Millimeter wave radar point cloud generation method with freezing mechanism static filtering and local interpolation angle measurement

Abstract

The invention discloses a millimeter wave radar point cloud generation method with freezing mechanism static filtering and local interpolation angle measurement. According to the method, exponential moving average (EMA) static clutter suppression with a freezing mechanism is carried out on a distance-Doppler unit, and a smoothing coefficient is temporarily improved to freeze background update when a suspected human body scattering target is detected, so that a slow target is prevented from being filtered out; and calculating a coarse grid spatial spectrum for the candidate units by adopting adaptive diagonal loading two-dimensional minimum variance undistorted response (MVDR), reconstructing a fine grid spatial spectrum in a peak neighborhood based on local Kriging interpolation of a Gaussian variation function so as to obtain a high-resolution azimuth angle and a pitch angle, and finally generating a three-dimensional point cloud in combination with the distance. Compared with the prior art, the method has the advantages that clutter suppression robustness is improved and false alarm and leak detection are reduced under the complex indoor slowly-changing background, and meanwhile, the angle measurement resolution is improved on the premise that the calculated amount is controlled through local interpolation.

Description

Technical Field

[0001] This invention belongs to the field of millimeter-wave radar signal processing and three-dimensional perception technology, specifically relating to a millimeter-wave radar point cloud generation method for human motion recognition, and more particularly to a three-dimensional point cloud generation method that combines adaptive freezing mechanism static filtering with two-dimensional minimum variance distortionless response (MVDR) angle measurement based on local Kriging interpolation. Background Technology

[0002] Human motion recognition relies on high-quality point cloud data to characterize the spatial structure and temporal continuity of the human body. Millimeter-wave radar has advantages such as all-weather operation, privacy protection, and insensitivity to occlusion. It can acquire human body scattering point clouds without acquiring optical images, thus becoming an important sensing method for motion recognition.

[0003] However, existing millimeter-wave point cloud generation processes still face the following problems in indoor environments: (1) Slow background drift and slow target absorption: Indoor temperature drift, slight changes in radar attitude, and slow multipath variation can cause static clutter background to drift slowly over time; Traditional exponential moving average (EMA) / first-order low-pass background modeling, while suppressing static clutter, can easily absorb slow or continuous human body scattering into the background, resulting in sparse point clouds and broken targets. (2) False alarms / missed detections caused by detection and angle measurement coupling: Detection strategies that rely solely on single-channel or single-scale thresholds are prone to false alarms when noise fluctuates and the background is unstable, or miss detections when human body scattering is weak. (3) Contradiction between angle measurement resolution and computational complexity: High-resolution angle measurement algorithms such as two-dimensional MVDR need to search on fine angle grids, resulting in high computational complexity; if coarse grids are used instead, it will lead to insufficient angle resolution, point cloud jitter, and spatial spectrum peak position jumps, which is not conducive to the spatiotemporal continuity modeling of action recognition.

[0004] Therefore, there is an urgent need for a point cloud generation method that is robust under gradually changing background conditions, can suppress slow target absorption, and improves the two-dimensional angular resolution and point cloud continuity with controllable computational load. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a millimeter-wave radar point cloud generation method that combines adaptive freezing mechanism EMA static filtering and local Kriging interpolation 2D MVDR angle measurement, aiming to provide stable, continuous, and high-resolution 3D point cloud data for human motion recognition.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for generating millimeter-wave radar point clouds by combining adaptive filtering and interpolation angle estimation, comprising the following steps:

[0007] Step S1: Receive the raw sampled data from the millimeter-wave radar echo and organize it frame-by-frame into multi-channel data for a multiple-transmit, multiple-receiver (MMR) system. Perform a range-to-fast Fourier transform (FFT) on each chirp to obtain the range spectrum, and perform an FFT on the same range gate in the slow time dimension to obtain the range-Doppler spectrum. This forms a data structure containing... Distance of virtual array element channels - Doppler complex data cube;

[0008] Step S2: Perform EMA static filtering with adaptive freezing mechanism on the multi-channel range-Doppler complex data cube to obtain background estimation and filtering residual. When a range-Doppler cell is determined to be a suspected human body scattering cell, the smoothing coefficient is increased to freeze the background update of the cell.

