1 As shown, it includes the following steps:

[0044] Step S1: Acquisition of raw radar signal and construction of sliding window: The raw signal is obtained by transmitting a linear frequency modulated continuous wave signal and receiving multi-channel echo signals through millimeter-wave radar, and then a three-dimensional complex data cube is obtained through preprocessing.

[0045] In this embodiment, the millimeter-wave radar system uses frequency-modulated continuous wave (FMCW) to acquire target echo signals. The system periodically transmits linearly frequency-modulated chirped signals, which are received by multiple receiving antennas after reflection from the target. After mixing, filtering, and analog-to-digital conversion, discrete time-domain sampled data is generated.

[0046] Specifically, the preprocessing method is as follows: the data from the entire acquisition period is organized into a four-dimensional complex matrix, with the dimensions being the number of channels N. ch The number of chirp pulses N contained in each frame c The number of distance sampling points N in each chirp s And the total number of frames N in the entire acquisition process. F ,Right now:

[0047]

[0048] Where N ch The product N of the number of transmitting antennas and the number of receiving antennas TX ×N RX The calculation yielded the result. To improve the system's responsiveness to dynamic time processes, a sliding window mechanism is introduced in the time dimension. Let the duration of each sliding window be T. w seconds, radar frame rate f r Frames per second, then each sliding window T w The number of frames contained is denoted as N. w =T w ·f r The sliding window constructs continuous time segments by advancing a fixed step size Δt (the time interval between two adjacent sliding windows, which is also the update period of the Kalman filter).

[0049] Within each sliding window, the system selects the first chirp pulse from each frame for processing, thereby concatenating the specified pulses in the original four-dimensional matrix along the time dimension to construct a new three-dimensional complex matrix:

[0050]

[0051] Among them, Xwin This represents the complex data cube within the current sliding window used for subsequent processing. This matrix serves as the input data for subsequent frequency domain transformation and target detection modules, providing fundamental support for constructing the range-Doppler spectrum.

[0052] Step S2: Two-dimensional distance-Doppler spectrum generation and static clutter suppression: The three-dimensional data cube is subjected to window function weighting and fast Fourier transform along the distance dimension. In the Doppler dimension, the transformed data is subjected to mean removal to suppress static clutter. Then, window function weighting and fast Fourier transform are performed again along the Doppler dimension to generate a two-dimensional distance-Doppler spectrum.

[0053] Specifically, after extracting data within the sliding window, the system processes the three-dimensional complex matrix X. win The signal is processed using a two-dimensional frequency domain transformation to obtain the target's distribution information in both range and Doppler dimensions. This process involves performing Fast Fourier Transform (FFT) along both the range and time dimensions, supplemented by window function weighting and static clutter suppression operations to construct a two-dimensional range-Doppler spectrum (RDM). Figure 2 The diagram shows the RDM obtained by a single antenna and the RDM obtained after incoherent accumulation by multiple antennas. The horizontal axis represents the radial distance of the target relative to the radar, and the vertical axis represents the target's velocity number. The signal-to-noise ratio is improved after incoherent accumulation.

[0054] Specifically, the system first, along the distance N from the sampling point on each receiving antenna channel. s The data is weighted using a window function (preferably a Hamming window) and then subjected to a one-dimensional FFT to obtain the spectral response of the target in each range cell. This process yields the initial frequency domain representation of each channel in the range domain.

[0055] Next, to suppress DC component and static background interference, before performing the Doppler-dimensional FFT, this invention performs a Doppler transformation on the complex data after the range-dimensional transformation along the Doppler direction (i.e., the sliding window frame sequence direction, with a length of M). w A mean removal operation is performed, calculating the complex mean of each distance cell along this dimension and subtracting the mean point-by-point from each corresponding distance cell to form an intermediate result for subsequent Doppler processing. This step primarily suppresses spectral clutter generated by stationary targets or constant scattering sources (such as walls or static reflectors), improving the frequency domain separability of moving targets. This processing is strictly limited to the current sliding window, avoiding temporal discontinuities caused by cross-window background updates, ensuring the algorithm's response sensitivity in dynamic scenes and robustness against interference in static backgrounds. In the final difference spectrum, the Doppler response of dynamic targets is highlighted, while static background interference is effectively suppressed.

