Top-mounted human trajectory tracking system and method based on millimeter wave 4D point cloud

Abstract

This invention discloses a top-mounted human trajectory tracking system and method based on millimeter-wave 4D point clouds. The system includes a millimeter-wave radar module, an MCU processor, and a communication module. The method includes: converting the radar point cloud from spherical coordinates to Cartesian coordinates; using one-dimensional DBSCAN high-level clustering to filter out mirror noise in top-mounted scenes; performing DBSCAN clustering on the XY plane to obtain target clusters; achieving cross-frame target association through Hungarian matching, and extending the lifespan of unmatched historical targets by utilizing the noise point density in their neighborhood (noise extension); for targets with height abrupt changes, adaptively adjusting the Kalman filter observation noise covariance to suppress position jumps; managing the target lifespan based on the number of consecutive matching frames, and supporting regional rules such as prohibited areas and rest areas. This invention effectively solves the problems of mirror interference, easy loss of cross-layer tracking, and poor scene adaptability of top-mounted millimeter-wave radar in human trajectory tracking in privacy-sensitive locations.

Description

Technical Field

[0001] This invention relates to the field of radar target tracking technology, specifically to an indoor human trajectory tracking system and method based on 4D millimeter-wave radar point clouds, suitable for ceiling-mounted scenarios. Background Technology

[0002] In places with high privacy requirements, such as nursing homes, hotels, and conference rooms, monitoring the presence, number, and activity patterns of people indoors is of significant application value. Traditional visual sensors (such as cameras) are severely limited in their deployment in these scenarios due to the potential for infringing on personal privacy.

[0003] 4D millimeter-wave radar can provide information on the position, velocity, and intensity of a target in three-dimensional space (i.e., 4D point cloud) without capturing biological visual details, making it an ideal choice for solving personnel monitoring problems in privacy-sensitive scenarios. However, existing radar tracking algorithms face the following challenges when applied to top-mounted systems: (1) Point cloud noise and mirror interference: The electromagnetic waves of the top-mounted radar are reflected by the ground and walls, which easily generate a large number of mirror points and background noise, interfering with the detection of real targets.

[0004] (2) Problem of sudden changes in height across layers: When a target (human body) moves at different height layers indoors (such as from the ground to the sofa or bed), the height of its point cloud ( The coordinates may change significantly, which can easily lead to incorrect switching of the target identity (ID) or interruption of the trajectory.

[0005] (3) Target temporary occlusion and disappearance: The target may be temporarily occluded by furniture, resulting in temporary loss of point cloud. The algorithm needs to have the ability to "remember" the target and resume tracking within a certain period of time.

[0006] (4) Scenario adaptability: Different application scenarios (such as conference rooms and nursing homes) have different requirements for the sensitivity of target tracking and the waiting time for the target to disappear, requiring a flexible configuration mechanism.

[0007] Therefore, there is an urgent need for a human trajectory tracking solution based on top-mounted 4D millimeter-wave radar that has been optimized for real-world scenarios and can effectively address the above-mentioned problems. Summary of the Invention

[0008] To address the technical problems of noise interference, easy loss of tracking across floors, and poor scene adaptability of existing top-mounted millimeter-wave radar in indoor human trajectory tracking, this invention provides a robust and configurable tracking system and method.

[0009] A top-mounted human trajectory tracking system based on millimeter-wave 4D point clouds includes: Millimeter-wave radar module: used to transmit and receive electromagnetic waves and generate raw 4D point cloud data.

[0010] MCU processor: Receives and processes the point cloud data and executes the trajectory tracking algorithm.

[0011] Communication module: Used to receive configuration parameters and upload tracking results.

[0012] Its core innovation lies in the trajectory tracking method executed in the MCU processor, which specifically includes the following steps: For ease of description, the following symbols are defined: Let the original point cloud set obtained by radar analysis in frame t (where t is an integer representing the time sequence) be... ,in This represents the nth point in the frame. This represents the total number of points in the frame.

