Real-time human body detection method and system based on microwave radar

Abstract

The invention provides a real-time human body detection method and system based on a microwave radar, and is applied to the field of signal data processing. By constructing a sparse reconstruction model and combining multi-dimensional feature decomposition, the method effectively relieves the recognition difficulty caused by target signal aliasing, continuously outputs and receives radar signals, extracts distance, Doppler and angle features, constructs target distribution point clouds in a three-dimensional feature space, further utilizes angle continuity judgment and point cloud density analysis, and improves the recognition accuracy of target distribution point clouds. According to the method, the number and distribution of overlapped targets are accurately estimated, a Kalman filter is introduced for dynamic tracking for a scene in which a continuous area is not formed, continuous separation and pseudo-color image reconstruction of multiple targets are realized, and thus the resolution precision and target positioning capability of a radar in a multi-person close-range aggregation scene are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of signal data processing, and in particular to a real-time human body detection method and system based on microwave radar. Background Art

[0002] Microwave radar transmits electromagnetic waves at specific frequencies (such as 24 GHz, 60 GHz, or 77 GHz). These waves reflect off the human body and receive echo signals. Microscopic movements of the human body (such as breathing, heartbeats, and micro-movements) cause minute variations in the echo signal's frequency (micro-Doppler effect), amplitude, and phase. By analyzing these variations, human presence detection, motion recognition, and even vital sign monitoring can be achieved.

[0003] However, when two or more targets are very close in space (such as only a few tens of centimeters apart), the radar waves reflected by them almost overlap in time delay (distance information) and may also be very close in spectrum and angle, resulting in serious superposition of received signals. The superimposed signals are difficult to distinguish through traditional matched filters or beamformers, resulting in incorrect quantity estimation or misidentification. Summary of the Invention

[0004] The present invention aims to solve the problem that the echo time, frequency and amplitude characteristics of multiple human body reflections detected by radar are similar, making it impossible to accurately estimate the number of targets and their respective positions. It provides a real-time human body detection method and system based on microwave radar.

[0005] The present invention adopts the following technical means to solve the technical problem:

[0006] The present invention provides a real-time human body detection method based on microwave radar, comprising:

[0007] Based on a detection scene preset by the microwave radar, repeatedly outputting and receiving the echo signal of the detection scene within a preset time period;

[0008] Determining whether a combined signal peak is detected in the echo signal;

[0009] If so, based on the pre-collected original radar signal, the original radar signal is used as superimposed data of the sparse source to construct a corresponding sparse reconstruction model, and a preset sparse coding is used to decompose the aliased signal into corresponding sparse units. Multidimensional features are extracted from the sparse units, and a target distribution point cloud of the multidimensional features is established through a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features;

[0010] Determining whether the angular direction of the target distribution point cloud forms a continuous area;

[0011] If not, the local point cloud density of the target distribution point cloud is identified, and the corresponding target number is dynamically estimated based on the local point cloud density. A preset Kalman filter is applied to continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

[0012] Furthermore, the step of decomposing the aliased signal into corresponding sparse units by using a preset sparse coding and extracting multidimensional features from the sparse units further includes:

[0013] Calculating the amplitude square integral of the sparse unit, normalizing the amplitude square integral to a preset standard scale, applying a preset continuous wavelet transform to the sparse unit, and generating a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate;

[0014] Determining whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat;

[0015] If so, based on the echo delay of the sparse unit, the target distance corresponding to the echo delay is collected, and according to the target distance, the internal point distribution form of the sparse unit is counted, and within a preset time window, the time trajectory evolution of the sparse unit is tracked, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

[0016] Furthermore, after the step of establishing the target distribution point cloud of the multi-dimensional feature through the preset three-dimensional feature space, the method further includes:

[0017] Based on the point cloud density of the target distributed point cloud, identifying the number of neighboring points of each point cloud within a preset radius;

[0018] Determine whether the number of neighbor points is less than a preset value;

[0019] If so, the extreme outliers of the multidimensional features are detected, and based on the extreme outliers, the discreteness of each point cloud is evaluated, the topological characteristics of the target distribution point cloud are counted, and additional features are added to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

[0020] Furthermore, the step of identifying the local point cloud density of the target distribution point cloud and dynamically estimating the corresponding number of targets based on the local point cloud density further includes:

[0021] Identifying continuous frame point clouds from the target distribution point cloud, performing temporal fusion on the continuous frame point clouds, and generating corresponding point cloud high-density areas;

[0022] Determining whether the high-density area of ​​the point cloud appears repeatedly within a preset frequency;

[0023] If so, the position information of the high-density area of ​​the point cloud is obtained, the preset micro-motion spectrum features are extracted from the high-density area of ​​the point cloud, and the target types of the continuous frame point clouds are dynamically divided based on the position information and the micro-motion spectrum features, wherein the micro-motion spectrum features specifically include breathing frequency and arm swing frequency, and the target types specifically include dynamic human bodies, interference objects and non-human objects.

[0024] Furthermore, the step of determining whether the echo signal detects a combined signal peak further includes:

[0025] Obtaining the full width at half maximum of the main peak of the echo signal;

[0026] Determining whether the main peak half-width exceeds a preset multiple of the historical single target waveform width;

[0027] If so, several small peaks of the half-width at half maximum of the main peak are detected, the integrated energy on the left and right sides of the small peak is identified, and based on the degree of symmetry of the integrated energy, the superimposed waveform caused by the corresponding reflection amplitude of each target is dynamically collected, wherein the small peaks specifically include shoulder peak secondary valleys, platform waveform secondary valleys and inter-peak secondary valleys.

[0028] Furthermore, the step of determining whether the angular direction of the target distribution point cloud forms a continuous area further includes:

[0029] Sorting the target distribution point cloud according to preset angle values, constructing a corresponding angle sequence by sorting, and calculating the angle increment of the angle sequence;

[0030] Determining whether the angle increment increases;

[0031] If so, identify the breakpoint after the angle increment becomes larger, perform cluster cutting at the breakpoint, and use it as the separation boundary of the point cloud target. Based on the position of the breakpoint, divide the point cloud target into several continuous segments, and use the several continuous segments as the angle area of ​​a single candidate target.

[0032] Furthermore, the step of repeatedly outputting and receiving the echo signal of the detection scene within a preset time period based on the detection scene preset by the microwave radar further includes:

[0033] Adaptively switching the beam coverage strategy of the microwave radar based on a preset scene type of the detection scene, wherein the scene type specifically includes a corridor, a room, and a doorway, and the beam coverage strategy specifically includes omnidirectional, directional, and variable beams;

[0034] Determine whether the beam coverage strategy can meet the preset target dynamic detection;

[0035] If not, the preset static shielding area is dynamically activated according to the preset detection parameters of the microwave radar, and the obstacle echoes pre-collected by the microwave radar are blocked through the static shielding area, wherein the detection parameters specifically include the horizontal scanning angle range, vertical coverage range and effective distance.

[0036] The present invention also provides a real-time human body detection system based on microwave radar, comprising:

[0037] A receiving module is used to repeatedly output and receive echo signals of a detection scene preset by the microwave radar within a preset time period;

[0038] A judgment module, used to judge whether the echo signal detects a combined signal peak;

[0039] an execution module, configured to, if so, construct a corresponding sparse reconstruction model based on pre-collected original radar signals, using the original radar signals as superimposed data of sparse sources, decompose the aliased signals into corresponding sparse units using preset sparse coding, extract multidimensional features from the sparse units, and establish a target distribution point cloud of the multidimensional features through a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features;

[0040] A second judgment module is used to judge whether the angle direction of the target distribution point cloud forms a continuous area;

[0041] The second execution module is used to identify the local point cloud density of the target distribution point cloud if it has not been formed, dynamically estimate the corresponding target number based on the local point cloud density, apply a preset Kalman filter, continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

[0042] Furthermore, the execution module further includes:

[0043] a calculation unit, configured to calculate the amplitude square integral of the sparse unit, normalize the amplitude square integral to a preset standard scale, apply a preset continuous wavelet transform to the sparse unit, and generate a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate;

[0044] a judging unit, configured to judge whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat;

[0045] An execution unit is configured to, if yes, collect the target distance corresponding to the echo delay of the sparse unit based on the echo delay, count the internal point distribution form of the sparse unit according to the target distance, and track the time trajectory evolution of the sparse unit within a preset time window, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

[0046] Furthermore, it also includes:

[0047] an identification module, configured to identify the number of neighboring points of each point cloud within a preset radius based on the point cloud density of the target distribution point cloud;

[0048] A third judgment module is used to judge whether the number of neighbor points is less than a preset value;

[0049] The third execution module is used to detect the extreme outliers of the multidimensional features, evaluate the discreteness of each point cloud according to the extreme outliers, count the topological characteristics of the target distribution point cloud, and add additional features to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

[0050] The present invention provides a real-time human body detection method and system based on microwave radar, which has the following beneficial effects:

[0051] The present invention effectively alleviates the recognition difficulties caused by target signal aliasing by constructing a sparse reconstruction model and combining it with multi-dimensional feature decomposition. It continuously outputs and receives radar signals, extracts distance, Doppler and angle features, and constructs a target distribution point cloud in a three-dimensional feature space. It further uses angle continuity judgment and point cloud density analysis to accurately estimate the number and distribution of overlapping targets. For scenes where no continuous area is formed, a Kalman filter is also introduced for dynamic tracking to achieve continuous separation of multiple targets and pseudo-color image reconstruction, thereby significantly improving the radar's resolution accuracy and target positioning capabilities in scenes where multiple people gather in close proximity. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of an embodiment of a real-time human body detection method based on microwave radar according to the present invention;

[0053] Figure 2 This is a structural block diagram of an embodiment of a real-time human body detection system based on microwave radar of the present invention. DETAILED DESCRIPTION

[0054] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The implementation, functional features and advantages of the present invention will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0056] Reference Attachment Figure 1 , a real-time human body detection method based on microwave radar in one embodiment of the present invention, comprising:

[0057] S1: Based on a preset detection scenario of the microwave radar, repeatedly output and receive echo signals of the detection scenario within a preset time period;

[0058] S2: Determine whether a combined signal peak is detected in the echo signal;

[0059] S3: If yes, then based on the pre-collected original radar signal, use the original radar signal as superimposed data of the sparse source, construct a corresponding sparse reconstruction model, use a preset sparse coding to decompose the aliased signal into corresponding sparse units, extract multidimensional features from the sparse units, and establish a target distribution point cloud of the multidimensional features through a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features;

[0060] S4: Determine whether the angle direction of the target distribution point cloud forms a continuous area;

[0061] S5: If not, identify the local point cloud density of the target distribution point cloud, dynamically estimate the corresponding target number based on the local point cloud density, apply a preset Kalman filter, continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