[0056] Subsequently, along the sliding window frame sequence M on the data processed above. w The direction is weighted by a window function, and a Fast Fourier Transform (FFT) is performed, followed by a one-dimensional FFT to extract the Doppler frequency components, thus reflecting the target's variation characteristics in the radial velocity dimension, resulting in a two-dimensional complex spectrum for each receiving channel. This spectrum shows the energy distribution characteristics of the echo signal in the range-Doppler joint domain, where dynamic targets will exhibit significant peaks in the Doppler frequency direction.

[0057] After completing the frequency domain processing and static clutter suppression processing for each channel, the system performs incoherent power accumulation on the spectra of all channels, that is, summing the squares of the complex spectra of each channel after taking their modulo, to form a unified two-dimensional spectrum:

[0058]

[0059] Among them, F r Let F represent the FFT along the distance dimension. d Indicates the FFT along the Doppler curve. This represents the data received from the i-th channel.

[0060] The final distance – Doppler spectrum It will serve as the main input for subsequent target detection and feature extraction modules, used to identify potential moving target regions.

[0061] Step S3: Target detection and candidate region extraction based on SO-CFAR: The Smallest-Of-CFAR algorithm is used in the two-dimensional distance-Doppler spectrum to calculate the background noise estimate using a one-dimensional sliding window in the Doppler direction, and candidate target points are extracted based on the false alarm rate. Figure 3 The result of CFAR detection is shown in the target distance-Doppler spectrum.

[0062] Specifically, in obtaining the distance-Doppler spectrum R rdm Subsequently, to effectively extract potential moving targets, the system employs the SO-CFAR strategy from the constant false alarm rate (CFAR) detection algorithm for target detection. This method is particularly suitable for handling scenes with strong local clutter or non-uniform backgrounds, and can improve the performance of multi-target detection.

[0063] In its implementation, the system uses each distance-Doppler unit in the spectral graph as the detection center and constructs a symmetrical sliding window in its Doppler direction, with the window containing several training unit regions.

[0064] Specifically, in the range-Doppler map, the system constructs a one-dimensional sliding detection window along the Doppler direction, centered on each target cell. This window is symmetrically distributed on both sides of the target cell, and includes N... traintraining units and N guard One protection unit. The training unit is used to estimate the background noise level, and the protection unit is used to avoid interference from the target main lobe.

[0065] For each central unit (i,j), the system calculates the average power of the training units to its left and right, respectively, defined as:

[0066]

[0067] The background noise estimate is taken as the smaller of the two averages, i.e.:

[0068] P min =min(P left ,P right )

[0069] Based on the preset false alarm rate P fa The system calculates the threshold coefficient α, thereby obtaining the detection threshold of the unit:

[0070] T = α·P min ,in N = 2N train

[0071] If the current unit power R rdm If (i,j)>T, then the target signal is considered to exist at that position, and it is determined to be a candidate detection unit.

[0072] The system performs the above judgment operation on the entire distance-Doppler map, extracts all candidate target points that meet the conditions, and records their corresponding distances and Doppler indices. These candidate points will serve as input for subsequent spatial clustering and target trajectory estimation.

[0073] S4: Target direction angle estimation and polar coordinate transformation: Extract the multi-channel complex echo signal of the candidate target point, perform coherent correction through phase compensation factor, obtain the target direction angle using high-resolution direction estimation algorithm, and convert it into two-dimensional plane coordinates by combining radial distance.

[0074] To achieve spatial localization of the target, the system uses the complex snapshot data tensor X within the sliding window constructed in step S1. win The complex echo information of candidate target points in each channel is extracted, and their phase characteristics are finely modeled and their direction is estimated.