[0013] Let the currently maintained set of historical targets be... ,in This represents the m-th historical target that existed in the previous frame. Total number of historical goals. Each historical goal... It must contain at least the following attributes: Position coordinates (Cartesian coordinate system); initial position ; Continuous matching frame count (Indicates the number of times the target is matched consecutively by the current cluster); Remaining lifespan (in frames) (Target is deleted when ≤0). Target state (0 indicates candidate / unconfirmed, 1 indicates confirmed); Trajectory queue It is used to store historical location sequences.

[0014] It should be noted that the tracking method of the present invention is based solely on the three-dimensional spatial position of the point cloud. It does not rely on information from other dimensions such as Doppler velocity and signal strength. Therefore, this invention is applicable to low-cost 4D millimeter-wave radar that only outputs position information, and has good versatility.

[0015] Step S1, Point Cloud Input and Coordinate Transformation: Receive the raw point cloud of frame t, and parse it according to the protocol to obtain the point set in spherical coordinates. ;in The radial distance from the radar to the nth point; This is the azimuth angle (horizontal angle). This is the pitch angle (altitude angle).

[0016] Using coordinate transformation formulas: , , To obtain the point set in the Cartesian coordinate system .

[0017] Step S2, noise filtering based on one-dimensional high-density clustering: For mirror points and background noise in top-mounted scenarios, perform the following sub-steps: like ≥3, for all points in the current frame One-dimensional DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering is performed on the coordinates to obtain multiple height sub-layers, each denoted as . And count the number of points in each layer. and average coordinate .

[0018] Excluding average Coordinates greater than 1.2× The sublayer, in which Set the radar installation height (preset parameters), these points are the mirror noise of the ceiling or high places.

[0019] Valid point cloud layers are selected from the remaining sub-layers, and the selection logic is as follows: If the average of all sub-layers All greater than (If the current frame consists entirely of ground-level noise or extremely high mirror points), then the sub-layer with the most points is selected as the effective layer. Otherwise, prioritize the option that satisfies the requirements. < And the sublayer with the most points; If no sub-layer meets the above conditions, then the sub-layer with the most global points is selected.

[0020] A Boolean mask is generated based on the point index of the selected valid layer. The valid point cloud of the current frame is marked, and the point set is obtained after filtering out invalid points. .

[0021] Step S3, Two-dimensional spatial DBSCAN clustering: The effective point cloud after filtering DBSCAN clustering is performed on a plane, with the cluster radius... The neighborhood search radius is fixed at 0.5 m. This value is based on the typical clustering range of the human body's torso in a single-frame point cloud, while also taking into account the average point cloud density in indoor environments. In cases of extremely sparse or dense point clouds, this value can be fine-tuned according to the actual radar angular resolution (e.g., 0.3 m ~ 0.8 m). The fine-tuning principle is: when the radar's horizontal angular resolution is high (e.g., ≤ 2°) and the installation height is low (≤ 3 m), the value can be appropriately reduced. (e.g., 0.3 m); conversely, when the angular resolution is low (e.g., ≥ 4°) or the installation height is high (≥ 4 m), the value can be appropriately increased. (e.g., 0.8 m) to ensure that human targets can be clustered into a coherent cluster.

[0022] Get the current frame target clusters and noise point set Each cluster Including its internal point set, the number of points in the cluster is .

[0023] The minimum number of samples for clustering is dynamically adjusted based on the total number of valid points and the number of currently active targets; for example, let the total number of valid points in the current frame be... The number of targets confirmed in the previous frame was (i.e., state) If the target number is 1, then the minimum sample size is... Determined according to the following rules: like <50, then =3; If 50≤ <200, then =5; like ≥200, then =7; like >5, then Add 2 extra; The above values ​​can be adjusted according to the actual radar point cloud density.

[0024] Step S4, cross-frame target association and "noise continuation" mechanism: Location extraction: For each current cluster Calculate its plane center: Previous frame historical target The position is .

