Rotary millimeter wave radar human body tracking and recognition device and method

Abstract

The invention belongs to the technical field of radar human body tracking and recognition, and particularly relates to a rotary millimeter wave radar human body tracking and recognition device and method. The millimeter-wave radar module is mounted on the rotating holder module, and the main beam direction of an antenna in the millimeter-wave radar module is perpendicular to the rotating shaft; the millimeter wave radar module is connected with the control module, and the embedded processor is internally provided with a point cloud data processing algorithm and an identity recognition neural network which are used for overcoming the influence of environmental noise and multipath effect on data quality. The problems of fixed field angle and limited monitoring range of the traditional millimeter wave radar are effectively solved; and through an improved data processing algorithm and an identity recognition neural network, the influence of environmental noise and a multipath effect on data quality is overcome, and the reliability and adaptability of the system in a complex scene are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar human body tracking and identification, and in particular relates to a rotary millimeter wave radar human body tracking and identification device and method. Background Art

[0002] With the rapid development of the Internet of Things and intelligent sensing technology, the importance of real-time human tracking and identity recognition in the fields of smart home, security monitoring, health management, etc. has become increasingly prominent. At present, many sensing methods have been applied to this, including vision-based methods, wearable devices and other RF-based methods (such as Wi-Fi). However, these technologies have obvious shortcomings in specific scenarios, and millimeter-wave radar technology has gradually attracted attention due to its unique advantages. In environments with dim light, smoke or obstructed vision, the detection effect of vision-based methods is significantly reduced; in addition, monitoring methods that rely on image and video data may cause privacy protection issues in continuous monitoring. In contrast, millimeter-wave radar has low dependence on ambient light and strong penetration. The collected point cloud data information does not expose the specific details of the human body, and can provide higher privacy protection during the monitoring process. Although wearable devices can obtain individual information, they usually require the monitored person to actively cooperate, making it difficult to achieve non-contact and unconscious monitoring, while millimeter-wave radar can achieve target tracking and identification without active cooperation. Compared with other RF-based methods such as Wi-Fi, millimeter-wave radar has higher resolution and stronger environmental adaptability, and can provide more accurate detection results in complex scenes.

[0003] Although millimeter-wave radar technology has many advantages, there are still certain technical limitations in practical applications. Due to the operation in the high frequency band, the inherent physical characteristics of the millimeter-wave radar antenna itself lead to its strong directionality and narrow beam width, which is specifically manifested in its limited field of view. This limitation not only weakens the radar's panoramic monitoring capability in a large range of scenes, but also when the target is out of the field of view, the radar cannot continue to track, affecting the continuous identification and tracking of the target. The traditional solution is usually to deploy multiple radar devices to work together to expand the coverage range, but this not only greatly increases the system cost, but also puts strict requirements on the synchronization between devices, significantly increasing the complexity of implementation. Another solution is to use a rotating gimbal.

[0004] The following problems exist when using a rotating pan-tilt platform: the point cloud data generated by the millimeter-wave radar is relatively sparse, and it is difficult to use it directly for high-precision dynamic target tracking and identity recognition. Sparse point cloud data is difficult to provide sufficient local and global feature information, which affects the accuracy of feature extraction and classification; and due to the different movement speeds and directions of different parts of the human body, the point cloud density changes in different parts of the body, resulting in unstable feature extraction and difficulty in capturing consistent movement patterns. At the same time, when the target deviates from the front of the radar at a large angle, even if it is still within the detection range, the signal strength will be significantly weakened, making the sparse point cloud data more unreliable, which will have an adverse effect on subsequent judgments. In addition, in complex environments, millimeter-wave radars are susceptible to multipath signal interference. When the radar signal is reflected multiple times on the surface of an object in the environment, the multipath signal and the direct reflection signal are superimposed on each other, which may introduce false targets and noise, further reducing the quality of the point cloud data. The multipath effect also causes the spatial distribution of the point cloud data to become irregular, increasing the difficulty of feature extraction, thereby affecting the detection accuracy and overall reliability of the system. Summary of the invention

[0005] In order to solve the technical limitations of the above-mentioned millimeter-wave radar, the present invention proposes a rotating millimeter-wave radar human body tracking and recognition device, which solves the problem of small field of view angle of existing millimeter-wave radar by integrating a rotating pan-tilt platform.

[0006] The present invention also proposes a rotating millimeter-wave radar human body tracking and identification method. By designing a pan-tilt control algorithm, the pan-tilt can dynamically adjust the radar's viewing angle according to the target position to ensure that the target is always within the monitoring range, further improving the tracking efficiency and flexibility of the system; at the same time, an improved data processing algorithm and neural network model are introduced to effectively suppress the influence of multipath interference and improve the accuracy of identity recognition, thereby significantly improving the stability and overall performance of the system.

[0007] The technical solution provided by the present invention is a rotating millimeter-wave radar human body tracking and identification device, comprising: a millimeter-wave radar module, which adopts a 77GHz-81GHz millimeter-wave radar based on frequency modulation continuous wave technology, and is used to transmit and receive millimeter-wave signals and collect point cloud data of the target; a rotating pan-tilt module, which adopts a servo motor, supports high-precision, high-speed rotation and precise angle control, and the single rotation angle of the pan-tilt is 0~180°; a control module, which adopts an embedded processor with high computing power and low power consumption, and is suitable for performing complex real-time signal processing and target identification tasks; the millimeter-wave radar module is installed on the rotating pan-tilt module, and the main beam direction of the antenna in the millimeter-wave radar module is perpendicular to the rotation axis; the millimeter-wave radar module is connected to the control module, and the embedded processor has a built-in point cloud data processing algorithm and an identity recognition neural network for overcoming the influence of environmental noise and multipath effects on data quality.

[0008] Preferably, the identity recognition neural network includes: a feature extraction module, which extracts features from point cloud data, including a PointNet module, a self-attention module, a Bi-LSTM model and an MLP layer connected in series, the MLP layer performs target identity recognition based on multi-dimensional features and outputs an identity label; attention pooling, which is used for feature aggregation to form multi-dimensional features, and is respectively arranged between the PointNet module and the self-attention module and between the Bi-LSTM model and the MLP layer.

[0009] The rotating millimeter wave radar human body tracking and recognition method has the following steps: S1. Deploy a hardware module, where the hardware module consists of a millimeter wave radar module, a rotating pan-tilt module, and a control module. After the deployment is completed, the rotating millimeter wave radar human body tracking and recognition device as claimed in claim 1 or 2 is obtained; S2, the millimeter wave radar transmit signal and the reflected signal are mixed to generate an intermediate frequency signal, and the frequency offset is extracted by fast Fourier transform to calculate the target distance. At the same time, the target angle is estimated by the phase difference between the receiving antennas, and the target distance and target angle information are integrated into three-dimensional point cloud data to represent the spatial position of the target in the radar field of view; S3, preprocess the original point cloud data, then apply the DBSCAN clustering algorithm to separate the target point cloud cluster and remove outliers, and merge the target point cloud cluster into the trajectory set to form an active trajectory; S4. For the time series point cloud data of active trajectories, input it into the pre-trained deep neural network model for feature extraction and classification; S5. Based on the results of classification and identification in S4, select the set high-priority target for tracking.