[0075] Specifically, the system uses the distance-Doppler coordinate index for each candidate target point from the tensor X. winThe complex response signal is extracted from all receiving channels. Since the system adopts a multi-transmit channel configuration, the signal may have an inherent phase shift in different channels. This phase shift is caused by the different beam propagation paths / time delays between the transmit channels. Therefore, the system constructs a phase compensation factor based on the transmit channel number and Doppler index, and performs phase rotation on the data of each receiving channel in the complex domain, thereby achieving coherent correction of the multi-channel data.

[0076] After completing inter-channel phase alignment, the system performs two-dimensional orientation angle estimation on candidate target points. Orientation estimation employs the MUSIC algorithm based on array signal processing, which constructs a steering vector a(θ) and, combined with the array covariance matrix R, solves for its spatial spectral function.

[0077]

[0078] Among them, E n Let θ represent the noise subspace matrix, and θ be the angle scanning variable. The system traverses the angle space to find the spectral peak position, which is the estimated direction angle corresponding to the target.

[0079] After obtaining the orientation angle, the radial distance r of the target in the range-Doppler image is combined to convert its polar coordinate information into two-dimensional Cartesian plane coordinates. The calculation method is as follows:

[0080] x = r·cos(θ), y = r·sin(θ)

[0081] The system converts the polar coordinates of all valid target points into rectangular coordinates and constructs a distribution map of candidate targets in a horizontal two-dimensional space, providing basic input for subsequent clustering and tracking.

[0082] S5: Spatial Clustering and Measurement Extraction: Input candidate target points into the DBSCAN clustering algorithm to identify target clusters and remove noise points, and extract the geometric center and two-dimensional bounding box of the clusters. Figure 4 The results of two-dimensional target localization and DBSCAN clustering achieved by combining steps S4 and S5 are shown. The horizontal and vertical axes represent the horizontal and vertical distances of the target relative to the radar, respectively. The target boundaries are delineated with red boxes, and the target number is marked at the center of the target.

[0083] After obtaining the two-dimensional location coordinates of the candidate targets, the system continues to perform spatial clustering processing on the target point set to identify multiple individual targets and eliminate isolated noise points. In this embodiment, the system uses the density-based spatial clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to perform preliminary clustering, extraction, and screening of candidate targets.

[0084] First, the detected set of candidate target points is mapped to a Cartesian coordinate system, and the position of each point is denoted as (x, y). i ,y i The system generates a point cloud distribution. Then, it uses density reachability as the clustering criterion to connect the neighborhoods of target points. Given a neighborhood radius ε and a minimum neighbor number threshold MinPts, if a point contains at least MinPts neighbor points within its ε neighborhood, it is considered a cluster core point (a sample point that satisfies the density threshold), and all points satisfying density reachability are clustered into the same cluster. Conversely, if a point neither satisfies the core condition nor belongs to the density reachability region of any core, it is considered spatial noise and discarded.

[0085] For each cluster, the system further calculates the geometric center positions of all its member points, which are used as the estimated location of the target. Assume the k-th cluster contains N... k For each valid point, the system estimates its center coordinates using an average method, as follows:

[0086]

[0087] After extracting the cluster centers, the system constructs a two-dimensional rectangular bounding box based on the minimum and maximum coordinates of all points within each cluster. This bounding box delineates the geometric distribution of the target's spatial region and provides positional constraints for subsequent beamforming guidance and target labeling. The bounding box can be used for system visualization and data output. Ultimately, each cluster outputs the following information: the candidate target number, the spatial coordinates of the cluster center, the size of the constructed bounding box, and region attributes. The system uses this information as input to the target measurement module for subsequent Kalman filter state estimation and trajectory management.