[0025] Hungarian match: and The cost matrix is ​​constructed using the Euclidean distance between them, and a maximum matching distance threshold is set. (Typical value 2.0m). In the cost matrix, if the Euclidean distance between a pair of historical targets and the current cluster is greater than... If the cost at that position is set to infinity (or a maximum value), the Hungarian algorithm is forced to prohibit the match. This threshold is set based on the maximum possible distance a human can move between two adjacent frames (considering frame rate and human walking speed) and can be adjusted according to the actual scenario.

[0026] After matching is completed, if the distance of a matching pair is still greater than 1 / 3, then... If a match is possible due to multiple finite values ​​in the cost matrix, then the match is rejected, and the match is removed from the set of matched pairs and assigned to the unmatched historical target and the unmatched current cluster, respectively.

[0027] The Hungarian algorithm was applied to complete the matching, and the result was: Matching pair set ,in For the number of matching pairs; Unmatched historical target set , ; No match in the current cluster set , ; High mutation detection: If in a matching pair, the average value of the current cluster is... Coordinates and historical goals The difference exceeds the preset threshold If the target is detected, it is marked, and an adaptive observation noise covariance is used in subsequent Kalman filtering. Specifically, when a height abrupt change (i.e., ...) is detected, the target is marked. When the Kalman filter observation noise covariance matrix corresponding to the matching pair is... Temporarily enlarged to the original size times (e.g.) =10), which makes the predicted values ​​more trusted than the observed values ​​during filter updates, thereby suppressing position jumps. After three consecutive frames without detecting any further height changes of the target, recovery is initiated. Set to the initial value to prevent trajectory jumps caused by cross-layer activities.

[0028] It should be noted that each objective It independently maintains its own observation noise covariance matrix The initial values ​​are all =diag(0.05,0.05); This represents a diagonal matrix. When the target is marked as having a height abrupt change, the current frame update uses... = · Otherwise use This flag is only valid for the current frame; the next frame will reassess the flag based on whether another height mutation is detected.

[0029] Noise matching (“life extension” mechanism): Unmatched current cluster All points are merged into the noise point set. It should be noted that the inclusion of points that do not match the current cluster into the noise set is only used to assist in the continuation determination of unmatched historical targets, and does not affect the subsequent candidate status of these clusters as new targets.

[0030] Iterate through each unmatched historical target : in its position Surrounding radius (Typical value 1m, parameter) The selection of 1m is based on the following: the uncertainty of the human body's position in a single-frame point cloud due to occlusion or movement is usually less than 1 meter, and this radius is sufficient to cover the area that the human body may occupy; at the same time, an excessively large radius would introduce too many noise points, leading to false positives. Within the circular neighborhood (where the human body is within the range of 1m), the following values ​​are statistically analyzed: Points .like > ( Noise point threshold, typically 2, parameter The selection of 2 is based on the following: even if the point cloud is relatively sparse, as long as the target actually exists in the area, due to radar reflection characteristics, there will usually be at least two noise points (or sidelobe points) falling around the target; if there are fewer than two noise points, it is highly probable that there is no actual target in the area. The above parameters have been optimized through extensive field testing (users can also fine-tune them according to the radar model and installation height), thus setting the remaining lifetime. Increase Frames (typically 5), with an upper limit constraint applied: ← ,in This is the maximum lifespan limit (typically 100 frames).

[0031] It should be noted that incorporating points from unmatched current clusters into the noise set is only used to assist in determining the continuation of unmatched historical targets. This operation does not affect the fact that these clusters are subsequently considered as candidates for new targets in step S6. That is, points from the same cluster can serve as noise points to help other targets extend their lifespan, or they can be used in subsequent steps to generate new historical targets. The two are executed sequentially and do not conflict with each other.