[0010] Preferably, step S4 comprises: S51, normalizing the point cloud data to ensure the consistency of data at different times; S52. Use PointNet combined with Self-Attention mechanism and Bi-LSTM model to extract features from point cloud data; S53, input the multi-dimensional features into the classifier, perform target identity recognition, and output the target recognition result and its confidence.

[0011] Preferably, in step S5, a suitable threshold is selected as the edge of the field of view according to the beam characteristics of the millimeter-wave radar antenna, the current position and current speed of the target are obtained through the target's Kalman filter, and the position and angle of the target at the next moment are predicted. When the target exceeds the edge of the field of view, a motion decision is made, and the angle to which the gimbal needs to rotate is calculated according to the target motion characteristics and the gimbal rotation speed, and the gimbal angle is adjusted so that the target is in the center of the radar field of view at the end of the rotation, thereby ensuring continuous and effective target monitoring and tracking.

[0012] Preferably, step S52 includes: extracting the shape, distribution and local features of the target through the PointNet module and the self-attention module; modeling the dynamic characteristics in the time dimension through the Bi-LSTM model, capturing the motion changes of the target, and mining the behavior pattern and time series relationship.

[0013] Preferably, an edge-triggered rotation strategy is used to control the gimbal rotation. The gimbal rotation takes a certain amount of time, during which the user usually continues to move. If the radar is simply rotated to the current position of the target, it is very likely that the user has exceeded or is about to exceed the edge of the field of view when the gimbal rotation is completed, causing the radar to rotate again, thereby affecting the data quality. In order to avoid this situation and ensure that the user is in the center of the radar field of view when the rotation ends, the following precise control strategy is formulated: First, the Kalman filter is used to estimate the current position (x, y) of the target and its approximate velocity ( , ); Then, calculate the possible position and deviation angle of the target in the next frame ,in is the radar sampling interval. If the calculation result shows that the target will not exceed the set threshold, the target will continue to be monitored; otherwise, the radar needs to rotate to put the target in the center of the field of view again; To determine the angle the radar needs to rotate, we solve the following system of equations: ; in, It's time to spin. is the angle of pan / tilt rotation, is the angle at which the target currently deviates from the front of the field of view in the radar coordinate system. is the angle that the target will cross, is the angular velocity of the servo; This system of equations is nonlinear and can be solved numerically by , , according to the rotation time obtained , the number of frames required for the rotation can be determined. During this period, the association of the point cloud data is unstable due to the rotation. The trajectories in the trajectory set are only predicted by the Kalman filter without being updated, ensuring that inaccurate measurement data will not be introduced during the rotation process; Finally, according to , get the angle to which the radar should finally rotate , the system issues instructions based on this to control the gimbal to rotate to ensure that the target returns to the center of the field of view when the radar stops rotating.

[0014] Preferably, it also includes: using the point cloud cluster data obtained in step S3 to match and maintain the trajectory, for a newly detected point cloud cluster, determining whether it is a new target; if it is a new target, creating a trajectory record; for an existing trajectory, using Kalman filtering and Hungarian algorithm, according to the position, speed and historical trajectory of the target, predicting the current position and associating it with the point cloud cluster; updating the target state parameters of the matched trajectory, and removing the trajectory that has not been matched for a long time.

[0015] Preferably, in step S3, preprocessing includes removing static background noise and invalid point clouds.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention designs a millimeter-wave radar solution based on a rotating pan-tilt platform. By designing a control algorithm to dynamically adjust the detection direction of the radar, the field of view is expanded at low cost and high efficiency, solving the problem of large-scale monitoring.

[0017] 2. The present invention adopts a deep learning algorithm and introduces an attention mechanism to enhance the model's perception of key areas, extracts the motion characteristics of the human body from sparse point clouds, captures forward and backward information in time series, and realizes real-time tracking and recognition of targets.

[0018] 3. The present invention combines the point cloud processing method with the motion trajectory processing algorithm, suppresses the influence of the multipath effect, improves the quality of the point cloud data, and enhances the reliability of the data and the robustness of the system.

[0019] 4. The present invention dynamically adjusts the radar field of view by rotating the pan-tilt platform to ensure that the target is always within the monitoring range, thereby achieving continuous tracking of the target; compared with the traditional radar system with a fixed viewing angle, the present invention has higher flexibility and adaptability, and can better cope with the dynamic characteristics of targets in different scenarios.

[0020] 5. The present invention uses a low-cost commercial millimeter-wave radar module and a simple rotating pan-tilt head, which not only realizes real-time tracking and identification of multiple targets, but also has the characteristics of easy deployment and is suitable for large-scale promotion and application. Especially in the fields of smart home and health monitoring, the present invention has shown broad application prospects and significant advantages. Its low cost and high performance make it an ideal choice and can meet the needs of various application scenarios.

[0021] In summary, by combining the high-precision sensing capability of millimeter-wave radar with the dynamic viewing angle adjustment function of the rotating gimbal, the present invention effectively solves the problems of fixed field of view and limited monitoring range of traditional millimeter-wave radar; and through improved data processing algorithms and identity recognition neural networks, it overcomes the influence of environmental noise and multipath effects on data quality, and significantly improves the reliability and adaptability of the system in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the composition of the rotating millimeter wave radar human body tracking and recognition device; Figure 2 This is the working flow chart of the rotating millimeter wave radar human body tracking and recognition device; Figure 3 It is a schematic diagram of the neural network model structure; Figure 4 It is a schematic diagram of multipath effect; Figure 5 This is a schematic diagram of a typical seven-segment S-shaped speed curve of a servo; Figure 6 This is a schematic diagram of the pan / tilt rotation; Figure 7 Schematic diagram of the relationship between the number of point clouds and distance and the fitting function. DETAILED DESCRIPTION

[0023] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0024] Reference Figure 1 Understand that the rotating millimeter-wave radar human tracking and identification device includes: Millimeter-wave radar module 1: installed on the top of the rotating gimbal, used to transmit and receive millimeter-wave signals, and collect point cloud data of the target. Support frame 2: used to connect the millimeter-wave radar module 1 and the rotating mechanism 3, provide stable support, and ensure the mechanical stability of the device during operation. Rotating mechanism 3: composed of a servo motor and its drive unit, supports high-speed rotation and precise angle control, and expands the monitoring range by dynamically adjusting the direction of the antenna. Base 4: used to fix the entire rotating gimbal device on the working platform or support structure, provide stable support for the equipment, and ensure the balance of the rotating mechanism 3 and the millimeter-wave radar module 1 during operation.