[0062] In this embodiment, the system is based on the detection scene pre-set by the microwave radar, and repeatedly outputs and receives the echo signal of the detection scene within a pre-set time period. The system then determines whether these echo signals detect the combined signal peak to execute the corresponding steps; for example, when the system determines that the repeated output and reception of the echo signal of the detection scene does not detect the combined signal peak, the system will believe that there is a certain distance separation between the human targets, and the radar echo does not produce obvious overlap or aliasing in the time, frequency and angle dimensions. The system will continue to maintain the current output and reception cycle, increase the detection capability of weak targets or distant targets, consider adjusting the radar's receiving gain or changing the transmission power to enhance the detection capability of distant or small targets. At the same time, since there is no merging of signal peaks, it means that the targets corresponding to the echo signals are relatively scattered. The system can directly process each independent echo, extract the multi-dimensional features of each echo (such as distance, speed, Doppler effect, etc.), and perform target recognition and positioning in the three-dimensional feature space. If the number of targets is small, the system can identify each target separately through a simple separation method based on distance and Doppler features to ensure the accurate position and motion state of each target. According to the target distribution, the working parameters of the radar system (such as scanning angle, transmission frequency, etc.) are dynamically adjusted to better adapt to the scene requirements. If the system finds that the detection area is relatively static, it can also choose to reduce the output frequency to reduce power consumption. For example, when the system The system determines that the echo signal of the repeated output and reception detection scene has detected the combined signal peak. At this time, the system will think that the distance between the human targets is close, which is easy to cause the radar echo to overlap or alias. The system will use the original radar signals collected in advance as the superposition data of sparse sources to build a corresponding sparse reconstruction model, and use the pre-set sparse coding to decompose the aliased signal into corresponding sparse units, and extract multi-dimensional features from these sparse units. The multi-dimensional features specifically include distance features, Doppler features and angle features. Through the pre-set three-dimensional feature space, a target distribution point cloud with multi-dimensional features is established; the system regards the original radar signals as the superposition data of multiple sparse sources, and uses the sparse reconstruction model to decompose the aliased signals into corresponding sparse units. The signal deconstruction model can effectively separate the merged signal into sparse units with identifiable features, thus breaking through the limitation of traditional methods that cannot distinguish individuals in close-range multi-person detection. At the same time, the three-dimensional features of distance, Doppler and angle are extracted from the sparse units to form richer target information, which is conducive to distinguishing targets from multiple dimensions. In particular, when the targets have similar shapes and similar movement directions, the accuracy of target classification and recognition is improved. In addition, the target distribution point cloud established through the three-dimensional feature space can intuitively display the relative position and density relationship of the targets in space, providing the basic data structure for subsequent operations such as target number estimation, trajectory separation and dynamic tracking, and improving the system's multi-target management capabilities in crowded or occluded scenes.The system then determines whether the angular directions of these target distribution point clouds form a continuous area to execute the corresponding steps; for example, when the system determines that the angular directions of the target distribution point clouds form a continuous area, the system will believe that relatively complete target reflection information is detected in this angular area, which can more accurately represent the directional expansion range of the target in space. The system will adjust the processing method of the echo signal and merge the echo signals in the continuous area to avoid unnecessary signal separation. At this time, the target recognition strategy should focus on the analysis of group behavior rather than the separation of individual targets in order to improve processing efficiency. At the same time, according to information such as density analysis, motion patterns and multidimensional features, the number of targets is dynamically estimated, and appropriate classification algorithms (such as cluster-based classification) are used. Classification method) to classify and identify these targets. Through the dynamic position changes of the targets, the relationship between targets can be judged more accurately to avoid misidentification in multi-target scenes. For the occasional continuous angle direction area, the system should carefully exclude noise interference and set reasonable thresholds and signal enhancement mechanisms to ensure that the continuity of target distribution is not due to misjudgment caused by interference signals or environmental noise. For example, when the system determines that the angle direction of the target distribution point cloud does not form a continuous area, the system will think that no complete target reflection information is detected in this angle area, and it cannot represent the direction expansion range of the target in space. The system will identify the local point cloud density of the target distribution point cloud, and dynamically estimate the corresponding local point cloud density according to different local point cloud densities. The number of human targets is determined by applying a pre-set Kalman filter to continuously track the target parameters of each human target, including the target speed and target speed change, to generate separated multi-target echoes, and reconstruct these multi-target echoes into a pseudo-color radar image; by further analyzing the local point cloud density of the target distribution point cloud, the system can capture the existence of multiple high-density reflection sub-areas even in the absence of extension in the angular direction, thereby dynamically estimating the actual number of targets. This density-based supplementary mechanism significantly improves the recognition robustness of the system in weak signal areas or non-ideal reflection conditions. At the same time, when the distance between multiple targets is close or the moving speed is similar, their radar echoes often overlap in time, frequency, and angle dimensions. , making it difficult to distinguish instantaneous targets. At this point, the Kalman filter is introduced. Based on the target's historical motion state and combined with new observation data, it continuously estimates and corrects the speed and speed change parameters of each target. This not only effectively avoids target trajectory interruptions caused by short-term aliasing, but also maintains the consistency of target trajectory in the face of partial occlusion or signal loss. After completing sparse reconstruction, feature extraction, point cloud generation, and target tracking for multiple targets, the system synthesizes the separated target echo signals into a pseudo-color radar image. Each target is assigned a different color code based on its characteristic parameters. This visualization method improves the readability of the system results and the efficiency of human-computer interaction, allowing end users to intuitively identify the number, location, and movement direction of targets.

[0063] It should be noted that, based on the pre-collected original radar signal, the original radar signal is used as the superposition data of the sparse source to construct a corresponding sparse reconstruction model, and the aliased signal is decomposed into corresponding sparse units using a preset sparse coding. Multidimensional features are extracted from the sparse units. A specific example is as follows:

[0064] Suppose a microwave radar is installed on the ceiling or side wall of an office corridor to monitor the number, direction, and speed of people passing through. At a certain moment, two employees, A and B, walk side by side, less than 40 centimeters apart, at the same speed and of similar size. A clear echo peak is formed at the radar receiving end, but only one set of echo data is generated, as shown below:

[0065] There is only one concentrated reflection (e.g. 2.5 meters) away from the reflection peak;

[0066] The Doppler shift is 0.4 m / s (forward direction);

[0067] Angle estimates are close to overlapping (e.g. +5°);

[0068] If a radar system uses traditional CFAR or FFT+angle estimation, it is usually unable to distinguish between two people and may mistakenly identify them as a "single target," which will cause subsequent statistical errors. The following steps are then followed:

[0069] 1. Upon detecting the combined signal peak, the radar system analyzes the echo energy distribution at that moment and finds that the reflected signal is wide (waveform broadening), indicating signs of signal overlap, thus triggering the aliasing detection mechanism.

[0070] 2. Build the sparse reconstruction model system and call the "Original Radar Signal Library", which contains the following collected data:

[0071] Single person reflection templates A1, A2, and A3: represent people of different heights, shapes, directions, and speeds;

[0072] Each template signal structure contains radar raw IQ data of multiple channels (such as range channel, velocity channel, and angle channel);

[0073] Build a model and express the current echo signal Y as a linear combination (sparse form) of a set of original templates A:

[0074] Y≈α1*A1+α2*A2+...+αn*An+e

[0075] Where α is the sparse coefficient and e is the residual;

[0076] 3. Using sparse coding decomposition, the system uses the orthogonal matching pursuit (OMP) algorithm for decoding: in the sparse dictionary, the most relevant templates are matched one by one; ultimately, two sets of templates A5 and A9 with significant responses are matched;

[0077] The decoding results show that the current signal can be interpreted as the overlapping result of two sparse target sources;

[0078] 4. Multi-dimensional feature extraction. For the original templates of A5 and A9, the system extracts the following features:

[0079] feature Target A (corresponding to A5) Target B (corresponding to A9) distance 2.43m 2.56m speed 0.39m / s 0.41m / s angle +4.3° +6.2°

[0080] The system represents these two sparse units as two independent point cloud clusters in the three-dimensional feature space, preparing for subsequent point cloud connectivity judgment, target number estimation, trajectory tracking, etc.

[0081] To summarize, the above examples show that the system ultimately determines that there are actually two targets at that moment, not one, successfully avoiding the "target merging" or "missed detection" problems of traditional processing algorithms; even if two people are close to each other and their movements are consistent, they are still separated into independent signals through sparse modeling; it is particularly suitable for human target recognition and statistical analysis in high-density environments such as elevator doors and classroom foyers.

[0082] It should be supplemented that the local point cloud density of the target distribution point cloud is identified, the corresponding target number is dynamically estimated based on the local point cloud density, a preset Kalman filter is applied, the target parameters of each target are continuously tracked, and separated multi-target echoes are generated. The multi-target echoes are reconstructed into a pseudo-color radar image. A specific example is as follows:

[0083] Imagine a microwave radar deployed above an elevator door in an office building to detect people waiting to board. At a certain time, as the elevator approaches its destination, several people gather at the door. The radar system needs to determine the number of people waiting and continuously track their movements.

[0084] At this time, four people enter the radar field of view almost at the same time, two of them are standing side by side, and the other two are approaching one after the other. The detailed system processing flow is as follows:

[0085] 1. Point cloud data generation: The radar detects multiple reflected signals, but due to the close distance between people (only about 0.3 to 0.5 meters), some echoes merge, initially forming a relatively dense but blurred 3D point cloud area.

[0086] 2. Local point cloud density recognition: The system divides the 3D point cloud into voxel grids, for example, a unit block of 0.2 meters x 0.2 meters x 10 degrees. It detects multiple high-density areas: two point cloud cores (two people side by side) near the front of the elevator door; and two sparse but stable point cloud clusters (one in front and one behind) slightly further back. The system calculates the density of each point cloud cluster (for example, >20 points per voxel is considered high density). Based on the density and its spatial distribution, it dynamically estimates the presence of four independent targets.

[0087] 3. Kalman filter tracking, create a Kalman filter for each estimated target and initialize the state vector:

[0088] Position coordinates (X, Y);

[0089] Velocity vector (Vx, Vy);

[0090] The system obtains the updated point cloud center position and Doppler velocity in each frame. The Kalman filter continuously predicts and corrects based on the time series to solve the problem of partial target occlusion or temporary signal disappearance.

[0091] For example:

[0092] In the first frame, the four targets are clear;

[0093] In the second frame, the third target is blocked from view, but the system continues to track its motion through prediction.

[0094] In the third frame, the third target reappears, and its path is highly consistent with the predicted path;

[0095] 4. Multi-target echo separation: The system reverse maps the tracking results back to the original radar signal and separates the reflection timing of each target to obtain 4 sets of time-frequency echo data of independent targets;

[0096] 5. Pseudo-color image reconstruction: the distance, Doppler velocity, and reflection intensity of each target are mapped to color elements on the pseudo-color image:

[0097] Distance → Hue: close is red, far is blue;

[0098] Speed ​​→ Saturation: Fast means high saturation, slow means low saturation;

[0099] Reflection intensity → Brightness: strong is bright, weak is dark;

[0100] The final image shows four "human shadow areas" with distinct colors, which facilitates subsequent system or manual identification;

[0101] To summarize, in the above examples, the system can still identify the number of targets with high precision in close-range crowded scenes, overcoming the problem of traditional radar where "overlapping waveforms make it difficult to distinguish individuals." Kalman filtering enhances the robustness against short-term occlusion or interference signals, making the tracking process more coherent. Pseudo-color images intuitively express changes in target status, facilitating equipment maintenance, intelligent identification, or data archiving and analysis.