[0032] Step S5, Kalman filter update: For each matching pair The observation is taken as the center of the current cluster plane: , Based on the Kalman prediction state of the historical target in the previous frame, this invention employs a uniform motion model (CV model). For each historical target... Its state vector is defined as = ,in , ρ represents the velocity in the x and y directions, respectively. State transition matrix. and observation matrix The settings are as follows: , ; in The time interval between two adjacent frames is determined by the radar's operating frame rate (e.g., when the frame rate is 10Hz). =0.1s). The prediction and update process follows the standard Kalman filter recursive formula. Specifically, let the target... The estimated state vector in the previous frame (frame t-1) is: The estimated covariance matrix is The prediction and update process of Kalman filtering is as follows: Prediction steps: ; ; Update steps: ; ; ; in The process noise covariance matrix is ​​set as follows: =diag(0.01,0.01,0.1,0.1) (Typical values, units correspond to state variables; this setting means: the process noise variance of the position component is relatively small (0.01 m) 2 This reflects the smooth change in the human body's position between adjacent frames; the process noise variance of the velocity component is relatively large (0.1 (m / s)). 2 (Speed ​​can vary to some extent (e.g., acceleration or deceleration); the above values ​​are obtained through optimization using actual motion data, and users can adjust them appropriately according to the expected intensity of the exercise); all targets use the same... ; To observe the noise covariance matrix, For the goal The observation noise covariance matrix, initial value =diag(0.05,0.05) (unit: m) 2If the target is marked as highly abrupt in step S4, then the current frame update will take... = · ( =10); otherwise take = .

[0033] Updated location estimation Taken from The first two components.

[0034] Perform a standard Kalman filter update to obtain a new position estimate for the target. And write that position to its trajectory queue: ← .

[0035] Step S6, Target Lifecycle Management: New target: For each unmatched current cluster Create new historical goals ,initialization: = MC=0, L= F=0, where The initial lifespan of the candidate target is set (typically 10 frames), and the trajectory queue is initialized to the current center.

[0036] Simultaneously, the state vector of the Kalman filter is initialized for the new target. = The position component is initialized to the cluster center. The velocity components are initialized to 0. The initial state covariance matrix... Set to: , Wherein the variance =0.1 (unit: m) 2 ), velocity variance =1.0 (unit (m / s)) 2 This reflects the uncertainty of the initial velocity. In each subsequent frame update, if the target is successfully matched, the state is corrected according to the Kalman update formula in step S5; if the target is not matched, only the prediction step is performed (no update), and the velocity state decays naturally over time.

[0037] Matched target update: For each matched historical target : ← +1; like =0 and > ( To determine the threshold for the number of consecutive matching frames required for the target (typically 10 frames), then: ←1, ← ,in The lifespan of a confirmed target (typically 100 frames).

[0038] Unmatched target update and removal: For each unmatched historical target : ←0, ← -1; like If ≤0, then remove the target from the historical target set. Remove from the queue and complete its trajectory. (Including all historical locations of the target from its appearance to its disappearance) is output to the communication module as the target's trajectory log for uploading or local storage.

[0039] Regionalization rules (configurable): For a target marked as being located within the "forbidden zone", if its consecutive number of unmatched frames exceeds [a certain threshold], [further action will be taken]. (Typical value 10 frames), then set directly. =0 and remove.

[0040] For those located in the "Rest Area" and in a certain state The confirmed target with a value of 1 can be... Lock it to the maximum value to ensure it never disappears.

[0041] To implement the above regionalization rules, special areas are pre-defined on the two-dimensional plane (x,y) using polygons (such as rectangles or convex polygons) and labeled with their types (prohibited areas, rest areas, etc.).

[0042] Specifically, each region consists of a set of polygon vertex coordinates. Define and associate a type label (forbidden zone, rest area, etc.). After each frame update, based on the target location... The ray casting algorithm is used to determine whether a target is located within a certain area. If a target is located in multiple areas simultaneously, it is processed according to priority: prohibited areas have higher priority than rest areas.