[0025] Workflow: Combination Figure 2It is understood that when the target enters the radar field of view, first, the millimeter-wave radar point cloud is acquired, which means that the millimeter-wave radar first collects the reflected signals in the environment and converts these signals into point cloud data; then, the point cloud data is preprocessed, mainly point cloud clustering and screening, and the point cloud data is clustered, screened and filtered to generate independent target point cloud clusters; on this basis, the target point cloud cluster is merged into the trajectory set to form an active trajectory. Specifically, the system uses the preprocessed point cloud data to extract the target's location information. If it is a new target, it combines multiple frames of data to create a continuous trajectory of the target. Otherwise, the Hungarian algorithm is used to find the optimal match, and then the Kalman filter is used to update it, and the point cloud cluster is matched and associated with the existing trajectory, and the trajectory set is maintained according to the matching information, where the trajectory set is a set of existing trajectories, and the existing trajectory is also formed according to the point cloud data; then, according to the state of the tracking target, the system makes a decision to control the rotation of the gimbal, and the system sends a control instruction to the gimbal to instruct the gimbal to rotate, so that the target is always kept within the field of view, thereby achieving continuous tracking of the target. In addition, when multiple trajectories exist, the point cloud data of the trajectories are input into the pre-trained neural network for identity recognition, and the tracking target is selected based on the output of the neural network to ensure that high-priority targets are tracked in multi-target trajectory scenarios.

[0026] Figure 3 The structure diagram of the neural network model for identity recognition. The input radar data is the point cloud sequence data collected by the millimeter wave radar in a continuous time period. The PointNet module performs preliminary feature extraction on each point cloud data. The extracted features are then processed by the self-attention module to optimize the focus on key point cloud features and suppress redundant information. The processed features are sent to the bidirectional long short-term memory network Bi-LSTM to capture the characteristics of the target's historical trajectory and future trends. Finally, the target identity label is output through the MLP layer.

[0027] The following describes a conventional fixed millimeter wave radar human body tracking method and a rotating millimeter wave radar human body tracking method to facilitate a better understanding of the technical solution of the present invention.

[0028] 1. Traditional fixed millimeter wave radar human body tracking method As a non-contact, environmentally adaptable sensing technology, millimeter-wave radar has broad application prospects in the field of human tracking. Millimeter-wave radar human tracking mainly captures the point cloud data of the human body and uses this data for target recognition and trajectory tracking. The extraction of human point cloud mainly includes the following key steps: distance measurement, speed measurement, dynamic threshold screening (CFAR) and angle estimation, and finally generates point cloud data containing the target's three-dimensional spatial position information. These steps extract target features from different dimensions (distance, speed, direction), providing a high-quality data foundation for subsequent intelligent processing.

[0029] Millimeter wave radar is based on frequency modulated continuous wave (FMCW) technology. It measures the distance to the target by transmitting a linear frequency modulated pulse signal (also called a chirp signal) and measuring the frequency change of the echo signal. A chirp signal is a signal whose frequency changes linearly with time. Its frequency changes from a starting frequency to a starting frequency. Start in a cycle The coverage bandwidth B increases at a constant slope S. When the reflected signal is received, the radar front end calculates the frequency difference between the transmitted signal and the received signal through the mixer to generate an intermediate frequency (IF) signal. Based on the intermediate frequency signal, the distance between the object and the radar can be calculated: , where c is the speed of light, is the frequency of the intermediate frequency signal. By performing a fast Fourier transform (FFT) on the intermediate frequency signal, the time domain signal can be converted into a frequency domain signal, and each FFT peak corresponds to the distance of a target. This process is called Range-FFT.

[0030] For multiple targets at the same distance, it is impossible to distinguish them directly through Range-FFT, so they need to be distinguished by the difference in radial velocity when they move toward the radar. A small change in the distance of the target will cause a significant shift in the phase of the intermediate frequency signal, so we transmit multiple chirp and measure the phase difference To obtain the relative speed of the object being measured : ,in It is the wavelength of the radar signal. For multiple objects at the same distance but at different speeds, a set of multiple equally spaced linear frequency modulation pulses needs to be emitted to form a "frame". In each frame, the range-Doppler spectrum of the target is first obtained through Range-FFT. In the spectrum, the phase change information can be obtained by extracting the peak phase of each chirp signal and performing a fast Fourier transform (FFT). The peaks of these phase changes correspond to the radial velocity information of the target object. This process is called Doppler-FFT. Through Doppler-FFT, targets with different radial velocities can be distinguished even at the same distance.

[0031] After completing the Doppler-FFT processing, the radar generates a set of dense data containing target distance and speed information. However, these data usually contain redundant information and are interspersed with a large number of noise points, which increases the complexity of subsequent processing. In order to improve processing efficiency and accurately extract target points, the Doppler-FFT data needs to be sparsely processed. In the sparse process, the Constant False Alarm Rate (CFAR) algorithm plays a key role as a dynamic threshold screening method. CFAR sets a dynamic threshold based on the local environment for each data point to effectively distinguish target points from noise points. Specifically, CFAR first calculates the average energy level in the neighborhood of the target point, and then multiplies the average by a preset proportional factor as the dynamic threshold. If the energy value of the current point exceeds the threshold, it is considered to be a target point; otherwise, it is considered to be a noise point. In addition, targets in actual scenes usually occupy multiple adjacent data units. In order to avoid misjudgment of target points, CFAR introduces the concept of protection area. A certain range is set around the selected target point so that the points within the range do not participate in the statistics of background noise, thereby further reducing the misjudgment rate.

[0032] For multiple targets at the same distance and speed, the angle of arrival (AoA) information is needed to distinguish them. Millimeter-wave radar uses multiple receiving antenna arrays to generate phase differences through differences in signal propagation paths to estimate the target's arrival angle. Taking a simple dual-antenna configuration as an example, the signal transmitter sends a linear frequency modulation signal with the same initial phase, and the phase difference between the signals received by the two receiving antennas is Can be used to estimate the angle of arrival of a target : ; in is the signal wavelength, l is the antenna spacing, is the phase difference. For more complex multi-antenna arrays that detect multiple targets, multiple phase differences can be calculated by performing FFT on the signals received by all antennas at the same distance and speed , and further derive the arrival angles of multiple targets, a process called "angle-FFT". In addition, for linear antenna arrays, the angle resolution The theoretical limit can be estimated by the following formula: ; Where L is the length or aperture size of the antenna array, is the angle of the target relative to the antenna array. The angular resolution is when the target is directly in front of the antenna array ( =0) and as decreases with the change of .