[0009] Step S3: Perform multi-channel energy aggregation on the filter residuals from step S2 to construct a range-Doppler two-dimensional power map, and perform two-dimensional constant false alarm rate (CFAR) detection on the two-dimensional power map to obtain a set of candidate units for human body scattering targets;

[0010] Step S4: For the set of candidate units for human body scattering targets, construct the array covariance matrix, and use two-dimensional MVDR spatial spectrum estimation with adaptive diagonal loading to calculate the coarse spatial spectrum on the sparse angle grid.

[0011] Step S5: Based on the coarse spatial spectrum from step S4, determine the energy peak region. Within this region, use local Kriging interpolation to perform fine-grid reconstruction of the two-dimensional MVDR spatial spectrum to obtain the target azimuth and elevation angle estimates.

[0012] Step S6: Combine the target distance obtained in step S1 with the azimuth and pitch angles obtained in step S5 to complete the mapping from distance-azimuth-pitch angle to the Cartesian coordinate system and output a 3D point cloud.

[0013] Furthermore, in S2, it is necessary to... Frame in Doppler Index Distance Index Observations at the location Establish an observation model:

[0014]

[0015] in The background item that drifts slowly over time. For fast variables such as human body scattering, This represents the noise term. The background is recursively estimated using EMA:

[0016]

[0017] in It is the smoothing coefficient. For background estimation, the filtered residual is:

[0018]

[0019] Basic smoothing coefficient By frame period With background expected time constant Mapping:

[0020]

[0021] To prevent slow / persistent targets from being absorbed into the background, an adaptive freeze mechanism is introduced: for each distance gate Robust background baseline is calculated on the Doppler band after removing the DC Doppler neighborhood. Calculate the power for each unit. ,when:

[0022] Trigger freeze, ,in Otherwise take This allows for almost no background updates on suspected human scattering units, thus protecting the target while retaining minimal leakage in the frozen state to suppress long-term drift distortion.

[0023] In the freeze decision of step S2, a threshold is given. The statistical meaning of this can be interpreted under the pure noise assumption regarding the probability of false triggering: after removing the DC Doppler neighborhood, if a certain distance gate... The residual complex echo can be approximated as zero-mean complex Gaussian noise, then its power (i.e., amplitude squared) is:

[0024]

[0025] It follows an exponential distribution with a mean of 1 / 2. The median of an exponential distribution satisfies:

[0026]

[0027] Therefore, a freeze decision is equivalent to setting a power threshold:

[0028]

[0029] Under pure noise conditions, the probability of a single distance-Doppler cell falsely triggering a freeze is:

[0030]

[0031] Therefore, the upper limit of the expected single-frame false trigger probability can be determined based on this. choose satisfy:

[0032]

[0033] On the other hand, the present invention adopts a median-based approach. The robust baseline estimation can be obtained, and the false trigger probability can be further reduced under conditions such as removing DC Doppler neighborhood and cross-frame consistency constraints, so that the results given by the above derivation are interpretable and controllable for engineering parameter selection.

[0034] Furthermore, in S3, the multi-channel filtering residuals obtained in step S2 for each frame need to be processed. exist Energy aggregation is performed on each channel dimension to obtain a two-dimensional power map:

[0035]

[0036] Two-dimensional CFAR detection is performed on the two-dimensional power graph: a two-dimensional training window and a guard window are constructed for each cell to be tested, the noise level is estimated using the training window and an adaptive threshold is formed, and a set of candidate cells that pass the threshold decision is output. , as a distance-Doppler candidate point for human body scattering targets.

[0037] Furthermore, S4 targets candidate units. In its corresponding multi-channel Estimate the covariance matrix based on the snapshot vectors:

[0038]

[0039] in For the first A snapshot is taken. To enhance robustness under varying noise and finite snapshot conditions, an adaptive diagonal loading covariance matrix is ​​introduced. :

[0040]

[0041]

[0042] in As a scaling factor, It is an identity matrix.

[0043] In sparse angular grid Calculate the two-dimensional MVDR spatial spectrum:

[0044]

[0045] Obtain coarse spatial spectrum sampling set .in For array guide vector, It is the azimuth angle. The pitch angle.