[0102] In this embodiment, the aliased signal is decomposed into corresponding sparse units by using a preset sparse coding, and the step S3 of extracting multidimensional features from the sparse units further includes:

[0103] S31: Calculating the amplitude square integral of the sparse unit, normalizing the amplitude square integral to a preset standard scale, applying a preset continuous wavelet transform to the sparse unit, and generating a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate;

[0104] S32: Determine whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat;

[0105] S33: If so, based on the echo delay of the sparse unit, the target distance corresponding to the echo delay is collected, and according to the target distance, the internal point distribution form of the sparse unit is counted, and within a preset time window, the time trajectory evolution of the sparse unit is tracked, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

[0106] In this embodiment, the system calculates the amplitude square integral of the sparse unit, normalizes these amplitude square integrals to a preset standard scale, applies a preset continuous wavelet transform to these sparse units, and generates a corresponding time-frequency spectrogram. The time-frequency spectrogram specifically includes the main frequency, bandwidth and frequency change rate. Then the system determines whether the time-frequency spectrogram detects the preset low-frequency periodic motion characteristics, which specifically include breathing and heartbeat, to execute the corresponding steps; for example, when the system determines that the time-frequency spectrogram does not detect the preset low-frequency periodic motion characteristics, the system will believe that multiple human body micro-motion signals are superimposed on each other in the sparse unit, resulting in a special If the energy in the eigenfrequency band is diffused or the waveform is distorted, the low-frequency features cannot be clearly identified. The system will temporarily classify the reflection target corresponding to the current sparse unit as a "non-biological target" or "inactive target", and perform target classification and filtering based on the position and speed characteristics to avoid mistaking it for a human object that can interact or require an alarm. At the same time, if the target is located in the area of ​​interest (such as a bed, sofa, or monitoring range), the analysis period will be automatically extended (such as from 5 seconds to 10 seconds), and time-frequency analysis will be performed again to improve the weak signal detection capability. If no low-frequency periodic features are detected in multiple sparse units, you can try to adjust the radar gain, antenna beam, or activate auxiliary channels (such as vertical beam groups) to improve detection sensitivity. If such "targets without low-frequency features" appear repeatedly, they should be marked and uploaded to the cloud or back-end database as part of behavioral risk identification (such as sudden fainting or abnormal stillness); for example, when the system determines that the time-frequency spectrum detects the pre-set low-frequency periodic motion characteristics, the system will think that the radar signal quality is good, the micro-motion components (such as breathing, heartbeat, etc.) are not masked by noise or interference, the system can clearly distinguish the target's vital signs, and the signal processing effect is better. The system will collect the target distance corresponding to the echo delay based on the sparse unit, and statistically calculate the internal point distribution of the sparse unit according to different target distances. The morphology specifically includes the main axis length, flatness and density distribution characteristics. Within a pre-set time window, the time trajectory evolution of these sparse units is tracked. Based on the echo delay of the sparse unit, the system can collect the delay information of the echo signal, thereby accurately calculating the distance to the target. Through the echo delay, the system can determine the position of the target. Especially when the target is at different distances, the system can dynamically adjust its detection range to further improve the tracking accuracy. At the same time, the internal point distribution morphology of the sparse unit (including the main axis length, flatness, density distribution characteristics, etc.) is statistically analyzed to further reveal the morphological characteristics of the target. For example, the density distribution characteristics of the target can help to infer the size, shape and motion state of the target.Changes in the main axis length and flatness can be used to determine the target's trajectory and posture. Within a set time window, the system can track the temporal trajectory evolution of sparse units. By continuously tracking the target's trajectory, the system can obtain the target's motion trends (such as speed and direction changes) in real time, allowing for dynamic monitoring and prediction of the target. This is particularly important for applications such as smart homes, health monitoring, and security systems.

[0107] It should be noted that based on the echo delay of the sparse unit, the target distance corresponding to the echo delay is collected, and according to the target distance, the internal point distribution form of the sparse unit is counted, and the time trajectory evolution of the sparse unit is tracked within a preset time window. The specific example is as follows:

[0108] Imagine a microwave radar system installed at the entrance of a shopping mall to monitor for unidentified individuals and ensure mall safety. By detecting, identifying, and tracking targets, the system can effectively distinguish between different targets and determine their direction of movement.

[0109] 1. Radar detection delay for multiple sparse units. In this scenario, a radar system is installed at the entrance of a shopping mall and can detect multiple targets. Assume that the radar system detects two reflected waves with echo delays of 15 nanoseconds and 25 nanoseconds, corresponding to target distances of 2.5 meters and 3.2 meters, respectively.

[0110] Examples:

[0111] The first echo has a delay of 15 nanoseconds, and the calculated distance is 2.5 meters. The target is likely the first person approaching the entrance.

[0112] The second echo has a delay of 25 nanoseconds, and the calculated distance is 3.2 meters. The target may be another intruder near the entrance.

[0113] 2. Analyze the point cloud morphology to determine the target state. The system then analyzes the point clouds inside the sparse cells of the two targets to determine their target states.

[0114] Examples:

[0115] The first target: The main axis length of its point cloud is short (about 0.4 meters), with high flatness and density, indicating that the target is in a stable standing state and its posture is relatively stable. It may be staying at the door to observe or perform some actions.

[0116] The second target: The main axis length of its point cloud is longer (about 0.8 meters), the flatness is lower, the density is lower, and the target is moving significantly, indicating that the target is quickly passing through the entrance of the mall and may intend to enter.

[0117] 3. Time trajectory tracking and behavior identification: The system uses tracking algorithms (such as Kalman filtering) to track the time trajectories of the two targets to identify their motion trajectories and behaviors;

[0118] Examples:

[0119] First target: Within the 10-second window, the first target remained relatively stationary at the entrance and did not move significantly; the system determined this to be normal behavior.

[0120] Second target: The point cloud position of the second target moves rapidly from the entrance to the interior of the mall within 10 seconds. The system determines it as a fast-moving target based on its speed, acceleration, and other dynamic changes. Combined with historical data, the system may determine this target as an intruder and trigger a security alarm.

[0121] To sum up, in the above examples, the system can accurately determine the target's position, posture, and behavior status through echo delay and internal point cloud analysis, greatly improving the accuracy of target recognition. At the same time, by tracking the target's time trajectory and monitoring the target's changes in real time, it can effectively identify the target's behavior pattern and avoid false alarms. It can also handle interference between multiple targets, accurately distinguish the behavior trajectory of each target, and adapt to complex monitoring scenarios.

[0122] In this embodiment, after step S3 of establishing the target distribution point cloud of the multi-dimensional feature through the preset three-dimensional feature space, the following steps are further included:

[0123] S301: Based on the point cloud density of the target distribution point cloud, identifying the number of neighboring points of each point cloud within a preset radius;

[0124] S302: Determine whether the number of neighbor points is less than a preset value;

[0125] S303: If so, detect the extreme outliers of the multidimensional features, evaluate the discreteness of each point cloud based on the extreme outliers, count the topological characteristics of the target distribution point cloud, and add additional features to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

[0126] In this embodiment, the system identifies the number of neighboring points of each point cloud within a preset radius based on the point cloud density of the target distribution point cloud, and then the system determines whether the number of neighboring points is less than a preset value to execute the corresponding steps; for example, when the system determines that the number of neighboring points of each point cloud within a preset radius is not less than a preset value, the system will consider that the current point cloud has a high local density and stability and does not belong to discrete noise points. There is no need to remove noise points. The system will use a clustering algorithm to cluster these dense point cloud areas and distinguish different targets. The clustering results can help the system identify different targets and at the same time The clustered point cloud can be extracted to extract its geometric shape, boundary, center of mass position and other features, and the direction, volume and shape of the target can be further analyzed. If there is time series data, the system can calculate the target's motion parameters, such as speed, acceleration, direction, etc., based on the target's changing trajectory, and use algorithms such as Kalman filtering to track the target and further predict the target's future position. Based on the target's geometric features and dynamic changes, classification algorithms (such as deep learning models, support vector machines, etc.) are used for target recognition, such as identifying whether the target is a pedestrian, a stationary object or other dynamic object; for example, when the system determines that each point cloud is within a pre-set radius, If the number of neighbor points is less than the preset value, the system will consider the current point cloud to be unstable and may be discrete noise points, which need to be removed. The system will detect the extreme outliers of multi-dimensional features, which include excessive reflection energy and frequency jumps. Based on these extreme outliers, the discrete degree of each point cloud is evaluated, and the topological characteristics of the target distribution point cloud are statistically analyzed. The topological characteristics include the main direction of the point cloud and the overall shape of the point cloud. Additional features are attached to each point cloud, including local curvature and local velocity gradient. The system can determine the number of neighbor points of the point cloud within the preset radius, combined with the extreme outliers of reflection energy and frequency, to improve the point cloud’s stability. It accurately distinguishes valid target points from discrete noise points, thereby avoiding the accidental deletion of edge targets or weakly reflective targets, improving the purity and credibility of point cloud data, and further statistically analyzing the topological characteristics of the point cloud (such as the main direction and overall shape) to extract additional features such as local curvature and velocity gradient, which helps to accurately model the geometry and motion form of the target, improve the target's multi-dimensional feature recognition ability and the stability of continuous tracking. With the help of multi-dimensional outlier detection and topological analysis, the system can not only eliminate invalid point clouds, but also identify abnormal point clouds with characteristics such as violent reflection and speed mutation, which helps to warn of potential safety risks or atypical behaviors and enhance the system's intelligent monitoring and response capabilities.

[0127] In this embodiment, the step S5 of identifying the local point cloud density of the target distribution point cloud and dynamically estimating the corresponding number of targets based on the local point cloud density further includes:

[0128] S51: identifying continuous frame point clouds from the target distribution point cloud, performing temporal fusion on the continuous frame point clouds, and generating corresponding point cloud high-density areas;

[0129] S52: Determine whether the high-density area of ​​the point cloud appears repeatedly within a preset frequency;

[0130] S53: If yes, obtain the position information of the high-density area of ​​the point cloud, extract the preset micro-motion spectrum features from the high-density area of ​​the point cloud, and dynamically divide the target types of the continuous frame point cloud based on the position information and the micro-motion spectrum features, wherein the micro-motion spectrum features specifically include breathing frequency and arm swing frequency, and the target types specifically include dynamic human bodies, interference objects and non-human objects.