[0033] After obtaining the target distance r, radial velocity v and horizontal angle Angle to vertical After that, the three-dimensional coordinates (x, y, z) of the target can be calculated by the conversion formula from polar coordinates to rectangular coordinates. ; ; ; In this way, we get the radar point cloud data we need, which provides a basis for subsequent target recognition and tracking.

[0034] 2. Rotating millimeter wave radar human body tracking method Due to the directivity of the radar antenna, the further the target deviates from the radar's front direction, the lower the received signal strength. This will cause the signal strength gap between the real data and the noise data to become smaller, thus affecting the accuracy of the CFAR algorithm and may even cause the target to be unable to be effectively detected. In addition, the angular resolution will also decrease with the increase of the target distance and deviation angle, further affecting the performance of the system.

[0035] In order to overcome this problem, the present invention adopts a rotating pan-tilt technology to enhance the field of view coverage capability of the radar and expand its detection range. The rotating pan-tilt not only effectively overcomes the limitations brought by the directivity of the radar antenna, but also can dynamically adjust the scanning direction of the radar to ensure that the target is always within the monitoring range. By adjusting the direction of the radar, the rotating pan-tilt can also optimize the accuracy of the angle of arrival measurement, especially in the case of large angles, and can significantly improve the accuracy of the angle estimation. This design significantly improves the flexibility and adaptability of the system, and is suitable for a variety of complex application scenarios, such as smart homes, security monitoring, and health monitoring, providing a broader application space for radar systems.

[0036] In order to achieve efficient target tracking, the rotational tracking process is divided into the following key steps. First, the radar collects point cloud information and corrects the coordinates according to the rotation state; then, the collected point cloud information is processed to eliminate anomalies and extract the target point cloud cluster; then, the point cloud data is associated with the trajectory information to build a stable and continuous trajectory; finally, the system selects the target trajectory and ensures that the target is within the radar field of view by controlling the gimbal rotation to achieve accurate tracking.

[0037] 1. Coordinate correction The radar collects the signal generated by the target movement through millimeter wave reflection, and then generates sparse point cloud information based on the signal. The characteristics of these point cloud data include the location of the target ( coordinates), speed towards the radar, and signal strength. However, since the radar is constantly rotating and the rotation axis does not overlap with the radar receiving antenna, and the point cloud data collected by the radar is based on the coordinate information generated by the current direction of the radar. Therefore, in order to convert these point cloud data into the real coordinate system, we need to correct the coordinates of the point cloud.

[0038] Specifically, the front of the radar's initial orientation is set as the x-axis of the real coordinate system and follows the left-hand coordinate system rule. Since the gimbal rotates horizontally, there is no need to consider the z-axis information during the correction process. It is assumed that the radar rotates clockwise by an angle of , then the point cloud coordinates should also rotate clockwise , the rotation matrix as follows: ; Through this rotation matrix, the point cloud data coordinates in the radar coordinate system ( ) is converted to coordinates in the real coordinate system ( ), the rotation formula is as follows: ; , is the distance deviation between the rotation axis and the radar receiving antenna.

[0039] 2. Point cloud processing When using millimeter-wave radar for environmental perception, the radar may receive non-target signals or erroneous data due to various factors. Non-target signals refer to interfering reflections from the environment, such as millimeter-wave signals reflected by non-target objects such as walls, floors, ceilings, and furniture. These signals may be superimposed on the real echo of the target, causing confusion. Erroneous data includes false target points caused by multipath effects, isolated points caused by environmental noise, and random errors caused by insufficient measurement accuracy of the equipment. Erroneous data usually manifests itself as false targets generated by multipath effects, redundant information in point cloud data, or measurement deviations caused by external interference or internal noise in the equipment. These problems will significantly affect the quality of point cloud data and reduce the system's recognition accuracy and tracking reliability of targets.

[0040] The multipath effect is the most serious and common problem. It will cause multiple false targets to appear, interfering with the radar's detection and identification of real targets. The multipath effect refers to the reflection, refraction or scattering of radar signals when they encounter different surfaces (such as walls, ground, etc.) during the propagation process, causing the signal to propagate through multiple paths. These multipath signals are superimposed together to generate false targets, affecting the accuracy of target tracking and identification. Figure 4 As shown, taking the multipath effect caused by the reflection surface on one side of the ground as an example, for the target at position P0, the correct signal transmission path should be , However, due to the presence of the ground, the signal may , Launched to P0, it may also be along Back to the radar, the combination of multiple paths will produce a false target point P1 ( , ), P2 ( , ), P3 ( , , ). Therefore, in order to improve the accuracy of subsequent target tracking and identity recognition, the point cloud data must be preprocessed to suppress the influence of multipath effects.

[0041] First, if Figure 4 As shown in the figure, the false point cloud generated by the multipath effect is usually farther than the real position, and some background noise may also be outside the monitoring range. Based on the prior knowledge of the environment or the detection range preset by the user, we perform range filtering on the point cloud data of each frame and remove points beyond the detection range to eliminate anomalies and simplify data processing.

[0042] Secondly, the noise points generated by environmental noise and background interference are usually isolated, and the distribution of human point clouds is irregular. There may be multiple targets in the environment, which brings challenges to the processing of point cloud data. To this end, we clustered the point cloud data and adopted the density-based clustering algorithm DBSCAN algorithm, which can divide areas with sufficient density into clusters and can handle clusters of any shape and size. We set the clustering parameters as: neighborhood radius =0.5, the minimum number of samples MinPts=5. Experiments show that this value can cover the neighborhood range of a typical human point cloud, and effectively avoid the mistaken aggregation of multiple targets into one cluster due to a large neighborhood radius. Through clustering, we divide the point cloud data into multiple target point cloud clusters and remove isolated points that may be noise or false targets.

[0043] Furthermore, due to the angular resolution limitation of the millimeter-wave radar, the point cloud of the same target may be incorrectly segmented into multiple independent point cloud clusters. Therefore, we merge the point cloud clusters that may be the same target in each frame of data after clustering to improve the accuracy and completeness of the data. The threshold is set to dis× If the distance between the centers of two point cloud clusters is less than the threshold and the distance between the centers of the two point cloud clusters and the radar is small, they will be merged. Here, dis is the average distance between the two point cloud clusters and the radar. Slightly larger than the horizontal angle resolution of the radar, it is used to generate an adaptive threshold to process point cloud data at different distances from the radar.