[0088] Step S6: Kalman filtering and target state estimation: Establish a state vector containing position and velocity for the cluster, use Kalman filtering to predict the target state and update the observation, and stop tracking after consecutive target loss. Figure 5 This diagram illustrates target location estimation based on Kalman filtering. Blue clusters represent target clusters obtained through clustering, while red "×" marks indicate target centers obtained through Kalman estimation. The left image shows the case where the measured and estimated values ​​match, while the right image shows the case where they do not. Subsequent target locations will use the estimated values.

[0089] After locating the spatial cluster centers of candidate targets, the system updates the states of all targets in real time within each sliding window period to further establish the temporal continuity and dynamic trajectory modeling of multiple targets. This embodiment employs a classic linear Kalman filter structure to perform state prediction and observation fusion for each candidate target, constructing its two-dimensional motion trajectory.

[0090] Specifically, the system establishes the following four-dimensional state vector for each target:

[0091]

[0092] in, The estimated spatial coordinates of the target at the current moment. It estimates the velocity components in the horizontal and depth directions.

[0093] State prediction phase:

[0094] At each sliding window time, the system first uses the state vector of the target in the previous window. The state transition process for recursive state prediction is described by the following linear model:

[0095]

[0096] P k∣k-1 =F·P k-1 ·F T +Q

[0097] in, Let P be the state transition matrix. k-1 Let Q be the state estimation covariance at the previous time step, and Q be the process noise covariance matrix. The state transition matrix is ​​in the form of:

[0098]

[0099] The definition of Δt is the same as that in S1.

[0100] Observational correlation and state update:

[0101] Using the current sliding window cluster center coordinates as the observation input, the observation vector is defined as follows:

[0102]

[0103] The measurement matrix H is defined as follows:

[0104]

[0105] By calculating the predicted state With all current measurements z k The Euclidean distance between them is used to determine whether the association threshold condition is met. If the threshold condition is met, the following state update is performed:

[0106] Kalman gain calculation:

[0107] K k =P k∣k-1 ·H T ·(H·Pk∣k-1 ·H T +R) -1

[0108] State vector update:

[0109]

[0110] Covariance update:

[0111] P k =(IK k ·H)·P k∣k-1

[0112] Where R is the measurement noise covariance matrix, which reflects the radar echo measurement error.

[0113] Status output and direction angle calculation:

[0114] After completing the filtering update, the system outputs the current posterior state estimate for each target, including its two-dimensional position. With speed

[0115] Simultaneously, based on the estimated current state of the target, the system calculates its spatial orientation angle.

[0116]

[0117] This angle is used as the azimuth guide input for subsequent directional gain enhancement modules to achieve more refined beam processing of the target spatial distribution.

[0118] If a target is not associated with an observation in multiple consecutive sliding windows, the system will accumulate its lost frame count and remove the corresponding trajectory after exceeding a set threshold; otherwise, the target trajectory will be maintained and participate in the next round of state recursion.

[0119] S7: Target-guided beamforming enhancement: Construct a guide vector based on the target orientation angle, and apply the minimum variance distortion-free response criterion to beamform the array observation data to enhance the target signal. Figure 6 The image shows the results of beamforming algorithm used to enhance target signals and achieve multi-target segmentation based on target range and velocity information. The image displays the range and velocity information of three targets.

[0120] To achieve precise signal enhancement and interference suppression in the target area, after completing Kalman filter prediction and spatial clustering matching, the system performs local adaptive beamforming for each tracked target. The core of this process lies in:

[0121] The signal power in the direction of interest is increased by using array observation data, while suppressing interference echoes from other directions.

[0122] First, based on the predicted location of the target and the geometric centers of all clusters, the Euclidean spatial distance is calculated, and a nearest neighbor strategy is used to complete the one-to-one association, that is, to determine a unique augmentation candidate region for each target. This step provides the input basis for subsequent region clipping and guide vector construction.