[0131] In this embodiment, the system identifies continuous frame point clouds from the target distribution point cloud, performs time-series fusion on these continuous frame point clouds, generates corresponding point cloud high-density areas, and then the system determines whether the point cloud high-density area recurs within a preset frequency to execute the corresponding steps; for example, when the system determines that the point cloud high-density area does not recur within the preset frequency, the system will believe that the point cloud changes in the area may not have obvious stability or periodicity, and there may be no continuous targets. The system will identify whether there are large changes or sudden fluctuations in the point cloud in the area. If the high-density area does not recur within the predetermined frequency, it means that the target may have moved out of the current detection range, or the target is masked by noise. At the same time, for high-density areas that do not recur, the system needs to update the target status, perform target tracking or leave the current area status update, which involves adjusting the target's motion model. and path prediction to avoid mistakenly tracking a non-existent or inactive target, and if the high-density area does not reappear, the system needs to further filter out potential environmental noise, eliminate false alarms, and dynamically adjust the sensitivity of the sensor, optimize the radar signal processing algorithm, reduce the interference of irrelevant factors on target detection, and improve the system's sensitivity and accuracy to valid targets; for example, when the system determines that a high-density area of ​​the point cloud reappears within a preset frequency, the system will consider that the point cloud changes in this area have obvious stability or periodicity, and there are continuous targets. The system will obtain the position information of the high-density area of ​​the point cloud, and extract the pre-set micro-motion spectrum features from the high-density area of ​​the point cloud. The micro-motion spectrum features specifically include breathing frequency and arm swing frequency. Based on the position information and micro-motion spectrum features, the target types of continuous frame point clouds are dynamically divided. The target types specifically include dynamic human bodies, interference objects, and non-human objects.The system can effectively distinguish dynamic human bodies from interference objects or non-human objects by extracting micro-motion spectrum features. For example, the human breathing frequency has a specific low-frequency stable feature, while the arm swing frequency reflects a rhythmic movement pattern. These features are obviously different in the spectrum. Compared with the judgment method that simply relies on static form or position point cloud, this method of integrating micro-motion features can significantly improve the accuracy of identifying dynamic human bodies, and avoid misidentifying wind-blown leaves, swinging curtains, etc. as human targets. At the same time, when high-density point cloud areas continue to repeat in frequency, the system can judge that the target is a long-term active or stationary object, rather than an instantaneous interference. This judgment based on frequency stability is helpful The system prioritizes processing resources for real, persistent human targets, improving overall monitoring efficiency and real-time response capabilities. This is particularly applicable to scenarios such as personnel retention monitoring and security deployment. Furthermore, the system integrates location information and micro-motion spectrum data to dynamically segment targets within continuous frame point clouds, identifying, for example, walking humans, periodic interference sources (such as fans), and stationary non-human objects (such as pillars and sofas). This classification approach, based on "target behavior + micro-motion features," empowers the system with enhanced target understanding capabilities, advancing the transition from "point cloud detection" to "semantic understanding," providing a solid foundation for subsequent behavior recognition, intelligent decision-making, and early warning control.

[0132] In this embodiment, the step S2 of determining whether the echo signal has detected a combined signal peak further includes:

[0133] S21: Obtaining the full width at half maximum of the main peak of the echo signal;

[0134] S22: Determine whether the main peak half-height width exceeds a preset multiple of the historical single target waveform width;

[0135] S23: If so, detect several small peaks of the half-width of the main peak, identify the integrated energy on the left and right sides of the small peak, and dynamically collect the superimposed waveform caused by the reflection amplitude of each target based on the degree of symmetry of the integrated energy, wherein the small peaks specifically include shoulder peak secondary valleys, platform waveform secondary valleys and inter-peak secondary valleys.

[0136] In this embodiment, the system obtains the half-width of the main peak of the echo signal, and then determines whether the half-width of the main peak exceeds a preset multiple of the historical single-target waveform width to execute the corresponding steps; for example, when the system determines that the half-width of the main peak of the echo signal does not exceed a preset multiple of the historical single-target waveform width, the system will believe that the current echo signal may come from a relatively concentrated single target, whose signal waveform is relatively stable and has not undergone excessive expansion or overlap. The system will confirm that the current echo signal comes from a single target, and perform single-target tracking and analysis, extract relevant information of the target, such as position, speed, motion trajectory, etc., and further predict the target's movement trend and possible behavioral changes. At the same time, since the echo signal indicates that the target is single and clear The system chooses to optimize the target detection algorithm to improve its tracking accuracy, such as using the Kalman filter or other tracking algorithms to accurately track the target to avoid misjudgment caused by multi-target interference. The system can reduce the need for complex calculations and focus on tracking and analyzing a single target to improve efficiency. When the system confirms that the echo signal source is a single target, it can appropriately increase the tolerance to signal noise and interference and ignore the impact of background noise interference on the signal, thereby further improving the accuracy and stability of detection. For situations with less interference, the system can more clearly analyze the behavioral characteristics of the target; for example, when the system determines that the half-height width of the main peak of the echo signal exceeds the preset multiple of the historical single-target waveform width, the system will consider the current echo to be a single target. The signal comes from multiple scattered targets, and the system will detect several small peaks of the half-height width of the main peak. The small peaks specifically include shoulder peak secondary valley, platform waveform secondary valley and peak-to-peak secondary valley, identify the integrated energy on the left and right sides of these small peaks, and dynamically collect the superimposed waveform caused by the reflection amplitude of each target based on the symmetry degree of different integrated energies; the system can effectively distinguish multiple reflection sources in the superimposed signal through fine identification of small peaks in the main peak (such as shoulder peak secondary valley, platform waveform secondary valley, etc.) and analysis of the integrated energy on the left and right sides. Compared with directly judging a wide main peak as a "fuzzy target", this method can more finely distinguish multiple close targets, significantly improve the accuracy of multi-target analysis, avoid missing or misidentifying individual targets due to waveform aliasing, and at the same time, based on each By analyzing the symmetry of the integrated energy on the left and right sides of the small peak, the system can dynamically infer the reflection amplitude of multiple targets in the composite waveform. This approach helps to reconstruct the independent echo characteristics of each target, enabling the system to extract the true reflection intensity, possible material differences or body distribution characteristics of each target, thereby improving the ability to judge the target category, form and behavior in subsequent identification. In complex scenarios where multiple targets are close, overlapping or frequently occluded (such as crowds approaching, walking through corridors or multiple people meeting, etc.), the system can not only maintain the continuity of echo recognition through refined processing of the half-height width of the main peak after it exceeds the limit, but also dynamically adjust the number and distribution information of targets in the perception model, thereby improving the stability and robustness of the system in tracking and classifying multiple targets in dynamic environments.

[0137] It should be noted that several small peaks of the half-maximum width of the main peak are detected, the integrated energy on the left and right sides of the small peak is identified, and based on the degree of symmetry of the integrated energy, the superimposed waveform caused by the corresponding reflection amplitude of each target is dynamically collected. The specific example is as follows:

[0138] Assume that a system monitors multiple moving targets in an environment, using radar signals for target detection. Assume that at a given moment, the echo signal waveform received by the system exhibits a main peak, surrounded by several smaller peaks (such as shoulder-shaped secondary valleys, plateau-shaped secondary valleys, and secondary valleys between peaks). The following step-by-step analysis illustrates how the system uses this information to separate multiple targets and dynamically acquire the reflection amplitude of each target.

[0139] Steps and analysis to calculate the full width at half maximum (FWHM) of the main peak:

[0140] The system first calculates the full width at half maximum (FWHM) of the echo signal's main peak. If this peak width is significantly larger than a preset multiple of the historical single-target waveform width, the system concludes that the signal may be a reflection from multiple targets. For example, if the system's historical single-target waveform width is 1 unit, and the current main peak's full width at half maximum is 2 units, exceeding a preset multiple of the historical width (e.g., 1.5), this indicates that the signal may be a superposition of echoes from multiple targets.

[0141] Identify small peaks:

[0142] Next, the system detects multiple small peaks around the main peak. It assumes that these small peaks are shoulder peak secondary valleys, platform waveform secondary valleys, and inter-peak secondary valleys. The system then identifies the possible sources of the target's echo signal from these small peaks.

[0143] Shoulder peak secondary valley: refers to a small peak that appears on the left side of the main peak. The system finds that the reflected energy distribution of this small peak is more concentrated;

[0144] Platform-shaped waveform secondary valley: refers to the stable area between the main peak and the wave crest, where a gentle small wave crest appears, indicating that there may be a continuous fluctuation of a reflection source;

[0145] Inter-peak secondary valley: refers to the valley point between two peaks, which may correspond to the overlap of echo signals between two targets. The system needs to further analyze its characteristics;

[0146] Calculate the integrated energy:

[0147] For each small peak, the system calculates the integrated energy on its left and right sides, that is, calculates the reflection intensity distribution of the waveform;

[0148] For the shoulder peak secondary valley, suppose the system calculates that the integrated energy on the left side of the small wave peak is 0.8, while the integrated energy on the right side is 0.2; this indicates that the echo of this small peak may come from a distant target;

[0149] For the secondary valley of the platform-shaped waveform, the system found that the integrated energy on the left and right sides was basically equal (0.5), which means that it may come from a relatively uniform reflection source, such as a human body or object at a close distance;

[0150] For the secondary valley between peaks, the system found that the integrated energy difference between the left and right sides was large (0.7 on the left and 0.3 on the right), which indicates that the echo signals reflected by the two targets may overlap and have a large distance difference;

[0151] Analyze the symmetry of the integrated energy:

[0152] The system further analyzes the energy symmetry of each small peak:

[0153] For the shoulder peak secondary valley, due to the energy asymmetry, the system believes that the small peak may come from a target at a distance;

[0154] For the secondary valley of the platform-shaped waveform, since the energy on the left and right sides is symmetrical, the system believes that the echo signal may come from a close-range target with uniform reflection;

[0155] For the secondary valley between peaks, due to the large energy difference between the left and right sides, the system considers it to be the superposition of the echo signals of two targets;

[0156] Dynamic acquisition of reflection amplitude:

[0157] Based on the analysis of the integrated energy, the system dynamically adjusts the reflection amplitude acquisition for each target:

[0158] For targets in the shoulder peak and sub-valley, the system dynamically adjusts the collected reflection amplitude, sets the target as a long-distance target, and reduces the estimation of its reflection amplitude;

[0159] For targets in the secondary valley of the platform-shaped waveform, the system considers that its reflection amplitude is larger, so the reflection amplitude estimation of the target is increased;

[0160] For targets in the secondary valley between peaks, the system dynamically calculates the reflection amplitudes of the two targets and separates their echoes;

[0161] Target echo separation and reconstruction:

[0162] The system can effectively separate the echoes of multiple targets and generate independent waveforms for each target. For example, it can separate the echo waveform of a distant target from the echo of the shoulder peak secondary valley; separate the echo waveform of a close target from the echo of the plateau waveform secondary valley; and separate the echo waveforms of two targets from the echo of the secondary valley between peaks, representing them as two independent targets. Based on the dynamically collected reflection amplitude information, the system further tracks the motion trajectory of each target and identifies and classifies the target type (such as dynamic human body, static obstacle, etc.).

[0163] To summarize, the above examples show that the system can separate the reflection waveforms of multiple targets from complex echo signals, and through integrated energy symmetry analysis, dynamically collect the reflection amplitude of each target, and then reconstruct the echo signals of multiple targets; this process greatly improves the accuracy and real-time performance of radar signal processing, especially in multi-target environments, and can effectively avoid the aliasing of target echoes and provide reliable basic data for subsequent target tracking and classification.