[0044] Then, the multipath effect will not only produce isolated noise points, but also form local clusters. Although the DBSCAN algorithm can effectively filter out isolated points, for noise points generated by multipath effects in dense point cloud areas, these noise points may not be clustered and filtered, but are regarded as point cloud clusters of a target. In order to further eliminate such misjudgments, we combine the spatial characteristics of the multipath effect and use angle analysis to eliminate point cloud clusters that do not conform to the real target. Calculate the center position of each point cloud cluster and its respective angle boundary (the minimum and maximum deviation angle interval of all points in the point cloud cluster relative to the radar), compare all point cloud clusters pairwise, and check the relative position and angle relationship between the two point cloud clusters. If the angle between the line between the centers of the two point cloud clusters and the line between the radar and the closer point cloud cluster is less than the set threshold, or the angle boundary of the point cloud cluster closer to the radar completely contains the angle boundary of the point cloud cluster farther from the radar, it can be inferred that the point cloud cluster farther from the radar is a false target caused by the multipath effect and is deleted from the point cloud data.

[0045] After that, the point clouds that do not conform to human noise are screened. The height threshold is set to 0.8 meters (half of the average height of a human body), and the height of each point cloud cluster is calculated. If it is less than the height threshold, it will be deleted. While eliminating point cloud clusters that do not conform to human characteristics, it can also avoid point clouds that are mistakenly deleted due to partial body occlusion or static state; then, the number of points in each point cloud cluster is counted. If the number of points in the point cloud cluster is less than 10, it is considered that the information provided by the point cloud is insufficient to represent an actual human target and may be a false noise target, so it is deleted.

[0046] 3. Trajectory management Through the previous filtering operation, the point cloud of each frame has been maintained and processed, but the point clouds between frames have not yet formed a connection, so it is necessary to associate the discrete point clouds to generate continuous tracks. We create and maintain a track set to manage all the current target tracks. The track set is initially empty and is dynamically updated as new tracks are added and inactive tracks are deleted.

[0047] For each point cloud cluster, we calculate its geometric center as the location of the cluster. However, due to the sparsity of point cloud data, the z-axis center of gravity of the point cloud will fluctuate in the vertical direction. This fluctuation may cause instability of the point cloud in the z-axis direction. Therefore, trajectory tracking pays more attention to the horizontal movement of the target in the field of view rather than the vertical change. Therefore, when matching point clouds and trajectories, we usually ignore the z attribute of the point cloud center of gravity and only consider the changes in the x and y directions.

[0048] For point cloud matching, we need to match the trajectory with the point cloud cluster of the next frame. A simple approach is to directly match the current position of the trajectory with the nearest point cloud cluster in the next frame. However, this method has certain problems. Since the target may move, matching based solely on the current position may lead to mismatching, especially when the target has a large displacement, simple nearest distance matching often cannot provide sufficient accuracy. So here we introduce the Kalman filter. The Kalman filter can estimate uncertain information, and can also better predict the next state in the presence of noise interference, and find the subtle correlation between multiple variables. In addition, the Kalman filter has low memory usage and fast inference speed, which is suitable for resource-constrained scenarios. Since the point cloud generated by the radar is unevenly distributed, and the speed of different parts of the body is different when the target moves, the point cloud density will also change. Simply using the center of the point cloud cluster as the target position will result in large errors. Introducing the Kalman filter can reduce errors and obtain a more stable and smooth trajectory. At the same time, when the gimbal rotates, static objects in the environment may be mistakenly identified as moving targets. In addition, the gimbal rotation control is asynchronous, and the accurate rotation angle cannot be obtained in real time, which will lead to large errors when performing point cloud matching. Therefore, when the gimbal rotates, we choose to abandon the point cloud cluster information obtained at this time, and instead use the Kalman filter to predict the target position as the current recording point of the trajectory, and use it to participate in the point cloud matching of the next frame.

[0049] When performing point cloud matching, if only the point cloud closest to the trajectory is selected, it may lead to repeated selection or too large matching distance. Therefore, we need a more accurate matching method to ensure that each point cloud and each trajectory can only be selected once. To solve this problem, we use the Hungarian algorithm. The Hungarian algorithm is an optimization algorithm for solving the assignment problem, which is used to find the best match in a group of objects. The core idea is to minimize the total cost of matching by constructing a cost matrix and optimizing the matrix.

[0050] Specifically, for each trajectory in the trajectory set, we create a Kalman filter and compare the trajectory’s current position (x, y) with its current velocity (v x , v y) is set as its state vector. For each frame of point cloud, Kalman filter prediction is performed on each trajectory in the trajectory set, and the possible position of the target in the current frame is predicted based on the state of the trajectory in the previous frame. For each point cloud cluster generated in the current frame, its geometric center position is calculated, and the distance between the predicted trajectory position and the center position of the point cloud cluster is calculated pairwise, and a cost matrix is ​​constructed based on these distances. If the distance in the matrix is ​​greater than 0.8 meters (the maximum distance that an adult can travel between two frames), the cost will be set to unreachable, indicating that the point cloud cluster cannot match the current trajectory. The Hungarian algorithm is used to optimize the cost matrix to complete the global optimal matching of the trajectory and the point cloud cluster, so that the total cost between the matched trajectory and the point cloud cluster is minimized.

[0051] 4. Track maintenance When the match is successful, we can assume that the target represented by the track is still in the field of view. However, when the target suddenly appears or disappears, the Hungarian algorithm cannot complete the match for the new point cloud cluster or the track that is no longer active. Therefore, in order to ensure the accuracy and real-time performance of the system, track maintenance is required, including completing track updates for targets that are still active within the detection range, and considering the tracks of targets that leave the field of view or stop being active as invalid tracks and removing them from the track set.

[0052] When a new target suddenly appears in the detection range, the trajectory set is empty or the point cloud cluster is not successfully matched to the trajectory. A new trajectory is created for the point cloud cluster and the geometric center position of the point cloud cluster is used as the starting point of the trajectory. The state vector of the trajectory (including the current position (x, y) and the current speed (v) is initialized. x , v y )), and add the new trajectory to the trajectory set.

[0053] When the target trajectory is successfully matched with the point cloud cluster in a new frame, the data of the new point cloud cluster is associated with the matched trajectory and the position information of the point cloud cluster is used to update the state vector and covariance matrix of the Kalman filter, further improving the accuracy of trajectory prediction.

[0054] When the target of the track is lost, the track fails to match any point cloud cluster successfully, and the time interval of the unsuccessful track matching is counted. If the track fails to match any point cloud cluster for more than 1 second, it is considered that the tracked target has left the tracking range or is no longer moving, and it is removed from the track set.

[0055] For temporary trajectories caused by noise, all trajectories in the trajectory set are checked. If a trajectory lasts less than 1 second and fails to successfully match a point cloud cluster, it is considered invalid and removed from the trajectory set to avoid the accumulation of false trajectories caused by short-term interference and improve the overall reliability of the system.