[0046] Furthermore, in S5, to improve angular resolution while controlling computational load, the region of interest (ROI) is first determined based on the peak and threshold rules of the coarse spatial spectrum, and fine-mesh reconstruction is performed only within the ROI. Using ordinary Kriging interpolation, for any fine-grid angle point Estimating the spatial spectrum:

[0047]

[0048] in for nearest neighbor A coarse grid point, Obtained from the Kriging equations. The semivariogram is modeled using a Gaussian model:

[0049]

[0050] in The angular plane distance This represents the nugget value (corresponding to systematic error and spectral noise floor). The arch height (corresponding to the global variance). These are the range parameters (related to beamwidth). Performed on a fine mesh within the ROI. After reconstruction, the target azimuth is obtained by searching only within the ROI. Pitch angle .

[0051] Furthermore, S6 will target distance With angle estimation Mapping to a Cartesian coordinate system generates a point cloud.

[0052] Compared with the prior art, the present invention has at least the following beneficial effects:

[0053] (1) Suppressing slowly changing background and avoiding absorption by slow targets: Through the EMA+ adaptive freezing mechanism, stable clutter suppression is maintained under the condition of slow background drift. At the same time, the background update is frozen for suspected human scattering units, which significantly reduces the point cloud loss caused by absorption by slow / continuous targets. (2) Improved detection robustness and transferability: Multi-channel aggregation improves the detectability of human scattering. CFAR adaptively estimates noise in the two-dimensional domain, reducing false alarms and missed detections under environmental changes with fixed thresholds. (3) Achieving both high resolution of angle measurement and controllable computing power: Sparse grid MVDR provides low-cost coarse spectrum localization. Local Kriging performs fine grid reconstruction and peak search within the ROI, achieving an angular resolution similar to fine grid MVDR, while significantly reducing the computational cost of fine grid in the full-angle domain. (4) Enhanced spatial and temporal continuity of point cloud: The Gaussian variogram model improves the smoothness of the spectrum and the continuity of peak positions, reduces angle jumps and point cloud jitter, and is more suitable for the requirements of spatiotemporal continuity for human action recognition. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0055] Figure 2 This is the energy decay curve of a single person walking-stopping-walking target in Example 1;

[0056] Figure 3 This is the dual-target resolution curve in Example 2;

[0057] Figure 4 This is the dual-target resolution angle error curve in Example 2;

[0058] Figure 5 This is the dynamic trajectory tracking diagram in Example 3;

[0059] Figure 6 This is the point cloud map of a single person standing in Example 4;

[0060] Figure 7 This is the point cloud diagram of a single person in a seated posture in Example 4;

[0061] Figure 8 This is the point cloud map of a single person lying flat in Example 4; Detailed Implementation

[0062] To facilitate understanding of the technical content of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] This invention is based on a frequency-modulated continuous wave radar system, in which the radar transmits a linear frequency-modulated signal and receives the echo. It should be noted that the following embodiments are only for explaining the invention and are not intended to limit the scope of protection of the invention; all equivalent substitutions or modifications made within the spirit and principles of the invention should be covered within the scope of protection of the invention.

[0064] Example 1

[0065] To verify the protective capability of the EMA static filter with adaptive freezing mechanism described in this invention against slow / continuous targets (such as a human body standing still with only slight breathing movements) from being absorbed by the background model, this embodiment constructs a set of controllable range-Doppler (RD) sequence simulation data and compares the method of this invention with standard EMA static filtering and inter-frame difference.

[0066] Step S1: Simulation Scene Setup

[0067] The scene contains only a single pedestrian target. Two strong static background scatterers (such as walls and furniture) are also set up, with their echo energy concentrated at distance indices 10 and 20, to simulate strong clutter backgrounds in real indoor scenes.

[0068] The pedestrian target's actions are segmented by time as follows:

[0069] 0–2 s: The target approaches the radar from a distance (movement state);

[0070] 2–6 s: The target stands still in place, but retains very slight breathing movements (approximately still).

[0071] 6–8 s: The target turns away (in motion).

[0072] The above sequence of actions deliberately includes long periods of stillness to create a typical failure scenario where the target is absorbed into the background, which is common in standard EMAs.

[0073] Step S2: RD Sequence Generation Method

[0074] This embodiment does not perform complex full-wave electromagnetic simulation, but directly constructs complex echo sequences in the RD domain. Each frame of RD observation can be represented as:

[0075]

[0076] in:

[0077] Static background items A complex Gaussian echo of fixed amplitude is superimposed in a specified range cell (Range Index=10, 20) to simulate strong static reflection and maintain relative stability across frames.