[0123] After matching, the system uses the clusters associated with the target as the basis for the augmentation region, locates the target's response range in the two-dimensional range-Doppler map, and extracts the corresponding complex echo signal. Since actual beamforming mainly depends on the horizontal dimension of the array, the system selects predefined horizontal sub-array channels from the full-channel data to construct a local observation snapshot matrix.

[0124]

[0125] Where, N h This represents the number of horizontal array elements, where k is the number of snapshots equal to the number of frames in the current sliding window. w X is the number of samples formed by stitching together target chirp snapshots selected from each frame along the time axis. ROI Each column represents a complex snapshot observation sample within an extraction region.

[0126] Subsequently, the system uses the estimated direction angle of the target. Construct its corresponding guide vector A loading factor δ is introduced into the covariance matrix for regularization to improve numerical stability and ensure the reliability of subsequent inversion operations. The estimated covariance matrix is ​​expressed as:

[0127]

[0128] Where, x i This represents the vector of the i-th snapshot. For N h ×N h The identity matrix.

[0129] The beamforming weight vector is calculated using the Minimum Variance Distortionless Response (MVDR) criterion:

[0130]

[0131] Apply the above weight vector to the observation matrix X ROI This allows us to obtain an enhanced response from the target in that direction:

[0132]

[0133] Result B kTo enhance the target image, the main lobe response can be significantly improved and side lobe clutter suppressed in the two-dimensional distance-Doppler plane, effectively improving the quality of subsequent target recognition and feature extraction.

[0134] S8: Dynamic and static state discrimination and result output: Extract the area of ​​the main connected region of the target based on the enhanced image after beamforming, compare it with the threshold to judge the dynamic and static state, and update the state label by majority voting in combination with the discrimination results of multiple frames.

[0135] To further improve the system's ability to recognize target states and interpret scenarios, the system performs dynamic and static state discrimination for each tracked target based on Kalman filter tracking and beamforming enhancement, and outputs its key position information and motion attributes.

[0136] The system uses beam-enhanced images as a basis, combined with the two-dimensional range-Doppler map response of each target in its direction, to construct connected regions and determine their structural characteristics. Since the Doppler dimension reflects the target's velocity information in the direction of radar beam illumination, the system uses this as a basis to identify whether the target is in a state of continuous motion.

[0137] Specifically, the system first acquires the beamforming enhancement image B pointed to by the Kalman prediction position. k The system performs amplitude normalization and regional connectivity analysis on each enhanced map. For each enhanced map, the system calculates the spatial span of all connected regions and sets a dynamic / static discrimination threshold based on the region area. This area corresponds to velocity extensibility in the frequency domain and can reflect the target motion characteristics.

[0138] If the area of ​​the main connected region is greater than the threshold T dyn If the target is dynamic, it is considered dynamic; otherwise, it is considered static.

[0139] Dynamic targets: Targets appear as extended structures in augmented maps, exhibiting significant velocity expansion or non-zero Doppler drift.

[0140] Static target: The target's response is concentrated, the main lobe region is stable, and the changes in position and velocity are minimal.

[0141] In addition, the system introduces a time consistency mechanism during the discrimination process: if a target is judged to be in the same state in multiple consecutive sliding windows, the state label is maintained; if the state changes, the previous window is used as the reference to keep the state stable and prevent short-term fluctuations from causing misjudgment.

[0142] Example 2

[0143] This embodiment provides a target detection system based on millimeter-wave radar that implements the method in Embodiment 1, including a signal acquisition module, a sliding window processing module, a spectrum analysis and clutter suppression module, a target point detection module, a two-dimensional positioning module, a spatial clustering module, a state tracking module, a beamforming module, and a state discrimination module;

[0144] The signal acquisition module generates a four-dimensional complex data matrix;

[0145] The sliding window processing module constructs a three-dimensional data cube;

[0146] The spectrum analysis and clutter suppression module outputs a distance-Doppler spectrum.