[0056] 5. Rotation control After completing the trajectory matching, we use the rotating gimbal to continuously track the target trajectory. In order to ensure the accuracy of tracking and the stability of the system, the control strategy of the present invention is inspired by the camera gimbal control method, that is, dynamically adjusting the viewing angle according to the target position. However, considering the sparsity of millimeter-wave radar point cloud data and the motion characteristics of the rotating gimbal, it is necessary to choose between the following two strategies: 1. Continuous adjustment strategy: The gimbal continuously adjusts when the target moves, keeping the target always in the center of the field of view; 2. Edge-triggered rotation strategy: The rotation is initiated only when the target approaches the edge of the field of view, and the target is placed back in the center of the field of view.

[0057] Next, we analyze the above two strategies. For strategy one, since the data collected by the millimeter-wave radar is a sparse point cloud and the motion characteristics of different parts of the human body vary greatly, the point cloud distribution is often uneven, which may cause the target position information to change in a jumpy manner. Although continuous rotation helps to maximize the target signal strength, frequent rotation may cause severe shaking of the gimbal, thereby reducing the quality of the point cloud data. In addition, the radar's ability to remove static objects will also be disturbed, and the system accuracy is difficult to guarantee. At the same time, since the servo movement is usually non-uniform and the control is not completely synchronized, the point cloud data during the rotation process may produce significant errors, resulting in a significant reduction in the reliability of the data frame.

[0058] In contrast, Strategy 2 avoids the jitter problem caused by frequent rotation by reducing unnecessary rotation movements, thereby effectively improving the tracking effect and the reliability of point cloud data. Although edge-triggered rotation requires discarding some data frames generated during the rotation process, these missing data can be predicted and supplemented by the Kalman filter to ensure the continuity of the trajectory. Therefore, the present invention selects an edge-triggered rotation strategy, sets a fixed angle as the threshold of the edge of the field of view, and when the target approaches the edge of the field of view, the pan-tilt rotation is started to bring the target back to the center of the field of view. This method reduces the invalid actions of the system while ensuring the tracking accuracy, effectively improving the quality of the point cloud data and the overall stability of the system.

[0059] In order to effectively implement the edge-triggered rotation control strategy, it is crucial to choose a suitable angle as the field of view threshold. This is because the field of view threshold directly affects the frequency of servo rotation and the overall efficiency of the system. If the field of view threshold is too small, once the target approaches the edge, the gimbal will frequently start rotating, resulting in reduced data frame quality and reduced tracking stability; on the contrary, if the threshold is too large, the signal strength of the target at the edge of the field of view may be significantly reduced, the data quality is unreliable, and the tracking effect is affected.

[0060] In order to further study the impact of the field of view threshold, we analyzed the motion characteristics of the servo. Rotary servos generally follow an S-curve acceleration and deceleration mode, which can smooth the motion trajectory of the servo and avoid system oscillation or instability caused by too fast changes in speed and acceleration. Figure 5 As shown in the figure, the S-curve acceleration and deceleration can be divided into: acceleration section (T1), uniform acceleration section (T2), deceleration section (T3), uniform speed section (T4), acceleration and deceleration section (T5), uniform deceleration section (T6), and deceleration and deceleration section (T7). To simplify the discussion, the acceleration during the acceleration and deceleration process of the servo is , maximum acceleration a max With maximum speed v max Set as a constant, and the radar reaches the maximum speed v each time it rotates max Then decelerate, the motion relationship can be obtained by the integral formula , set the angle of a single rotation ,but The speed curve is The area enclosed by the axis can be obtained from the symmetry of the acceleration and deceleration process. , then for the total angle n turned, the total time consumed can be expressed as ; From the formula, we can see that the total time consumed is Angle of single rotation The relationship is inversely proportional. The smaller the angle of a single rotation, the more times the servo needs to rotate to complete the same total angle n; the number of starts and accelerations and decelerations increases, resulting in an increase in the total time. Similarly, if the angle is too small and the servo cannot accelerate to the maximum speed in a single rotation, it will take more time. The system is prone to generate more data chaos frames caused by rotation, reducing tracking accuracy. Therefore, the angle should be as large as possible.

[0061] However, when the angle is too large, the signal strength of the target decreases when it moves to the edge of the field of view, which makes it impossible to obtain enough detailed information, which conflicts with the original intention of the gimbal design. In order to determine a suitable angle as the edge of the field of view, we designed an experiment, requiring the tester to move freely within the range of 6 to 8 meters of the radar and count the average number of point cloud data per frame. The angles in the experiment were set as: 0°, the horizontal 3dB beam width of the radar was 28°, and the 6dB beam width was 50°, corresponding to the maximum radiation direction, the angle width when the power density dropped to half and one-quarter relative to the maximum radiation direction, respectively. The experimental results are shown in Table 1. The results show that at the position where the power density drops by half, the number of points per frame decreases slightly relative to the maximum power position, while when the power density drops to one-quarter, the number of point clouds decreases significantly.

[0062] Based on the above analysis, the present invention selects 28 degrees as the field of view threshold to ensure the best tracking effect and data quality. This angle matches the horizontal 3dB beam width of the radar, which not only ensures sufficient field of view, but also avoids the problem of target signal attenuation caused by excessive angle.

[0063] Table 1: Relationship between the number of points per frame and angle of the free-motion radar at 6~8m

[0064] The pan / tilt rotation takes a certain amount of time, during which the user usually continues to move. If the radar is simply rotated to the current position of the target, it is very likely that the user has exceeded or is about to exceed the edge of the field of view when the pan / tilt rotation is completed, causing the radar to rotate again, thus affecting the data quality. In order to avoid this situation and ensure that the user is in the center of the radar field of view when the rotation ends, we formulate the following precise control strategy. First, the Kalman filter is used to estimate the current position (x, y) of the target and its approximate velocity ( , ). Then calculate the possible position and deviation angle of the target in the next frame ,in is the radar sampling interval. If the calculation result shows that the target will not exceed the set threshold, the target will continue to be monitored; otherwise, the radar needs to rotate to put the target in the center of the field of view again. Figure 6 As shown, , is the real coordinate system, , is the radar-based coordinate system used to evaluate the position of the target relative to the radar field of view. To determine the angle the radar needs to rotate, we solve the following system of equations: ; in, It's time to spin. is the angle of pan / tilt rotation, is the angle at which the target currently deviates from the front of the field of view in the radar coordinate system. is the angle that the target will cross, is the angular velocity of the servo.

[0065] This system of equations is nonlinear and can be solved numerically by , . According to the rotation time obtained , the number of frames required for the rotation can be determined. During this period, the association of the point cloud data is unstable due to the rotation. The trajectories in the trajectory set are only predicted by the Kalman filter without being updated, ensuring that no inaccurate measurement data is introduced during the rotation process. , is a known quantity, according to , we can get the angle that the radar should finally rotate to. Finally, the system will issue a command to control the gimbal to rotate to ensure that the target returns to the center of the field of view when the radar stops rotating.