[0078] Target Item This indicates the change in the distance and position of the target echo over time. Its phase The Doppler effect occurs as velocity changes.

[0079] • During the 0–2 s and 6–8 s (motion segments). A significant Doppler effect is produced as velocity changes;

[0080] • During the 2–6 s (stationary period), the velocity However, in order to simulate the micro-movements of breathing, a very weak perturbation is superimposed on the phase, so that the target exhibits the characteristic of continuous energy but near-zero Doppler.

[0081] Noise item Complex Gaussian white noise is added to simulate the system noise floor.

[0082] Step S3: Compare solutions and processing procedures

[0083] Input: Input the synthesized RD sequence into three filters respectively:

[0084] • Filter A (Contrast Group): Inter-frame differential filtering;

[0085] • Filter B (Comparison Group): Standard EMA static filter with fixed coefficients ;

[0086] • Filter C (in this invention): Adaptive frozen EMA static filtering.

[0087] Record the RD residual map of the output of the three filters for each frame, extract the energy of the target unit and plot the curve of change with the number of frames to characterize the attenuation phenomenon of whether the target is absorbed by the background.

[0088] Step S4: Experimental Results and Analysis

[0089] As attached Figure 2 As shown, the horizontal axis represents the number of frames, and the vertical axis represents the target energy (dB), with the target's static periods marked by shading. From the graph, we can observe that:

[0090] Inter-frame difference (green line) is generally weak during the static period, with a spike only appearing at the action switching boundary. It is difficult to continuously highlight the static human body echo, indicating that it is not suitable as a long-term robust clutter suppression strategy for static human body point cloud generation.

[0091] The standard EMA (blue line) shows a significant decline during the quiescent period, gradually decaying from an initial high energy to a lower plateau, exhibiting a typical phenomenon of persistent targets being absorbed by the background model;

[0092] The adaptive frozen EMA (red line) maintains a relatively stable platform during the quiescent period, with only minor fluctuations and almost no decay of target energy, indicating that the freezing mechanism effectively prevents background updates from engulfing slow targets.

[0093] In summary, this embodiment demonstrates that, compared to inter-frame difference and standard EMA static filtering with fixed coefficients, the adaptive frozen EMA of this invention can achieve significantly stronger preservation capabilities for stationary but persistent pedestrian targets in the presence of strong static background scattering, avoiding exponential decay of target energy during the static period. This provides a more reliable target energy basis for subsequent CFAR detection and two-dimensional MVDR angle measurement, and ultimately improves the point cloud generation quality and action recognition stability.

[0094] Example 2

[0095] To verify the super-resolution capability of the local Kriging interpolation 2D MVDR angle measurement algorithm described in this invention when the angle interval between two stationary targets is small, this embodiment constructs a stationary scene with two targets and compares it with the standard MVDR method with different grid step sizes, using the resolution success rate and angle difference measurement accuracy as indicators. Step S1: Experimental Scene

[0096] The scenario includes two stationary point targets whose radial distance and elevation angle are set to match those of the radar, with separation only in the angular domain. The azimuth interval between the two targets gradually increases from 1° to 20° to characterize the transition range from indistinguishable to stably resolvable by the algorithm.

[0097] Step S2: Comparison Algorithm Settings

[0098] To demonstrate the effectiveness of local interpolation in reducing computational load while maintaining super-resolution capability, this embodiment compares the following three methods:

[0099] Method A: Standard MVDR, angle search grid step size set to Grid=0.01 (direction cosine domain)

[0100] Method B: Standard MVDR, with the angle search grid step size set to Grid=0.1 (direction cosine domain).

[0101] Method C: Local Kriging Interpolation MVDR (This Invention)

[0102] Step S3: Evaluation Indicators and Statistical Methods

[0103] Resolution success rate: When the algorithm detects two obvious spectral peaks near the true angle of two targets (satisfying the peak significance and minimum peak spacing conditions), it is recorded as a success; repeat the process multiple times for each angle interval (multiple noise implementations / multiple snapshot segments), and the success rate is calculated as the resolution success rate.

[0104] Angle difference measurement accuracy: For successful samples, record the corresponding angles of the two spectral peaks. Calculate the measured angle difference And plot the curve by comparing it with the actual angle difference.