[0164] In this embodiment, the step S4 of determining whether the angular direction of the target distribution point cloud forms a continuous area further includes:

[0165] S41: sorting the target distribution point cloud according to preset angle values, constructing a corresponding angle sequence by sorting, and calculating the angle increment of the angle sequence;

[0166] S42: Determine whether the angle increment increases;

[0167] S43: If so, identify the breakpoint after the angle increment becomes larger, perform cluster cutting at the breakpoint, and use it as the separation boundary of the point cloud target. Based on the position of the breakpoint, divide the point cloud target into several continuous segments, and use the several continuous segments as the angle area of ​​a single candidate target.

[0168] In this embodiment, the system sorts the target distribution point cloud according to the preset angle value, constructs the corresponding angle sequence by sorting, calculates the angle increments of these angle sequences, and then the system determines whether these angle increments have become larger to execute the corresponding steps; for example, when the system determines that the angle increment of the angle sequence has not become larger, the system will consider that the target distribution is relatively uniform or there is a large stable area, and the system will sort the target distribution point cloud according to the angle value to obtain a corresponding angle sequence. Assuming that at a certain moment, the target point cloud of the system is relatively uniformly distributed, and the interval between each point in the angle sequence is small, the angle increments of these angle sequences are calculated, that is, the difference between adjacent angle values. If the target distribution is relatively uniform, the angle increments are relatively consistent and the difference is not large. At the same time, the calculated angle increments reflect the distribution of the target in space. If these angle increments are always consistent or change little, it means that the target distribution is relatively uniform, and the system can accurately judge the distribution characteristics of the target. For example, in a relatively uniform area, the target may have continuous reflection signals, and the angle increment changes little. If the angle increment always remains within a certain range and does not increase, it means that the target distribution is relatively stable, and the target distribution recognized by the system will not have large fluctuations or discrete point clouds. In other words, the relative positions between the targets change little, and there are no dense areas or scattered changes. In this case, the system will consider that the target is in a relatively stable or uniformly distributed state, and may not need further complex processing, such as target separation or noise removal; for example, when the system determines that the angle increment of the angle sequence has become larger, the system will consider that the target distribution is relatively concentrated, and there will be dense areas or scattered changes. The system will identify the breakpoints after the angle increment becomes larger, and perform cluster cutting at the breakpoints as the separation boundaries of the point cloud targets. Based on the positions of these breakpoints, the point cloud targets are divided into several continuous segments, and several continuous segments are used as the angle areas of single candidate targets; by identifying the breakpoints in the angle sequence where the angle increment is significantly larger, the system can accurately locate the distance between targets in the point cloud. The transition boundary effectively separates multiple targets that are close to each other or partially overlap, avoids target confusion, and improves the accuracy of subsequent classification and recognition. At the same time, the mutation characteristics of the angle increment are used as the basis for clustering and cutting, so that the system can flexibly divide the point cloud data into regions when there are multiple targets, especially when there are dense distribution and uneven intervals in angles, thereby enhancing the adaptability of point cloud processing to complex distributions. Moreover, by dividing the entire point cloud into several continuous segments as candidate target areas, the system can process each segment independently in a more targeted manner, such as dynamically setting filtering parameters, assigning tracking numbers, etc., thereby reducing the computational burden of irrelevant areas and improving the overall operation efficiency and real-time performance of the system.

[0169] It should be noted that the breakpoint after the angle increment becomes larger is identified, and cluster cutting is performed at the breakpoint as the separation boundary of the point cloud target. Based on the position of the breakpoint, the point cloud target is divided into several continuous segments, and the several continuous segments are used as the angle region of a single candidate target. The specific example is as follows:

[0170] Consider a radar system monitoring an indoor space. The radar transmits a signal and receives reflected point cloud data. Each point cloud represents the reflected signal from a target and contains information about the target's angle relative to the radar.

[0171] Step 1: Sorting and constructing the angle sequence. First, the system sorts the point cloud data according to the angle value of the target in space based on the reflected signal returned by the radar. Assume that the system obtains the following angle data (unit: degree):

[0172] [12.0, 12.2, 12.4, 12.5, 12.7, 20.1, 20.3, 20.5, 20.6, 30.0, 30.2, 30.3], these angle values ​​represent the angle information of multiple target reflection signals;

[0173] Step 2: Calculate the angle increment. Next, the system calculates the angle increment (angle difference) between each two adjacent points in the angle sequence:

[0174] Incremental sequence: [0.2, 0.2, 0.1, 0.2, 7.4, 0.2, 0.2, 0.1, 9.4, 0.2, 0.1]

[0175] Step 3: Identify the breakpoint. The system sets a threshold (for example, an increment exceeding 1.0° is considered a breakpoint) and identifies the breakpoint based on the angle increment sequence:

[0176] The angle increment between the fifth increment (12.7→20.1) is 7.4°, which is much more than 1.0°, so it is identified as the first break point;

[0177] The angle increment between the 9th increment (20.6→30.0) is 9.4°, which is also much larger than 1.0°, so it is identified as the second break point;

[0178] Step 4: Cluster cutting. Based on the identified breakpoints (the 5th and 9th increments), the system divides the point cloud data into 3 consecutive segments:

[0179] Segment 1 (Target 1): From the angle [12.0, 12.2, 12.4, 12.5, 12.7], this segment has a relatively tight angle and belongs to the first target area;

[0180] Segment 2 (Target 2): From the angle [20.1, 20.3, 20.5, 20.6], the angle values ​​of this segment are relatively close and belong to the second target area;

[0181] Segment 3 (target 3): From angle [30.0, 30.2, 30.3], the angle interval of this segment is slightly larger, but still belongs to the angle range of a single target;

[0182] Step 5: Target segmentation and subsequent processing. After target segmentation, the system treats each continuous angle segment as a separate candidate target; for example:

[0183] Target 1: Located in the angle interval [12.0, 12.7], this may represent the first target in the room (e.g., a person standing at the side of the room);

[0184] Target 2: Located in the angle interval [20.1, 20.6], representing another target (e.g., another person in the room);

[0185] Target 3: Located in the angle interval [30.0, 30.3], representing a distant target (such as an object or person in another corner);

[0186] Step 6: Result Analysis and Application. By dividing these angular regions, the system can separate each target from other targets and use these angular regions as input in subsequent target tracking or analysis. This enables the system to accurately: improve the accuracy of target separation, especially in environments with densely distributed targets; in real-time monitoring and tracking, continuously track and analyze based on these angular regions to avoid confusion between different targets; and in dynamic target recognition, further combine with other sensor information (such as distance, speed, etc.) to optimize target recognition accuracy.

[0187] To summarize, in the above examples, by performing incremental calculations on the angle sequence and identifying the breakpoints where the angle increments become larger, the system can effectively identify the boundaries between multiple targets and segment the point cloud data into multiple target areas. This method not only helps to improve the target separation capability in multi-target scenarios, but also provides more accurate data input for subsequent target recognition, tracking, and classification.

[0188] In this embodiment, based on a detection scenario preset by the microwave radar, step S1 of repeatedly outputting and receiving an echo signal of the detection scenario within a preset time period further includes:

[0189] S11: Adaptively switching the beam coverage strategy of the microwave radar based on a preset scene type of the detection scene, wherein the scene type specifically includes a corridor, a room, and a doorway, and the beam coverage strategy specifically includes omnidirectional, directional, and variable beams;

[0190] S12: Determine whether the beam coverage strategy can meet the preset target dynamic detection;

[0191] S13: If not, dynamically activate the preset static shielding area according to the preset detection parameters of the microwave radar, and block the obstacle echoes pre-collected by the microwave radar through the static shielding area, wherein the detection parameters specifically include the horizontal scanning angle range, the vertical coverage range and the effective distance.

[0192] In this embodiment, the system adaptively switches the beam coverage strategy of the microwave radar based on the pre-set scene type of the detection scene, which specifically includes corridors, rooms and doorways. The beam coverage strategy specifically includes omnidirectional, directional and variable beams. The system then determines whether these beam coverage strategies can meet the pre-set target dynamic detection to execute the corresponding steps; for example, when the system determines that the beam coverage strategy of the microwave radar can meet the pre-set target dynamic detection, the system will consider that the current radar beam configuration has effectively covered the area where the target may appear or move, and has sufficient detection sensitivity and spatial resolution, which can stably achieve the target in the scene. For dynamic behavior perception in the scene, the system will maintain the current omnidirectional, directional or variable beam coverage mode to ensure continuous and stable detection performance. At the same time, if it meets the detection needs and energy saving needs, it can enter low-frequency scanning or intermittent scanning mode to reduce energy consumption. And under the premise that the beam coverage meets the needs, the system will activate subsequent fine tracking functions, such as human posture estimation, micro-motion recognition (such as breathing, gait), etc., to further explore dynamic behavior characteristics; for example, when the system determines that the beam coverage strategy of the microwave radar cannot meet the pre-set target dynamic detection, the system will consider that the current radar beam configuration cannot effectively cover the area where the target may appear. Based on the microwave radar's pre-set detection parameters, which include the horizontal scan angle range, vertical coverage, and effective distance, the system dynamically activates pre-set static shielding zones. These static shielding zones block obstacle echoes pre-collected by the microwave radar. Activating static shielding zones shields echo signals from non-target areas, avoiding interference from obstacles. By setting the horizontal scan angle range, vertical coverage, and effective distance, the system can precisely locate the radar detection area, reduce unnecessary interference areas, and make the radar's detection range more focused, better covering the area required for dynamic target detection. The introduction of static shielding zones effectively reduces the impact of unnecessary obstacle echo signals on the radar system. By blocking obstacle echoes collected by the radar, these echo signals are prevented from being confused with target echoes, ensuring that the system can focus on detecting targets. Shielding zones are particularly important when target detection conditions change or the environment is complex. Dynamically activating static shielding zones allows the radar to make adaptive adjustments based on real-time environmental changes, thereby improving system stability and accuracy in complex environments. By blocking interference signals, the system can more accurately capture target signals, especially in situations with multiple targets or complex backgrounds, which helps improve detection reliability.