[0066] Through this precise control method, the present invention can reduce unnecessary rotational movements and the burden of the steering gear while ensuring tracking accuracy, thereby improving the response speed and stability of the entire system.

[0067] 3. Deep Neural Networks In the radar field of view, when there is only one track, the gimbal can automatically track the movement of the target. However, when there are multiple tracks, it is necessary to select a track with a higher priority for tracking to enhance the signal strength in the target direction and obtain more effective information. To solve this problem, the present invention designs a target recognition algorithm based on a deep neural network. The system can select high-priority targets and continue to track them according to preset priority labels and confidence levels, thereby achieving efficient target management and tracking optimization.

[0068] 3.1 Data Augmentation Considering that the target may be located anywhere in the room and move along any path, in order to ensure that the system can complete the recognition task, the neural network model needs to have a high generalization ability. However, the actual collected training data is difficult to cover all possible scenes and states. Therefore, the present invention adopts data enhancement technology to increase the diversity of data, thereby improving the robustness and adaptability of the deep neural network model.

[0069] Specifically, we randomly rotate each frame of point cloud data and generate diversified point cloud data by setting the global rotation angle range and local perturbation angle range. Global rotation simulates the overall displacement of the target under different viewing angles, while local perturbation introduces subtle angle changes to capture changes in local features of the target. This method not only increases the diversity of training data, but also effectively simulates the point cloud distribution characteristics of the target in different postures and motion states, improving the adaptability of the model.

[0070] In order to simulate the change of point cloud density at different distances of the target and enhance the model's recognition ability of sparse data, the present invention uses resampling technology to enhance the point cloud. Figure 7 As shown in the figure, the number of point clouds f(x) shows a nonlinear decreasing trend as the distance x between the target and the radar changes. This relationship reflects that in millimeter-wave radar, the farther the target is, the weaker the echo signal strength is, and the number of point clouds decreases accordingly. The approximate relationship between the number of point clouds and the distance is obtained through experimental fitting: ; Where a and b are fitting parameters, and their specific values ​​depend on the signal characteristics of the radar and the reflection intensity of the target. Based on the above relationship, the present invention designs a point cloud resampling method to dynamically adjust the number of point clouds per frame n in the time slice according to the change of the target distance. Specifically: the fitting function f(x) is obtained using the target training data to simulate the change of the target distance. , then the number of point clouds is adjusted to: ; Where n′ is the number of adjusted point clouds. When the distance becomes closer, the number of point clouds needs to be increased. At this time, the points in the original point cloud cluster will be copied and slightly offset in the direction of the nearest point to simulate a higher density point cloud distribution. When the distance becomes farther, the number of point clouds needs to be reduced. At this time, points are deleted from the original point cloud cluster by random sampling to adjust the point cloud density. This method of dynamically adjusting the point cloud density based on distance can not only improve the robustness of the model when processing sparse point clouds, but also better adapt to the point cloud density fluctuations caused by changes in target distance in actual application scenarios, thereby improving the generalization ability and application effect of the model.

[0071] In order to further eliminate scale differences and improve the model's sensitivity to data with uneven feature distribution, the present invention standardizes the spatial coordinates and attribute features of point cloud data, including decentralization and normalization. The x, y, and z spatial coordinates of the point cloud are decentralized, and the distribution of the point cloud data is concentrated near the origin by subtracting the average value of each coordinate dimension. This processing method can effectively eliminate the deviation between different coordinate ranges and reduce the interference caused by spatial position differences, thereby improving the stability of the model in capturing point cloud features. The signal strength and speed attribute features are normalized and their values ​​are mapped to a unified numerical range of [-1, 1]. The purpose of normalization is to avoid certain features from occupying a larger weight during the model training process due to feature scale differences, and to ensure that each feature is equally important to model learning.

[0072] 3.2 Network Architecture The input data is the enhanced point cloud data with a shape of b×k×n×5, where b is the batch size, k is the window size, and n is the number of points per frame. The point cloud has five feature dimensions, including x, y, z spatial coordinates, intensity, and velocity. If the number of points in each point cloud frame is greater than n, it is randomly sampled to n points. If the number of points in the point cloud is less than n, the number of points is expanded to n by zero padding.

[0073] In point cloud data processing, it is crucial to choose a suitable feature extraction method. Since point cloud data exists in the form of sparse and irregular three-dimensional points, its distribution characteristics are difficult to be directly processed by traditional convolutional neural networks. To this end, the present invention adopts the PointNet architecture as a feature extraction method. PointNet is a network structure designed specifically for processing point cloud data. It can extract local features of each point in the point cloud and aggregate global features to help the network understand the spatial distribution of the point cloud. Compared with other complex point cloud processing methods, it has higher computational efficiency and can meet real-time processing requirements. However, although the maximum pooling operation in PointNet can aggregate global features, it cannot fully highlight the information of key points, thereby affecting the quality of feature expression. To this end, the present invention replaces maximum pooling with attention pooling, generates weights for each point through a multi-layer perceptron (MLP), and gives higher attention to particularly important points in the point cloud. Subsequently, by weighted aggregation of point cloud data, key information is retained while suppressing interference from irrelevant or redundant points. This method can enhance the model's perception of key areas, thereby further improving the quality of feature extraction.

[0074] In order to further optimize the feature extraction effect, the present invention introduces the self-attention mechanism. This mechanism captures the interaction of key features in the point cloud by calculating the global relationship between feature points, while effectively suppressing redundant information, thereby improving the expressiveness of features. Combining attention pooling and self-attention mechanism, the network can more comprehensively understand the spatial characteristics of the point cloud and output higher quality feature tensors. ,in is the dimension of the extracted features.

[0075] Point cloud data not only has spatial distribution characteristics, but also contains dynamic changes in time. To this end, the present invention uses a bidirectional long short-term memory network (Bi-LSTM) to capture the dynamic changes of the target in the time series. Bi-LSTM can simultaneously capture the forward and backward information in the time series, enhancing the ability to model the target's motion trajectory. The process outputs the trajectory feature tensor ,in is the size of the LSTM hidden unit. In order to further optimize the time series features, the present invention continues to use the attention pooling technology after LSTM, so that the model pays more attention to the features of the key time steps in the time series and reduces the negative impact of long-term forgetting on the performance. Finally, the optimized time series features are compressed into The character vector of the shape.

[0076] Finally, the extracted trajectory features are input into the fully connected layer for classification, and the identity of the target and its confidence are output. This design enables the system to identify targets in real time and efficiently in complex dynamic scenes, providing a reliable basis for subsequent gimbal tracking decisions.

[0077] 3.3 Model Training The dataset is divided into training set and test set according to the ratio of 4:1 to ensure that the data in the training phase and the evaluation phase do not overlap. The cross entropy loss function is used for the classification task: ; Where b is the batch size, C is the number of categories, is the true label of the i-th sample in category c, is the predicted value of the i-th sample in category c.