[0105] Step S4: Experimental Results and Analysis

[0106] As attached Figure 3 As shown, the horizontal axis represents the angular difference between the two targets (in terms of direction cosine difference). (This is represented by the vertical axis, where the vertical axis represents the success rate of resolution.) It can be seen that:

[0107] Standard coarse-grid MVDR (Grid=0.1) fails to form two stable spectral peaks in the small angular difference region for a long time, and the resolution success rate only increases rapidly after the angular difference becomes larger, reflecting the significant resolution upper limit and spectral peak quantization effect brought by the coarse grid. The success rate curve of the method of this invention is relatively shifted to the left, that is, a higher resolution success rate can be obtained at a smaller angular difference. The overall performance is better than coarse-grid MVDR, and the success rate rises faster in the transition region. Furthermore, compared with fine-grid MVDR (Grid=0.01), the resolution success rate of the method of this invention is close to that of fine-grid MVDR, indicating that coarse grid + ROI local interpolation can maintain super-resolution capability close to that of fine-grid search while significantly reducing the number of global grid points.

[0108] As attached Figure 4 As shown, the horizontal axis represents the true angle difference (°), the vertical axis represents the measured angle difference (°), and the dashed line represents the ideal baseline. It can be seen that:

[0109] The measured angle difference of coarse-grid MVDR is obviously stepped, with typical quantization bias and jumps, making it difficult to accurately reflect the true angle difference. The method of this invention can follow the baseline change more continuously over a larger range, with smaller overall bias, and the improvement is particularly obvious in the small angle difference range, indicating that local interpolation reconstruction significantly improves the accuracy of spectral peak positioning.

[0110] In summary, this embodiment demonstrates that the method of the present invention simultaneously improves the resolution success rate and the angle difference measurement accuracy in dual-target scenarios, and avoids the spectral peak quantization error problem of coarse grid search while significantly reducing the number of global grid points.

[0111] Example 3

[0112] To verify the improvement effect of the method of the present invention on the continuity and accuracy of dynamic target point cloud trajectory in applications such as human motion recognition, this embodiment constructs a dynamic scene of a single target making circular motion in the direction cosine plane (uv), and compares the trajectory tracking capabilities of standard fine mesh MVDR and the local Kriging interpolation MVDR of the present invention.

[0113] Step S1: Experimental Scenario

[0114] The scene contains only one target, which moves in a circle in the direction cosine plane. Let the direction cosine be denoted as . The cosine of the pitch direction is The target trajectory can then be set as follows:

[0115]

[0116]

[0117] in radius of the circular trajectory Let be the angular velocity. This design allows for a direct verification of the continuity and fit of the estimated point within the two-dimensional angular domain.

[0118] Step S2: Comparison Algorithm Settings

[0119] Method A: Standard MVDR, angle search grid step size set to Grid=0.01 (direction cosine domain)

[0120] Method B: Local Kriging Interpolation MVDR (This Invention)

[0121] Step S3: Evaluation Indicators

[0122] Continuity: Observe the smoothness of the estimated points between frames and whether there are abrupt jumps, breaks, or missing points.

[0123] Accuracy: The degree of deviation between the estimated trajectory and the reference trajectory in the uv plane.

[0124] Step S4: Experimental Results and Analysis

[0125] As attached Figure 5 As shown, the gray curve represents the baseline circular trajectory, the blue markers are the standard MVDR (Grid=0.01) estimation points, and the red markers are the local Kriging interpolation MVDR estimation points of this invention. It can be seen that:

[0126] The method of this invention first locks the ROI with a coarse spectrum and then performs local interpolation reconstruction within the ROI. The spectral peak positioning is stable, and the estimated points are close to the reference trajectory as a whole, exhibiting better trajectory continuity and smaller local deviations.

[0127] In summary, this embodiment demonstrates that the method of the present invention can obtain stable and continuous two-dimensional angle estimation results in dynamic trajectory tracking scenarios, which is beneficial for subsequent three-dimensional point cloud generation and human motion recognition.

[0128] Example 4

[0129] To verify the improvement effect of the method of the present invention on the quality of human point cloud generation under real millimeter-wave radar acquisition data, this embodiment uses a Texas Instruments (TI) IWR6843ISK-ODS millimeter-wave radar board to collect raw sampling data of a single person in different postures and movements, and executes the complete point cloud generation process proposed in this invention on the collected data to obtain three-dimensional point cloud results under different human postures, which is used to prove that the present invention can provide a point cloud representation with clearer shape and stronger spatial consistency for human action recognition.