[0193] Reference Attachment Figure 2 , a real-time human body detection system based on microwave radar in one embodiment of the present invention, comprising:

[0194] The receiving module 10 is used to repeatedly output and receive the echo signal of the detection scene within a preset time period based on the detection scene preset by the microwave radar;

[0195] A judgment module 20, configured to judge whether a combined signal peak is detected in the echo signal;

[0196] an execution module 30 configured to, if yes, construct a corresponding sparse reconstruction model based on the pre-collected original radar signal, using the original radar signal as superimposed data of a sparse source, decompose the aliased signal into corresponding sparse units using a preset sparse coding, extract multidimensional features from the sparse units, and establish a target distribution point cloud of the multidimensional features in a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features;

[0197] The second judgment module 40 is used to judge whether the angle direction of the target distribution point cloud forms a continuous area;

[0198] The second execution module 50 is used to identify the local point cloud density of the target distribution point cloud if it has not been formed, dynamically estimate the corresponding target number based on the local point cloud density, apply a preset Kalman filter, continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

[0199] In this embodiment, the receiving module 10 repeatedly outputs and receives the echo signal of the detection scene within a preset time period based on the detection scene pre-set by the microwave radar, and then the judgment module 20 judges whether these echo signals have detected the merged signal peak to execute the corresponding steps; for example, when the system determines that the echo signal of the repeated output and reception of the detection scene has not detected the merged signal peak, the system will believe that there is a certain distance separation between the human targets, and the radar echoes do not produce obvious overlap or aliasing in the time, frequency and angle dimensions. The system will continue to maintain the current output and reception cycle, increase the detection capability of weak targets or long-distance targets, consider adjusting the radar's receiving gain or changing the transmission power to enhance long-distance or small targets. At the same time, since there is no merging of signal peaks, it means that the targets corresponding to the echo signals are relatively scattered. The system can directly process each independent echo, extract the multi-dimensional features of each echo (such as distance, speed, Doppler effect, etc.), and perform target identification and positioning in the three-dimensional feature space. If the number of targets is small, the system can identify each target separately through a simple separation method based on distance and Doppler features to ensure the accurate position and motion state of each target. According to the target distribution, the working parameters of the radar system (such as scanning angle, transmission frequency, etc.) are dynamically adjusted to better adapt to the scene requirements. If the system finds that the detection area is relatively static, it can also choose to reduce the output frequency to reduce power consumption; for example, When the system determines that the echo signal of the repeated output and reception detection scene has detected the combined signal peak, the execution module 30 will think that the distance between the human targets is close, which is easy to cause the radar echo to overlap or alias. The system will use the original radar signals collected in advance as the superposition data of sparse sources to build a corresponding sparse reconstruction model, and use the pre-set sparse coding to decompose the aliased signal into corresponding sparse units, and extract multi-dimensional features from these sparse units. The multi-dimensional features specifically include distance features, Doppler features and angle features, and establish a target distribution point cloud with multi-dimensional features through a pre-set three-dimensional feature space; the system regards the original radar signal as the superposition data of multiple sparse sources, and uses sparse The reconstruction model performs signal deconstruction and can effectively separate the merged signal into sparse units with identifiable features, thus overcoming the limitation of traditional methods in distinguishing individuals in close-range multi-person detection. At the same time, the three-dimensional features of distance, Doppler and angle are extracted from the sparse units to form richer target information, which is conducive to distinguishing targets from multiple dimensions. In particular, when the targets have similar shapes and similar movement directions, the accuracy of target classification and recognition is improved. In addition, the target distribution point cloud established through the three-dimensional feature space can intuitively display the relative position and density relationship of the targets in space, providing the basic data structure for subsequent operations such as target number estimation, trajectory separation and dynamic tracking, and improving the system's multi-target management capabilities in crowded or occluded scenes.Then the second judgment module 40 judges whether the angular directions of these target distribution point clouds form a continuous area to execute the corresponding steps; for example, when the system determines that the angular directions of the target distribution point clouds form a continuous area, the system will believe that relatively complete target reflection information is detected in this angular area, which can more accurately represent the directional expansion range of the target in space. The system will adjust the processing method of the echo signal and merge the echo signals in the continuous area to avoid unnecessary signal separation. At this time, the target recognition strategy should focus on the analysis of group behavior rather than the separation of individual targets in order to improve processing efficiency. At the same time, according to information such as density analysis, motion patterns and multidimensional features, the number of targets is dynamically estimated, and appropriate classification algorithms (such as clustering-based classification methods) are used to classify and identify these targets. The dynamic position changes of the targets can be used to more accurately judge the relationship between targets to avoid Avoid misidentification in multi-target scenarios, and for occasional continuous angular direction areas, the system should carefully eliminate noise interference, and ensure that the continuity of target distribution is not due to misjudgment caused by interference signals or environmental noise by setting reasonable thresholds and signal enhancement mechanisms; for example, when the system determines that the angular direction of the target distribution point cloud does not form a continuous area, the second execution module 50 will believe that no complete target reflection information is detected in this angular area, and it is impossible to represent the directional expansion range of the target in space. The system will identify the local point cloud density of the target distribution point cloud, and dynamically estimate the corresponding number of human targets based on different local point cloud densities, apply a pre-set Kalman filter, continuously track the target parameters of each human target, the target parameters specifically include target speed and target speed change, generate separated multi-target echoes, and reconstruct these multi-target echoes into a pseudo-color radar image;By further analyzing the local point cloud density of the target distribution point cloud, the system can capture the presence of multiple high-density reflection sub-regions even when there is a lack of angular extension, thereby dynamically estimating the actual number of targets present. This density-based supplementary mechanism significantly improves the system's recognition robustness in weak signal areas or non-ideal reflection conditions. At the same time, when multiple targets are close in distance or move at similar speeds, their radar echoes often overlap in time, frequency, and angle, making it difficult to distinguish instantaneous targets. In this case, a Kalman filter is introduced. Based on the target's historical motion state and combined with new observations, it continuously estimates and corrects the velocity and velocity change parameters of each target. This not only effectively avoids target trajectory interruptions caused by short-term aliasing, but also maintains target trajectory consistency in the face of partial occlusion or signal loss. After completing sparse reconstruction, feature extraction, point cloud generation, and target tracking for multiple targets, the system synthesizes the separated target echo signals into a pseudo-color radar image. Each target is assigned a different color code based on its characteristic parameters. This visualization method improves the readability of the system results and the efficiency of human-computer interaction, allowing end users to intuitively identify the number, location, and movement direction of targets. ;

[0200] In this embodiment, the execution module further includes:

[0201] a calculation unit, configured to calculate the amplitude square integral of the sparse unit, normalize the amplitude square integral to a preset standard scale, apply a preset continuous wavelet transform to the sparse unit, and generate a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate;

[0202] a judging unit, configured to judge whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat;

[0203] An execution unit is configured to, if yes, collect the target distance corresponding to the echo delay of the sparse unit based on the echo delay, count the internal point distribution form of the sparse unit according to the target distance, and track the time trajectory evolution of the sparse unit within a preset time window, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

[0204] In this embodiment, the system calculates the amplitude square integral of the sparse unit, normalizes these amplitude square integrals to a preset standard scale, applies a preset continuous wavelet transform to these sparse units, and generates a corresponding time-frequency spectrogram. The time-frequency spectrogram specifically includes the main frequency, bandwidth and frequency change rate. Then the system determines whether the time-frequency spectrogram detects the preset low-frequency periodic motion characteristics, which specifically include breathing and heartbeat, to execute the corresponding steps; for example, when the system determines that the time-frequency spectrogram does not detect the preset low-frequency periodic motion characteristics, the system will believe that multiple human body micro-motion signals are superimposed on each other in the sparse unit, resulting in a special If the energy in the eigenfrequency band is diffused or the waveform is distorted, the low-frequency features cannot be clearly identified. The system will temporarily classify the reflection target corresponding to the current sparse unit as a "non-biological target" or "inactive target", and perform target classification and filtering based on the position and speed characteristics to avoid mistaking it for a human object that can interact or require an alarm. At the same time, if the target is located in the area of ​​interest (such as a bed, sofa, or monitoring range), the analysis period will be automatically extended (such as from 5 seconds to 10 seconds), and time-frequency analysis will be performed again to improve the weak signal detection capability. If no low-frequency periodic features are detected in multiple sparse units, you can try to adjust the radar gain, antenna beam, or activate auxiliary channels (such as vertical beam groups) to improve detection sensitivity. If such "targets without low-frequency features" appear repeatedly, they should be marked and uploaded to the cloud or back-end database as part of behavioral risk identification (such as sudden fainting or abnormal stillness); for example, when the system determines that the time-frequency spectrum detects the pre-set low-frequency periodic motion characteristics, the system will think that the radar signal quality is good, the micro-motion components (such as breathing, heartbeat, etc.) are not masked by noise or interference, the system can clearly distinguish the target's vital signs, and the signal processing effect is better. The system will collect the target distance corresponding to the echo delay based on the sparse unit, and statistically calculate the internal point distribution of the sparse unit according to different target distances. The morphology specifically includes the main axis length, flatness and density distribution characteristics. Within a pre-set time window, the time trajectory evolution of these sparse units is tracked. Based on the echo delay of the sparse unit, the system can collect the delay information of the echo signal, thereby accurately calculating the distance to the target. Through the echo delay, the system can determine the position of the target. Especially when the target is at different distances, the system can dynamically adjust its detection range to further improve the tracking accuracy. At the same time, the internal point distribution morphology of the sparse unit (including the main axis length, flatness, density distribution characteristics, etc.) is statistically analyzed to further reveal the morphological characteristics of the target. For example, the density distribution characteristics of the target can help to infer the size, shape and motion state of the target.Changes in the main axis length and flatness can be used to determine the target's trajectory and posture. Within a set time window, the system can track the temporal trajectory evolution of sparse units. By continuously tracking the target's trajectory, the system can obtain the target's motion trends (such as speed and direction changes) in real time, allowing for dynamic monitoring and prediction of the target. This is particularly important for applications such as smart homes, health monitoring, and security systems.

[0205] In this embodiment, it also includes:

[0206] an identification module, configured to identify the number of neighboring points of each point cloud within a preset radius based on the point cloud density of the target distribution point cloud;

[0207] A third judgment module is used to judge whether the number of neighbor points is less than a preset value;

[0208] The third execution module is used to detect the extreme outliers of the multidimensional features, evaluate the discreteness of each point cloud according to the extreme outliers, count the topological characteristics of the target distribution point cloud, and add additional features to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

[0209] In this embodiment, the system identifies the number of neighboring points of each point cloud within a preset radius based on the point cloud density of the target distribution point cloud, and then the system determines whether the number of neighboring points is less than a preset value to execute the corresponding steps; for example, when the system determines that the number of neighboring points of each point cloud within a preset radius is not less than a preset value, the system will consider that the current point cloud has a high local density and stability and does not belong to discrete noise points. There is no need to remove noise points. The system will use a clustering algorithm to cluster these dense point cloud areas and distinguish different targets. The clustering results can help the system identify different targets and at the same time The clustered point cloud can be extracted to extract its geometric shape, boundary, center of mass position and other features, and the direction, volume and shape of the target can be further analyzed. If there is time series data, the system can calculate the target's motion parameters, such as speed, acceleration, direction, etc., based on the target's changing trajectory, and use algorithms such as Kalman filtering to track the target and further predict the target's future position. Based on the target's geometric features and dynamic changes, classification algorithms (such as deep learning models, support vector machines, etc.) are used for target recognition, such as identifying whether the target is a pedestrian, a stationary object or other dynamic object; for example, when the system determines that each point cloud is within a pre-set radius, If the number of neighbor points is less than the preset value, the system will consider the current point cloud to be unstable and may be discrete noise points, which need to be removed. The system will detect the extreme outliers of multi-dimensional features, which include excessive reflection energy and frequency jumps. Based on these extreme outliers, the discrete degree of each point cloud is evaluated, and the topological characteristics of the target distribution point cloud are statistically analyzed. The topological characteristics include the main direction of the point cloud and the overall shape of the point cloud. Additional features are attached to each point cloud, including local curvature and local velocity gradient. The system can determine the number of neighbor points of the point cloud within the preset radius, combined with the extreme outliers of reflection energy and frequency, to improve the point cloud’s stability. It accurately distinguishes valid target points from discrete noise points, thereby avoiding the accidental deletion of edge targets or weakly reflective targets, improving the purity and credibility of point cloud data, and further statistically analyzing the topological characteristics of the point cloud (such as the main direction and overall shape) to extract additional features such as local curvature and velocity gradient, which helps to accurately model the geometry and motion form of the target, improve the target's multi-dimensional feature recognition ability and the stability of continuous tracking. With the help of multi-dimensional outlier detection and topological analysis, the system can not only eliminate invalid point clouds, but also identify abnormal point clouds with characteristics such as violent reflection and speed mutation, which helps to warn of potential safety risks or atypical behaviors and enhance the system's intelligent monitoring and response capabilities.