[0147] The target point detection module performs SO-CFAR detection;

[0148] The two-dimensional positioning module realizes target orientation angle estimation and coordinate transformation;

[0149] The spatial clustering module outputs the target geometric center and bounding box;

[0150] The state tracking module updates the target trajectory using a Kalman filter;

[0151] The beamforming module improves the target signal-to-noise ratio;

[0152] The state discrimination module combines the area of ​​the connected region with the majority voting mechanism to output dynamic and static state labels.

[0153] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A target detection method based on millimeter-wave radar, characterized in that, Includes the following steps: S1. The original signal is obtained by transmitting a linear frequency modulated continuous wave signal through a millimeter-wave radar and receiving multi-channel echo signals. Then, a three-dimensional complex data cube is obtained through preprocessing. S2. Perform window function weighting and fast Fourier transform on the three-dimensional data cube along the distance dimension, perform mean removal operation on the transformed data in the Doppler dimension to suppress static clutter, and then perform window function weighting and fast Fourier transform along the Doppler dimension to generate a two-dimensional distance-Doppler spectrum. S3. The SO-CFAR algorithm is used in the two-dimensional distance-Doppler spectrum to calculate the background noise estimate using a one-dimensional sliding window in the Doppler direction, and candidate target points are extracted based on the false alarm rate by setting the detection threshold. S4. Extract the multi-channel complex echo signal of the candidate target point, perform coherence correction through phase compensation factor, obtain the target direction angle by high resolution direction estimation algorithm, and convert it into two-dimensional plane coordinates by combining radial distance. S5. Input the two-dimensional plane coordinates of the candidate target points into the DBSCAN clustering algorithm to obtain target clusters, identify target clusters and remove noise points, extract the geometric center and two-dimensional bounding box of the target clusters to determine the real target; S6. Establish a state vector containing position and velocity for the target cluster, use Kalman filtering to predict and update the target's position and velocity, and calculate the target's orientation angle based on the estimated current state of the target. If targets are lost consecutively, stop tracking. S7. Construct a steering vector based on the target orientation angle, and apply the minimum variance distortionless response criterion to beamform the array observation data to enhance the target signal; S8. Based on beamforming, extract the area of ​​the target main connected region from the image, compare it with a threshold to determine the dynamic and static state, and update the state label by majority voting in combination with the multi-frame discrimination results.

2. The method according to claim 1, characterized in that, The preprocessing method is as follows: the acquired raw signal is mixed, filtered and converted from analog to digital to generate a four-dimensional complex data matrix; then a sliding window mechanism is introduced in the time dimension to select a specified chirp pulse from each frame with a fixed step size to construct a three-dimensional complex data cube.

3. The method according to claim 2, characterized in that, In step S1, each element in the four-dimensional complex data matrix is ​​identified by four indices, which correspond to the channel number, chirp pulse number, distance sampling point number, and frame number, respectively.

4. The method according to claim 2, characterized in that, The sliding window mechanism: Let the coverage time of each sliding window be T. w seconds, radar frame rate f r Frames per second, then each sliding window T w The number of frames contained is denoted as N. w =T w ·f r The sliding window has a fixed step size of Δt, and the update period of the Kalman filter is the same for Δt.

5. The method according to claim 1, characterized in that, In step S2, the mean removal operation specifically involves: calculating the complex mean of each distance cell along the Doppler dimension, and subtracting the mean from each corresponding distance cell point by point.

6. The method according to claim 1, characterized in that, In step S3, the Smallest-Of-CFAR algorithm includes: Centered on the unit to be detected, a symmetrical window containing guard units and training units is constructed in the Doppler direction. The smaller of the average power of the training units on both sides is taken as the background noise estimate P. min According to the preset false alarm rate P fa The system calculates the threshold coefficient to obtain the detection threshold of the unit: T = α·P min Where α is the threshold coefficient, which is determined by the false alarm rate P fa It is calculated from the training unit N.