[0130] Step S1: Experimental setup and data collection conditions

[0131] Radar hardware: The TI IWR6843ISK-ODS radar board was selected as the data acquisition device. This radar has a 3Tx×4Rx time division multiplexing multiple input multiple output (TDM-MIMO) antenna array structure, which can form a virtual array within a single frame for two-dimensional angle measurement of azimuth and elevation angles.

[0132] Coordinates and Placement: The location of the radar is taken as the origin of the three-dimensional coordinate system, and the radar is installed 1 m above the ground. The subject is standing / sitting / lying down facing the radar, with the horizontal distance between the center of the body and the radar approximately 2 m.

[0133] Operating parameters: The radar frequency sweep range is set to 60–64 GHz. Each frame contains 128 chirs, each chirp contains 512 sampling points, and the frame period is 70 ms. Raw sampling data under different attitudes is collected under the above parameters as input for subsequent processing.

[0134] Step S3: Data Processing and Point Cloud Generation

[0135] The collected raw sampling data is processed frame by frame, and the following steps are performed sequentially:

[0136] (1) Range-Doppler Transform: Perform range-direction and Doppler-direction FFT on each frame of sampled data to obtain range-Doppler (RD) spectrum data, and form the RD observations required for subsequent static filtering and target detection.

[0137] (2) EMA static filtering with adaptive freezing mechanism: EMA modeling of background clutter is performed in the RD domain. At the same time, the freezing is triggered by threshold decision based on power-median baseline. Background updates are suppressed when the human body is approximately stationary or there are slight movements, so as to avoid human body scattering being absorbed into the background model, thereby improving the stability and detectability of target residuals under slight movement conditions.

[0138] (3) Two-dimensional CFAR human body scattering detection with multi-channel aggregation: The multi-channel RD residuals after static filtering are aggregated to construct energy, and two-dimensional CFAR detection is performed on the range-Doppler two-dimensional plane to obtain the candidate unit set corresponding to human body scattering, and output the multi-channel observation vector corresponding to each candidate unit.

[0139] (4) Two-dimensional MVDR angle measurement with adaptive diagonal loading: The spatial covariance matrix of the candidate cell is estimated by multi-channel data, and adaptive diagonal loading based on the trace of the covariance matrix is ​​introduced to enhance the robustness under array mismatch and noise environment changes. Then, the two-dimensional MVDR spatial spectrum is calculated to obtain the coarse resolution spatial spectrum estimation result in the angle domain.

[0140] (5) Local Kriging interpolation reconstruction of spatial spectrum: High-energy regions are extracted from the two-dimensional MVDR coarse spatial spectrum, and ordinary Kriging is used only in this region for local interpolation reconstruction. This significantly reduces the computational cost of global fine grid search while achieving high-resolution refinement of the spectral peak position, resulting in more accurate azimuth and elevation angle estimates.

[0141] (6) 3D point cloud construction: Combine the detected distance information with the interpolated and refined azimuth and elevation angles to calculate the corresponding 3D Cartesian coordinates. The data are then aggregated into a 3D point cloud of the human body for each frame; the point clouds of consecutive frames can be further used for downstream tasks such as human pose / motion recognition.

[0142] Step S4: Experimental Content and Results Presentation

[0143] This embodiment collects data under three typical human postures for a single person: standing, sitting, and lying down. After performing the complete process described above on the data for each posture, the results are shown in the attached figure. Figures 6-8 The 3D point cloud result is shown. The point cloud morphology reveals the following:

[0144] In a standing position ( Figure 6 Point clouds have a larger expansion range in the vertical direction and exhibit a spatial distribution characteristic consistent with the direction of human height.

[0145] In a seated position ( Figure 7 The vertical expansion range of point clouds is relatively reduced, and more concentrated scattering point clusters are formed in the upper body area of ​​the human body.

[0146] In a supine position ( Figure 8 The point cloud is significantly compressed in the vertical direction, while exhibiting a lower and more spread-out distribution pattern in the horizontal plane, reflecting the geometric characteristics of a lying human body parallel to the ground.

[0147] The spatial distribution differences of the three types of pose point clouds are clearly distinguishable, indicating that the method of the present invention can stably generate a three-dimensional point cloud shape consistent with human pose under real radar acquisition conditions, providing a more reliable point cloud input for subsequent human pose classification, action recognition and skeleton reconstruction.