[0043] After each frame is updated, based on the target position Determine its region. The priority rules are as follows: For targets located in the "restricted zone," their lifecycle management is entirely handled by independent counters. Control, global remaining lifespan The decrementing operation is ignored. Specifically: during the unmatch processing of each frame, it is not executed. ← -1, but rather added separately. ;when ≥ If the target is not found, remove it immediately. Once the target is matched again, Cleared and returned to normal Management (If the target is still in the prohibited zone, it will continue to be managed by the prohibited zone rules when it re-enters without a match).

[0044] For those located in the "Rest Area" and in a certain state The target with a value of 1 is reduced in each "Unmatched Target Update and Removal" execution. Before the operation, force Reset to (Or a very large value, such as 9999 frames), thus ensuring it never disappears. This rule has higher priority than the global rule. Decreasing.

[0045] Through the above steps, the system completes the processing of frame t and outputs the updated historical target set. And the trajectories of each target.

[0046] Compared with the prior art, the present invention has the following significant advantages: (1) Height-adaptive noise filtering for top-mounted scenarios significantly reduces mirror point and background noise interference: Existing methods typically employ fixed height thresholds or global filtering, which struggle to effectively distinguish between real ground target points and mirrored points generated by ceilings or walls. This invention introduces one-dimensional DBSCAN height clustering for the first time, and based on a layer selection logic that prioritizes the layer with the most points below the installation height, it adaptively identifies and retains the layer at the height of the real target while excluding invalid mirrored points higher than 1.2 times the installation height. Even in extreme frames with sparse point clouds or entirely noisy conditions, a backoff strategy selects the globally most reliable layer, ensuring the algorithm's robustness.

[0047] (2) An innovative "noise density lifespan" mechanism effectively addresses temporary target occlusion and sparse point clouds: Traditional tracking algorithms often rapidly reduce the target's lifetime and delete it when the target is briefly lost due to furniture occlusion or radar angle, leading to trajectory interruption. This invention addresses this by statistically analyzing the noise point density in the neighborhood of unmatched historical targets. When the number of noise points exceeds a threshold, it proactively increases the target's remaining lifetime, providing an "extension" for the target. This mechanism does not rely on continuous matching; it maintains the target's identity using only sparse point cloud information in the environment, significantly reducing the probability of target ID switching and trajectory breakage caused by short-term occlusion.

[0048] (3) Adaptive Kalman filtering for abrupt changes in cross-layer height to prevent target position jumps and mismatches: Under top-mounted radar, the point cloud of a human body moving at different heights (such as from the ground to a sofa or bed) The coordinates can change drastically, and traditional filters that directly fuse observations can cause abrupt changes in trajectory position or even misassociation with other targets. This invention detects height abrupt changes during the matching process and dynamically adjusts the observation noise covariance of the Kalman filter or introduces an adaptive correction factor to reduce the filter's confidence in new observations, thereby maintaining a smooth trajectory transition. This mechanism effectively solves the problem of unstable cross-layer tracking in top-mounted scenarios.

[0049] (4) Regionalized target lifecycle management greatly improves scenario adaptability: Existing systems generally use globally uniform target persistence parameters, which cannot accommodate the differentiated needs of various privacy-sensitive locations such as conference rooms, nursing homes, and hotels. This invention allows users to pre-configure special areas such as restricted zones and rest areas, and set independent lifecycle rules for targets in different areas (e.g., targets in restricted zones disappear after 10 consecutive frames lost, while confirmed targets in rest areas never disappear). This regional parameter configuration mechanism allows this invention to flexibly adapt to various application scenarios and meet different monitoring strategy requirements without modifying the core algorithm.

[0050] (5) Dynamically adjust clustering parameters to adapt to real-time changes in the number of targets and point cloud density: In the DBSCAN clustering step, this invention dynamically adjusts the minimum number of samples for clustering based on the total number of valid points and the number of active targets in the current frame. When the environment is open and there are few targets, a smaller minimum number of samples is used to detect weak targets; when targets are dense and the point cloud is dense, the minimum number of samples is increased to avoid over-segmentation. Compared with fixed-parameter clustering, this dynamic adjustment strategy has higher detection stability and computational efficiency.