7. The method according to claim 1, characterized in that, In step S4, the phase compensation factor is constructed based on the transmission channel number and Doppler index, and is used to correct the inherent phase shift of the multi-channel signal. The high-resolution orientation estimation algorithm is the MUSIC algorithm, which obtains the orientation angle θ by searching for spectral peaks using a spatial spectral function. The expression for the spatial spectral function is as follows: Among them, E n Let θ represent the noise subspace matrix, θ be the angle scanning variable and the searched direction angle, and a(θ) be the steering vector. The system traverses the angle space to find the position of the spectral peak, which is the estimated direction angle corresponding to the target.

8. The method according to claim 1, characterized in that, In step S5, DBSCAN clustering identifies density-reachable point clusters by neighborhood radius ε and minimum neighbor number MinPts, and points that do not meet the conditions are removed as noise. The geometric center coordinates of the cluster are: The two-dimensional bounding box is determined by the extreme values ​​of the geometric center coordinates of the cluster.

9. The method according to claim 1, characterized in that, In step S6 State prediction phase: At each sliding window time, first, based on the state vector of the target in the previous window... The state transition process for recursive state prediction is described by the following linear model: P k∣k-1 =F·P k-1 ·F T +Q in, Let P be the state transition matrix. k-1 Let Q be the state estimation covariance at the previous time step, and Q be the process noise covariance matrix. The state transition matrix is ​​in the form of: Observational correlation and state update: Using the current sliding window cluster center coordinates as the observation input, the observation vector is defined as follows: Define the measurement matrix H: By calculating the predicted state With all current measurements z k The Euclidean distance between them is used to determine whether the association threshold condition is met. If the threshold condition is met, the state is updated. Status output and direction angle calculation: After completing the filtering update, output the current posterior state estimate for each target, including its two-dimensional position. With speed Simultaneously, based on the estimated current state of the target, the system calculates its spatial orientation angle. If a target is not associated with an observation in multiple consecutive sliding windows, its lost frame count will be accumulated, and the corresponding trajectory will be removed after exceeding a set threshold; otherwise, the target trajectory will be maintained and participate in the next round of state recursion.

10. The method according to claim 9, characterized in that, In step S6, if the threshold condition is met, the following state update is performed: Kalman gain calculation: K k =P k∣k-1 ·H T ·(H·P k∣k-1 ·H T +R) -1 State vector update: Covariance update: P k =(I-K k ·H)·P k∣k-1 Where R is the measurement noise covariance matrix, which reflects the radar echo measurement error.

11. The method according to claim 1, characterized in that, In step S7, the beamforming weight vector calculation satisfies: Among them, R x To load the covariance matrix of the regularization factor δ, x i This represents the vector of the i-th snapshot. For N h ×N h The identity matrix.

12. The method according to claim 1, characterized in that, In step S8, the threshold comparison for determining the dynamic and static states satisfies the following condition: if the area of ​​the main connected region is greater than the threshold T... dyn If it is true, it is determined to be a dynamic target; otherwise, it is a static target. The majority voting mechanism works as follows: the label is updated when the judgment results of multiple consecutive frames are consistent; otherwise, the previous state is maintained.

13. A millimeter-wave radar target detection system based on the method of any one of claims 1-12, characterized in that, include: Signal acquisition module, sliding window processing module, spectrum analysis and clutter suppression module, target point detection module, two-dimensional positioning module, spatial clustering module, state tracking module, beamforming module, and state discrimination module; The signal acquisition module generates a four-dimensional complex data matrix; The sliding window processing module constructs a three-dimensional data cube; The spectrum analysis and clutter suppression module outputs a distance-Doppler spectrum. The target point detection module performs SO-CFAR detection; The two-dimensional positioning module realizes target orientation angle estimation and coordinate transformation; The spatial clustering module outputs the target geometric center and bounding box; The state tracking module updates the target trajectory using a Kalman filter; The beamforming module improves the target signal-to-noise ratio; The state discrimination module combines the area of ​​the connected region with the majority voting mechanism to output dynamic and static state labels.

Images