[0148] In summary, this embodiment demonstrates that, through verification using measured data from the TI IWR6843ISK-ODS, the point cloud generation process proposed in this invention can generate 3D human point clouds with clear morphology and strong spatial consistency under multi-pose conditions, and possesses good engineering feasibility and application value.

Claims

1. A millimeter wave radar point cloud generation method with frozen mechanism static filtering and local interpolation angle measurement, characterized in that, Includes the following steps: Step S1: receive raw sampling data of millimeter wave radar echo, organize into multi-channel data under multi-transmission and multi-reception system according to frame. Perform distance direction fast Fourier transform (FFT) on each chirp to obtain distance spectrum, and perform FFT on slow time dimension at the same distance gate to obtain distance-Doppler spectrum. Thus, a distance-Doppler complex data cube containing virtual array element channels is formed; Step S2: Perform static filtering with an adaptive freezing mechanism on the multi-channel range-Doppler complex data cube to obtain background estimation and filtering residuals. When a range-Doppler cell is determined to be a suspected human scattering cell, the smoothing coefficient is increased to freeze the background update of that cell. Step S3: Perform multi-channel energy aggregation on the filter residuals from step S2 to construct a range-Doppler two-dimensional power map, and perform two-dimensional constant false alarm rate (CFAR) detection on the two-dimensional power map to obtain a set of candidate units for human body scattering targets; Step S4: For the set of candidate units of human body scattering targets, construct the array covariance matrix, and use the two-dimensional minimum variance distortionless response (MVDR) spatial spectrum estimation with adaptive diagonal loading to calculate the coarse spatial spectrum on the sparse angle grid. Step S5: Based on the coarse spatial spectrum from step S4, determine the energy peak region. Within this region, use local Kriging interpolation to perform fine-grid reconstruction of the two-dimensional MVDR spatial spectrum to obtain the target azimuth and elevation angle estimates. Step S6: Combine the target distance obtained in step S1 with the azimuth and pitch angles obtained in step S5 to complete the mapping from distance-azimuth-pitch angle to the Cartesian coordinate system and output a 3D point cloud.

2. The method according to claim 1, characterized in that, The EMA background estimation in step S2 satisfies: wherein is the first frame at Doppler index , distance index , observation, is the background estimate, and the filtered residual is .

3. The method according to claim 2, characterized in that, The By frame period With background expected time constant Mapping results:

4. The method according to any one of claims 1 to 3, characterized in that, The adaptive freezing mechanism in step S2 includes: for each distance gate Robust baseline is calculated on Doppler bins with direct current neighborhood removed When the cell power Satisfies: When, will Promoted to To freeze the background updates, so as to protect the slow / continuous target from being absorbed, is the freeze threshold coefficient.

5. The method according to claims 3 to 4, characterized in that, In indoor static monitoring scenarios, In scenes where people move around frequently and the background changes rapidly, Freeze threshold coefficient Based on the probability of false triggering of target freeze The following settings are configured to satisfy the following under purely noisy conditions: in The false trigger probability of freezing for a single distance-Doppler unit.

6. The method according to claim 1, characterized in that, Step S3 uses a two-dimensional sliding window with training and protection units to estimate the noise level for each unit to be tested and generate an adaptive threshold, outputting a set of candidate units that pass the threshold decision.

7. The method according to claim 1, characterized in that, Step S4, adaptive diagonal loading, includes: setting the covariance matrix to... The number of array elements is The load size is: and with Alternative Participating in two-dimensional MVDR spatial spectrum calculations, among which As a scaling factor, It is an identity matrix.

8. The method according to claim 1, characterized in that, In step S4, the two-dimensional MVDR spatial spectrum is located at the angular position. The following conditions must be met: in For array guide vector, It is the azimuth angle. It is the pitch angle.

9. The method according to claim 1, characterized in that, The local Kriging interpolation in step S5 uses Ordinary Kriging constraints. And a Gaussian mutability function model is used: in The angular plane distance Value of a nugget. For the height of the arch, These are the range parameters.

10. The method according to claim 1, characterized in that, Step S5's local Kriging interpolation employs a coarse-to-fine strategy: first, calculations are performed on the coarse mesh. Then, local weights are extracted from the pre-calculated weight matrix within the peak region. And calculate: Target search is performed only within the peak region to output fine angle estimation.

Images