[0210] In this embodiment, the second execution module further includes:

[0211] A generating unit, configured to identify continuous frame point clouds from the target distribution point cloud, perform temporal fusion on the continuous frame point clouds, and generate corresponding point cloud high-density areas;

[0212] A second judgment unit is used to judge whether the high-density area of ​​the point cloud appears repeatedly within a preset frequency;

[0213] The second execution unit is configured to obtain the position information of the high-density area of ​​the point cloud, extract preset micro-motion spectrum features from the high-density area of ​​the point cloud, and dynamically divide the target types of the continuous frame point clouds based on the position information and the micro-motion spectrum features, wherein the micro-motion spectrum features specifically include breathing frequency and arm swing frequency, and the target types specifically include dynamic human bodies, interference objects, and non-human objects.

[0214] In this embodiment, the system identifies continuous frame point clouds from the target distribution point cloud, performs time-series fusion on these continuous frame point clouds, generates corresponding point cloud high-density areas, and then the system determines whether the point cloud high-density area recurs within a preset frequency to execute the corresponding steps; for example, when the system determines that the point cloud high-density area does not recur within the preset frequency, the system will believe that the point cloud changes in the area may not have obvious stability or periodicity, and there may be no continuous targets. The system will identify whether there are large changes or sudden fluctuations in the point cloud in the area. If the high-density area does not recur within the predetermined frequency, it means that the target may have moved out of the current detection range, or the target is masked by noise. At the same time, for high-density areas that do not recur, the system needs to update the target status, perform target tracking or leave the current area status update, which involves adjusting the target's motion model. and path prediction to avoid mistakenly tracking a non-existent or inactive target, and if the high-density area does not reappear, the system needs to further filter out potential environmental noise, eliminate false alarms, and dynamically adjust the sensitivity of the sensor, optimize the radar signal processing algorithm, reduce the interference of irrelevant factors on target detection, and improve the system's sensitivity and accuracy to valid targets; for example, when the system determines that a high-density area of ​​the point cloud reappears within a preset frequency, the system will consider that the point cloud changes in this area have obvious stability or periodicity, and there are continuous targets. The system will obtain the position information of the high-density area of ​​the point cloud, and extract the pre-set micro-motion spectrum features from the high-density area of ​​the point cloud. The micro-motion spectrum features specifically include breathing frequency and arm swing frequency. Based on the position information and micro-motion spectrum features, the target types of continuous frame point clouds are dynamically divided. The target types specifically include dynamic human bodies, interference objects, and non-human objects.The system can effectively distinguish dynamic human bodies from interference objects or non-human objects by extracting micro-motion spectrum features. For example, the human breathing frequency has a specific low-frequency stable feature, while the arm swing frequency reflects a rhythmic movement pattern. These features are obviously different in the spectrum. Compared with the judgment method that simply relies on static form or position point cloud, this method of integrating micro-motion features can significantly improve the accuracy of identifying dynamic human bodies, and avoid misidentifying wind-blown leaves, swinging curtains, etc. as human targets. At the same time, when high-density point cloud areas continue to repeat in frequency, the system can judge that the target is a long-term active or stationary object, rather than an instantaneous interference. This judgment based on frequency stability is helpful The system prioritizes processing resources for real, persistent human targets, improving overall monitoring efficiency and real-time response capabilities. This is particularly applicable to scenarios such as personnel retention monitoring and security deployment. Furthermore, the system integrates location information and micro-motion spectrum data to dynamically segment targets within continuous frame point clouds, identifying, for example, walking humans, periodic interference sources (such as fans), and stationary non-human objects (such as pillars and sofas). This classification approach, based on "target behavior + micro-motion features," empowers the system with enhanced target understanding capabilities, advancing the transition from "point cloud detection" to "semantic understanding," providing a solid foundation for subsequent behavior recognition, intelligent decision-making, and early warning control.

[0215] In this embodiment, the judgment module further includes:

[0216] An acquisition unit, configured to acquire the full width at half maximum of the main peak of the echo signal;

[0217] A third judgment unit is used to judge whether the half-maximum width of the main peak exceeds a preset multiple of the historical single target waveform width;

[0218] The third execution unit is used to detect several small peaks of the half-height width of the main peak, identify the integrated energy on the left and right sides of the small peak, and dynamically collect the superimposed waveform caused by the reflection amplitude of each target based on the degree of symmetry of the integrated energy, wherein the small peaks specifically include shoulder peak secondary valleys, platform waveform secondary valleys and inter-peak secondary valleys.

[0219] In this embodiment, the system obtains the half-width of the main peak of the echo signal, and then determines whether the half-width of the main peak exceeds a preset multiple of the historical single-target waveform width to execute the corresponding steps; for example, when the system determines that the half-width of the main peak of the echo signal does not exceed a preset multiple of the historical single-target waveform width, the system will believe that the current echo signal may come from a relatively concentrated single target, whose signal waveform is relatively stable and has not undergone excessive expansion or overlap. The system will confirm that the current echo signal comes from a single target, and perform single-target tracking and analysis, extract relevant information of the target, such as position, speed, motion trajectory, etc., and further predict the target's movement trend and possible behavioral changes. At the same time, since the echo signal indicates that the target is single and clear The system chooses to optimize the target detection algorithm to improve its tracking accuracy, such as using the Kalman filter or other tracking algorithms to accurately track the target to avoid misjudgment caused by multi-target interference. The system can reduce the need for complex calculations and focus on tracking and analyzing a single target to improve efficiency. When the system confirms that the echo signal source is a single target, it can appropriately increase the tolerance to signal noise and interference and ignore the impact of background noise interference on the signal, thereby further improving the accuracy and stability of detection. For situations with less interference, the system can more clearly analyze the behavioral characteristics of the target; for example, when the system determines that the half-height width of the main peak of the echo signal exceeds the preset multiple of the historical single-target waveform width, the system will consider the current echo to be a single target. The signal comes from multiple scattered targets, and the system will detect several small peaks of the half-height width of the main peak. The small peaks specifically include shoulder peak secondary valley, platform waveform secondary valley and peak-to-peak secondary valley, identify the integrated energy on the left and right sides of these small peaks, and dynamically collect the superimposed waveform caused by the reflection amplitude of each target based on the symmetry degree of different integrated energies; the system can effectively distinguish multiple reflection sources in the superimposed signal through fine identification of small peaks in the main peak (such as shoulder peak secondary valley, platform waveform secondary valley, etc.) and analysis of the integrated energy on the left and right sides. Compared with directly judging a wide main peak as a "fuzzy target", this method can more finely distinguish multiple close targets, significantly improve the accuracy of multi-target analysis, avoid missing or misidentifying individual targets due to waveform aliasing, and at the same time, based on each By analyzing the symmetry of the integrated energy on the left and right sides of the small peak, the system can dynamically infer the reflection amplitude of multiple targets in the composite waveform. This approach helps to reconstruct the independent echo characteristics of each target, enabling the system to extract the true reflection intensity, possible material differences or body distribution characteristics of each target, thereby improving the ability to judge the target category, form and behavior in subsequent identification. In complex scenarios where multiple targets are close, overlapping or frequently occluded (such as crowds approaching, walking through corridors or multiple people meeting, etc.), the system can not only maintain the continuity of echo recognition through refined processing of the half-height width of the main peak after it exceeds the limit, but also dynamically adjust the number and distribution information of targets in the perception model, thereby improving the stability and robustness of the system in tracking and classifying multiple targets in dynamic environments.

[0220] In this embodiment, the second judgment module further includes:

[0221] A second calculation unit is configured to sort the target distribution point cloud according to preset angle values, construct a corresponding angle sequence by sorting, and calculate an angle increment of the angle sequence;

[0222] a fourth judging unit, configured to judge whether the angle increment increases;

[0223] The fourth execution unit is used to identify the breakpoint after the angle increment becomes larger, perform cluster cutting at the breakpoint as the separation boundary of the point cloud target, divide the point cloud target into several continuous segments based on the position of the breakpoint, and use the several continuous segments as the angle area of ​​a single candidate target.

[0224] In this embodiment, the system sorts the target distribution point cloud according to the preset angle value, constructs the corresponding angle sequence by sorting, calculates the angle increments of these angle sequences, and then the system determines whether these angle increments have become larger to execute the corresponding steps; for example, when the system determines that the angle increment of the angle sequence has not become larger, the system will consider that the target distribution is relatively uniform or there is a large stable area, and the system will sort the target distribution point cloud according to the angle value to obtain a corresponding angle sequence. Assuming that at a certain moment, the target point cloud of the system is relatively uniformly distributed, and the interval between each point in the angle sequence is small, the angle increments of these angle sequences are calculated, that is, the difference between adjacent angle values. If the target distribution is relatively uniform, the angle increments are relatively consistent and the difference is not large. At the same time, the calculated angle increments reflect the distribution of the target in space. If these angle increments are always consistent or change little, it means that the target distribution is relatively uniform, and the system can accurately judge the distribution characteristics of the target. For example, in a relatively uniform area, the target may have continuous reflection signals, and the angle increment changes little. If the angle increment always remains within a certain range and does not increase, it means that the target distribution is relatively stable, and the target distribution recognized by the system will not have large fluctuations or discrete point clouds. In other words, the relative positions between the targets change little, and there are no dense areas or scattered changes. In this case, the system will consider that the target is in a relatively stable or uniformly distributed state, and may not need further complex processing, such as target separation or noise removal; for example, when the system determines that the angle increment of the angle sequence has become larger, the system will consider that the target distribution is relatively concentrated, and there will be dense areas or scattered changes. The system will identify the breakpoints after the angle increment becomes larger, and perform cluster cutting at the breakpoints as the separation boundaries of the point cloud targets. Based on the positions of these breakpoints, the point cloud targets are divided into several continuous segments, and several continuous segments are used as the angle areas of single candidate targets; by identifying the breakpoints in the angle sequence where the angle increment is significantly larger, the system can accurately locate the distance between targets in the point cloud. The transition boundary effectively separates multiple targets that are close to each other or partially overlap, avoids target confusion, and improves the accuracy of subsequent classification and recognition. At the same time, the mutation characteristics of the angle increment are used as the basis for clustering and cutting, so that the system can flexibly divide the point cloud data into regions when there are multiple targets, especially when there are dense distribution and uneven intervals in angles, thereby enhancing the adaptability of point cloud processing to complex distributions. Moreover, by dividing the entire point cloud into several continuous segments as candidate target areas, the system can process each segment independently in a more targeted manner, such as dynamically setting filtering parameters, assigning tracking numbers, etc., thereby reducing the computational burden of irrelevant areas and improving the overall operation efficiency and real-time performance of the system.