[0078] During the training process, the batch size is set to 32. After each training cycle, the accuracy is evaluated on the test set, and the hyperparameters are adjusted based on the evaluation results to optimize the model performance. The Adam optimizer is used, the initial learning rate is set to 0.0001, and the learning rate scheduler is used to adjust the learning rate according to the training process.

[0079] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present application and are not intended to be limiting of the present application. As long as they are within the spirit and scope of the present application, appropriate changes and modifications to the above embodiments are within the scope of protection claimed in the present application.

Claims

1. A rotating millimeter wave radar human body tracking and identification device, characterized in that: include: The millimeter wave radar module uses a 77GHz-81GHz millimeter wave radar based on frequency modulated continuous wave technology to transmit and receive millimeter wave signals and collect point cloud data of the target; The rotating gimbal module uses a servo motor to support high-precision, high-speed rotation and precise angle control. The single rotation angle of the gimbal is 0~180°; The control module uses an embedded processor with high computing power and low power consumption, suitable for performing complex real-time signal processing and target recognition tasks; The millimeter-wave radar module is installed on the rotating gimbal module, and the main beam direction of the antenna in the millimeter-wave radar module is perpendicular to the rotation axis; the millimeter-wave radar module is connected to the control module, and the embedded processor has a built-in point cloud data processing algorithm and identity recognition neural network to overcome the influence of environmental noise and multipath effects on data quality.

2. The rotary millimeter wave radar human body tracking and identification device according to claim 1, characterized in that: The identity recognition neural network includes: The feature extraction module extracts features from point cloud data, including a serially connected PointNet module, a self-attention module, a Bi-LSTM model, and an MLP layer. The MLP layer recognizes the target identity based on multi-dimensional features and outputs the identity label. Attention pooling is used to aggregate features to form multi-dimensional features. It is set between the PointNet module and the self-attention module and between the Bi-LSTM model and the MLP layer.

3. A rotating millimeter wave radar human body tracking and identification method, characterized in that: Here are the steps: S1. Deploy a hardware module, where the hardware module consists of a millimeter wave radar module, a rotating pan-tilt module, and a control module. After the deployment is completed, the rotating millimeter wave radar human body tracking and recognition device as claimed in claim 1 or 2 is obtained; S2, the millimeter wave radar transmit signal and the reflected signal are mixed to generate an intermediate frequency signal, and the frequency offset is extracted by fast Fourier transform to calculate the target distance. At the same time, the target angle is estimated by the phase difference between the receiving antennas, and the target distance and target angle information are integrated into three-dimensional point cloud data to represent the spatial position of the target in the radar field of view; S3, preprocess the original point cloud data, then apply the DBSCAN clustering algorithm to separate the target point cloud cluster and remove outliers, and merge the target point cloud cluster into the trajectory set to form an active trajectory; S4. For the time series point cloud data of active trajectories, input it into the pre-trained deep neural network model for feature extraction and classification; S5. Based on the results of classification and identification in S4, select the set high-priority target for tracking.

4. The rotary millimeter wave radar human body tracking and identification method according to claim 3 is characterized in that: Step S4 includes: S51, normalizing the point cloud data to ensure the consistency of data at different times; S52. Use PointNet combined with Self-Attention mechanism and Bi-LSTM model to extract features from point cloud data; S53, input the multi-dimensional features into the classifier, perform target identity recognition, and output the target recognition result and its confidence.

5. The rotary millimeter wave radar human body tracking and identification method according to claim 4, characterized in that: In step S5, a suitable threshold is selected as the edge of the field of view according to the beam characteristics of the millimeter-wave radar antenna, the current position and current speed of the target are obtained through the target's Kalman filter, and the position and angle of the target at the next moment are predicted. When the target exceeds the edge of the field of view, a motion decision is made, and the angle to which the gimbal needs to rotate is calculated according to the target motion characteristics and the gimbal rotation speed. The gimbal angle is adjusted so that the target is in the center of the radar field of view at the end of the rotation, thereby ensuring continuous and effective target monitoring and tracking.

6. The rotary millimeter wave radar human body tracking and identification method according to claim 4, characterized in that: Step S52 includes: extracting the shape, distribution and local features of the target through the PointNet module and the self-attention module; modeling the dynamic characteristics in the time dimension through the Bi-LSTM model, capturing the movement changes of the target, and mining the behavior pattern and time sequence relationship.

7. The rotary millimeter wave radar human body tracking and identification method according to claim 6, characterized in that: The edge-triggered rotation strategy is used to control the gimbal rotation. The gimbal rotation takes a certain amount of time, during which the user usually continues to move. If the radar is simply rotated to the current position of the target, it is very likely that the user has exceeded or is about to exceed the edge of the field of view when the gimbal rotation is completed, causing the radar to rotate again, thus affecting the data quality. In order to avoid this situation and ensure that the user is in the center of the radar field of view when the rotation ends, the following precise control strategy is formulated: First, the Kalman filter is used to estimate the current position (x, y) of the target and its approximate velocity ( , ); Then, calculate the possible position and deviation angle of the target in the next frame ,in is the radar sampling interval. If the calculation result shows that the target will not exceed the set threshold, the target will continue to be monitored; otherwise, the radar needs to rotate to put the target in the center of the field of view again; To determine the angle the radar needs to rotate, we solve the following system of equations: ； in, It's time to spin. is the angle of pan / tilt rotation, is the angle at which the target currently deviates from the front of the field of view in the radar coordinate system. is the angle that the target will cross, is the angular velocity of the servo; This system of equations is nonlinear and can be solved numerically by , , according to the rotation time obtained , the number of frames required for the rotation can be determined. During this period, the association of the point cloud data is unstable due to the rotation. The trajectories in the trajectory set are only predicted by the Kalman filter without being updated, ensuring that inaccurate measurement data will not be introduced during the rotation process; Finally, according to , get the angle to which the radar should finally rotate , the system issues instructions based on this to control the gimbal to rotate to ensure that the target returns to the center of the field of view when the radar stops rotating.

8. The rotary millimeter wave radar human body tracking and identification method according to claim 3, characterized in that: Also includes: Use the point cloud cluster data obtained in step S3 to match and maintain the trajectory, and determine whether the newly detected point cloud cluster is a new target; If it is a new target, a trajectory record is created; for existing trajectories, the Kalman filter and Hungarian algorithm are used to predict the current position and associate it with the point cloud cluster based on the target's position, speed and historical trajectory; Update the target state parameters of the matched trajectories and remove the trajectories that have not been matched for a long time.

9. The rotary millimeter wave radar human body tracking and identification method according to claim 3, characterized in that: Step S3, preprocessing includes removing static background noise and invalid point clouds.

Images