[0051] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description

[0052] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0053] Figure 1 This is a simplified system structure block diagram of the top-mounted human trajectory monitoring device according to an embodiment of the present invention.

[0054] Figure 2 This is a flowchart illustrating the operation of the top-mounted human trajectory monitoring device according to an embodiment of the present invention.

[0055] Figure 3 This is a circuit diagram of a millimeter-wave radar in an embodiment of the present invention; in the diagram, LAN represents a low-noise amplifier, PGA represents a programmable gain amplifier, PA represents a power amplifier, ADC represents an analog-to-digital converter, and FMCW Waveform Generator represents a linear frequency modulated continuous wave generator.

[0056] Figure 4 This is a flowchart of the human trajectory tracking algorithm executed inside the MCU processor in this embodiment of the invention, which specifically illustrates the complete processing flow from point cloud input to target lifecycle maintenance. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] I. System Composition: like Figure 1 , Figure 3 As shown, the human trajectory monitoring device system provided in this embodiment includes: Millimeter-wave radar module: adopts a 4T4R architecture, including radio frequency front-end (transmit link TX, receive link RX) and baseband processing end (ADC, microcontroller unit, point cloud generation unit) for generating 4D point clouds.

[0059] MCU processor: The core computing unit that runs the trajectory tracking algorithm described in this invention.

[0060] Communication module (such as 4G / WiFi module): Communicates with the cloud server via MQTT (Message Queuing Telemetry Transport) protocol, receives configuration instructions, and reports the tracking trajectory.

[0061] II. Workflow: like Figure 2 As shown, the system workflow is as follows: (1) Equipment installation and configuration: The equipment is mounted on the indoor ceiling with the radar facing downwards. Users can configure parameters such as installation height, monitoring range, and special areas (such as "rest area" and "no-entry zone") through a mini-program. The parameters are sent to the equipment via the cloud.

[0062] (2) Radar control and data reception: The MCU configures and starts the radar module through AT command (Attention Command) to continuously receive 4D point cloud data frames.

[0063] (3) Core tracking processing: The MCU executes the following steps S1 to S6 to process the point cloud of each frame and update the trajectory of each target.

[0064] (4) Result upload: Package the human body location, trajectory and other information obtained from tracking and upload it to the cloud for unified management through the communication module.

[0065] III. Detailed algorithm steps, see [link / reference] Figure 4 : In this embodiment, the typical values ​​of the relevant parameters are as follows: Radar installation height =3.0m (configurable); Maximum survival time =100 frames; Initial lifespan of candidate targets =10 frames; Target lifecycle confirmed =100 frames; Noise matching radius =1.0m, noise point threshold =2, number of frames for continued life =5 frames; Target confirmation threshold =10 frames; High mutation threshold =0.5m; Radar operating frame rate =10Hz, corresponding to the time interval between two adjacent frames. =1 / =0.1s. This frame rate can be adjusted according to the actual radar configuration (e.g., =5Hz~20Hz), Corresponding synchronous changes.

[0066] Taking a specific scenario as an example: at a certain moment, frame t contains =256 points. During step S2, one-dimensional DBSCAN clustering yields two height sub-layers: Sub-layer 1 average... =0.2m (corresponding to human body points near the ground), number of points 120; sub-layer 2 average =2.9m (approximately 3.0m installation height), 80 points. Average of factor layer 1. Less than Furthermore, due to the large number of points, it was selected as an effective layer. After filtering out sublayer 2, the effective point set... It contains 120 points.

[0067] Subsequently, DBSCAN clustering was performed on the XY plane to obtain... =2 clusters , Historical target set Includes =2 targets , After matching with Hungary, and match, and match, =2, no unmatched historical targets and no unmatched current clusters.