[0225] In this embodiment, the receiving module further includes:

[0226] a switching unit, configured to adaptively switch the beam coverage strategy of the microwave radar based on a scene type preset in the detection scene, wherein the scene type specifically includes a corridor, a room, and a doorway, and the beam coverage strategy specifically includes omnidirectional, directional, and variable beams;

[0227] A fifth judgment unit, configured to judge whether the beam coverage strategy can satisfy preset target dynamic detection;

[0228] The fifth execution unit is used to dynamically activate the preset static shielding area according to the detection parameters preset by the microwave radar, and block the obstacle echo pre-collected by the microwave radar through the static shielding area, wherein the detection parameters specifically include the horizontal scanning angle range, the vertical coverage range and the effective distance.

[0229] In this embodiment, the system adaptively switches the beam coverage strategy of the microwave radar based on the pre-set scene type of the detection scene, which specifically includes corridors, rooms and doorways. The beam coverage strategy specifically includes omnidirectional, directional and variable beams. The system then determines whether these beam coverage strategies can meet the pre-set target dynamic detection to execute the corresponding steps; for example, when the system determines that the beam coverage strategy of the microwave radar can meet the pre-set target dynamic detection, the system will consider that the current radar beam configuration has effectively covered the area where the target may appear or move, and has sufficient detection sensitivity and spatial resolution, which can stably achieve the target in the scene. For dynamic behavior perception in the scene, the system will maintain the current omnidirectional, directional or variable beam coverage mode to ensure continuous and stable detection performance. At the same time, if it meets the detection needs and energy saving needs, it can enter low-frequency scanning or intermittent scanning mode to reduce energy consumption. And under the premise that the beam coverage meets the needs, the system will activate subsequent fine tracking functions, such as human posture estimation, micro-motion recognition (such as breathing, gait), etc., to further explore dynamic behavior characteristics; for example, when the system determines that the beam coverage strategy of the microwave radar cannot meet the pre-set target dynamic detection, the system will consider that the current radar beam configuration cannot effectively cover the area where the target may appear. Based on the microwave radar's pre-set detection parameters, which include the horizontal scan angle range, vertical coverage, and effective distance, the system dynamically activates pre-set static shielding zones. These static shielding zones block obstacle echoes pre-collected by the microwave radar. Activating static shielding zones shields echo signals from non-target areas, avoiding interference from obstacles. By setting the horizontal scan angle range, vertical coverage, and effective distance, the system can precisely locate the radar detection area, reduce unnecessary interference areas, and make the radar's detection range more focused, better covering the area required for dynamic target detection. The introduction of static shielding zones effectively reduces the impact of unnecessary obstacle echo signals on the radar system. By blocking obstacle echoes collected by the radar, these echo signals are prevented from being confused with target echoes, ensuring that the system can focus on detecting targets. Shielding zones are particularly important when target detection conditions change or the environment is complex. Dynamically activating static shielding zones allows the radar to make adaptive adjustments based on real-time environmental changes, thereby improving system stability and accuracy in complex environments. By blocking interference signals, the system can more accurately capture target signals, especially in situations with multiple targets or complex backgrounds, which helps improve detection reliability.

[0230] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A real-time human body detection method based on microwave radar, characterized in that: The following steps are involved: Based on a detection scene preset by the microwave radar, repeatedly outputting and receiving the echo signal of the detection scene within a preset time period; Determining whether a combined signal peak is detected in the echo signal; If so, based on the pre-collected original radar signal, the original radar signal is used as superimposed data of the sparse source to construct a corresponding sparse reconstruction model, and a preset sparse coding is used to decompose the aliased signal into corresponding sparse units. Multidimensional features are extracted from the sparse units, and a target distribution point cloud of the multidimensional features is established through a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features; Determining whether the angular direction of the target distribution point cloud forms a continuous area; If not, the local point cloud density of the target distribution point cloud is identified, and the corresponding target number is dynamically estimated based on the local point cloud density. A preset Kalman filter is applied to continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

2. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: The step of decomposing the aliased signal into corresponding sparse units by using a preset sparse coding and extracting multidimensional features from the sparse units further includes: Calculating the amplitude square integral of the sparse unit, normalizing the amplitude square integral to a preset standard scale, applying a preset continuous wavelet transform to the sparse unit, and generating a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate; Determining whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat; If so, based on the echo delay of the sparse unit, the target distance corresponding to the echo delay is collected, and according to the target distance, the internal point distribution form of the sparse unit is counted, and within a preset time window, the time trajectory evolution of the sparse unit is tracked, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

3. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: After the step of establishing the target distribution point cloud of the multi-dimensional feature through the preset three-dimensional feature space, the method further includes: Based on the point cloud density of the target distributed point cloud, identifying the number of neighboring points of each point cloud within a preset radius; Determine whether the number of neighbor points is less than a preset value; If so, the extreme outliers of the multidimensional features are detected, and based on the extreme outliers, the discreteness of each point cloud is evaluated, the topological characteristics of the target distribution point cloud are counted, and additional features are added to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

4. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: The step of identifying the local point cloud density of the target distribution point cloud and dynamically estimating the corresponding number of targets based on the local point cloud density further includes: Identifying continuous frame point clouds from the target distribution point cloud, performing temporal fusion on the continuous frame point clouds, and generating corresponding point cloud high-density areas; Determining whether the high-density area of ​​the point cloud appears repeatedly within a preset frequency; If so, the position information of the high-density area of ​​the point cloud is obtained, the preset micro-motion spectrum features are extracted from the high-density area of ​​the point cloud, and the target types of the continuous frame point clouds are dynamically divided based on the position information and the micro-motion spectrum features, wherein the micro-motion spectrum features specifically include breathing frequency and arm swing frequency, and the target types specifically include dynamic human bodies, interference objects and non-human objects.

5. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: The step of determining whether the echo signal has detected a combined signal peak further includes: Obtaining the full width at half maximum of the main peak of the echo signal; Determining whether the main peak half-width exceeds a preset multiple of the historical single target waveform width; If so, several small peaks of the half-width at half maximum of the main peak are detected, the integrated energy on the left and right sides of the small peak is identified, and based on the degree of symmetry of the integrated energy, the superimposed waveform caused by the corresponding reflection amplitude of each target is dynamically collected, wherein the small peaks specifically include shoulder peak secondary valleys, platform waveform secondary valleys and inter-peak secondary valleys.

6. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: The step of determining whether the angular direction of the target distribution point cloud forms a continuous area further includes: Sorting the target distribution point cloud according to preset angle values, constructing a corresponding angle sequence by sorting, and calculating the angle increment of the angle sequence; Determining whether the angle increment increases; If so, identify the breakpoint after the angle increment becomes larger, perform cluster cutting at the breakpoint, and use it as the separation boundary of the point cloud target. Based on the position of the breakpoint, divide the point cloud target into several continuous segments, and use the several continuous segments as the angle area of ​​a single candidate target.

7. The real-time human body detection method based on microwave radar according to claim 1, characterized in that: The step of repeatedly outputting and receiving the echo signal of the detection scene within a preset time period based on the detection scene preset by the microwave radar further includes: Adaptively switching the beam coverage strategy of the microwave radar based on a preset scene type of the detection scene, wherein the scene type specifically includes a corridor, a room, and a doorway, and the beam coverage strategy specifically includes omnidirectional, directional, and variable beams; Determine whether the beam coverage strategy can meet the preset target dynamic detection; If not, the preset static shielding area is dynamically activated according to the preset detection parameters of the microwave radar, and the obstacle echoes pre-collected by the microwave radar are blocked through the static shielding area, wherein the detection parameters specifically include the horizontal scanning angle range, vertical coverage range and effective distance.

8. A real-time human body detection system based on microwave radar, characterized in that: include: A receiving module is used to repeatedly output and receive echo signals of a detection scene preset by the microwave radar within a preset time period; A judgment module, used to judge whether the echo signal detects a combined signal peak; an execution module, configured to, if so, construct a corresponding sparse reconstruction model based on pre-collected original radar signals, using the original radar signals as superimposed data of sparse sources, decompose the aliased signals into corresponding sparse units using preset sparse coding, extract multidimensional features from the sparse units, and establish a target distribution point cloud of the multidimensional features through a preset three-dimensional feature space, wherein the multidimensional features specifically include distance features, Doppler features, and angle features; A second judgment module is used to judge whether the angle direction of the target distribution point cloud forms a continuous area; The second execution module is used to identify the local point cloud density of the target distribution point cloud if it has not been formed, dynamically estimate the corresponding target number based on the local point cloud density, apply a preset Kalman filter, continuously track the target parameters of each target, generate separated multi-target echoes, and reconstruct the multi-target echoes into a pseudo-color radar image, wherein the target parameters specifically include target speed and target speed change.

9. The real-time human body detection system based on microwave radar according to claim 8, characterized in that: The execution module also includes: a calculation unit, configured to calculate the amplitude square integral of the sparse unit, normalize the amplitude square integral to a preset standard scale, apply a preset continuous wavelet transform to the sparse unit, and generate a corresponding time-frequency spectrogram, wherein the time-frequency spectrogram specifically includes a main frequency, a bandwidth, and a frequency change rate; a judging unit, configured to judge whether a preset low-frequency periodic motion feature is detected in the time-frequency spectrogram, wherein the low-frequency periodic motion feature specifically includes breathing and heartbeat; An execution unit is configured to, if yes, collect the target distance corresponding to the echo delay of the sparse unit based on the echo delay, count the internal point distribution form of the sparse unit according to the target distance, and track the time trajectory evolution of the sparse unit within a preset time window, wherein the internal point distribution form specifically includes the main axis length, flatness and density distribution characteristics.

10. The real-time human body detection system based on microwave radar according to claim 8, characterized in that: Also includes: an identification module, configured to identify the number of neighboring points of each point cloud within a preset radius based on the point cloud density of the target distribution point cloud; A third judgment module is used to judge whether the number of neighbor points is less than a preset value; The third execution module is used to detect the extreme outliers of the multidimensional features, evaluate the discreteness of each point cloud according to the extreme outliers, count the topological characteristics of the target distribution point cloud, and add additional features to each point cloud, wherein the extreme outliers specifically include excessive reflection energy and frequency jumps, the topological characteristics specifically include the main direction of the point cloud and the overall shape of the point cloud, and the additional features specifically include local curvature and local velocity gradient.

Images