[0068] At this point, noise matching is unnecessary. Kalman filtering is applied to each of the two matched pairs separately. Assume... of The coordinates changed from 0.3m in the previous frame to the current cluster. average The value is 0.2m, and the change is less than Normal updates; if a target suddenly rises 1.2m from the ground (e.g., climbs onto a bed), the change is 0.9m > Then, the observation noise covariance of the filter is marked and increased, so that the weight of the new position on the state correction is reduced, thus preventing position jumps.

[0069] After the update, , The consecutive matching frame count (MC) is incremented by 1. If the previous MC has exceeded 10 and has not yet been confirmed, its status F is set to 1 and its lifespan L is set to 100. Finally, the updated position is written to its respective trajectory queue, and it waits for the next frame of data.

[0070] Those skilled in the art should understand that the above parameter values ​​are merely examples and can be adjusted according to the actual installation height, indoor environment, and application requirements. The algorithm flow and symbol system protected by this invention follow the above definitions, and any equivalent substitutions or parameter adjustments that do not depart from the core steps of this invention fall within the protection scope of this invention.

Claims

1. A top-mounted human trajectory tracking method based on millimeter wave 4D point cloud, characterized in that, Includes the following steps: Step S1: Receive the t-th frame of the original 4D point cloud output from the millimeter-wave radar, and convert the point cloud from spherical coordinates to Cartesian coordinates to obtain the point set. ; Step S2: For all points in the current frame One-dimensional DBSCAN clustering was performed on the coordinates to obtain multiple height sub-layers, excluding the average. Coordinates greater than 1.2× The sublayer, in which The radar installation height is determined, and then the average value is preferentially selected from the remaining sublayers. Coordinates less than The sub-layer with the most points is selected as the valid layer. If no valid layer exists, the sub-layer with the most global points is selected. A Boolean mask is generated based on the valid layer's point index, and invalid points are filtered out to obtain the valid point cloud. ; Step S3: For the effective point cloud... Perform DBSCAN clustering on the plane to obtain the current frame. target clusters and noise point set The minimum number of samples for clustering is dynamically adjusted based on the total number of valid points in the current frame and the number of confirmed targets in the previous frame. Step S4: Calculate each current cluster plane center ,by Historical target position in the previous frame The cost matrix is ​​constructed using the Euclidean distance between them. The Hungarian algorithm is then applied for matching to obtain matched pairs, unmatched historical targets, and unmatched current clusters. For unmatched historical targets radius around its location Neighborhood statistics belong to Noise points ,like > , If the noise point count threshold is used, then its remaining lifespan is... Increase Frames, and not exceeding the upper limit. ; Step S5: For each matching pair, using the current cluster plane center as the observation value, update the position of the historical target using a uniform motion Kalman filter. If the average value of the current cluster is detected... Coordinates and Historical Goals The difference in coordinates exceeds a preset threshold Then the observation noise covariance matrix corresponding to the target will be... Temporarily enlarged to the original size After three consecutive frames without detecting any further height abrupt changes, the value returns to its initial value. Step S6: Update the lifecycle of historical targets based on the matching results: Create a new target for each unmatched current cluster, and initialize the remaining lifetime to 0. ; For a matched target, increase its consecutive matching frame count. When the consecutive matching frame count exceeds a threshold... Set its status to confirmed and extend its lifespan to [time period]. ; For unmatched historical targets, decrement their remaining lifespan by 1. When the remaining lifespan is ≤ 0, remove the target and output its trajectory queue. For targets within pre-defined no-match zones, an independent counter records the number of consecutive unmatched frames; once a threshold is reached, the target is removed directly, ignoring the global decrease in remaining lifespan. For confirmed targets within rest areas, their remaining lifespan is forcibly reset to [a certain value]. Or the maximum value.

2. The method according to claim 1, characterized in that, In step S3, the neighborhood radius of DBSCAN clustering The fixed setting is 0.5 m, and it can be finely adjusted within the range of 0.3 m to 0.8 m according to the radar horizontal angular resolution and installation height.

3. The method according to claim 1, characterized in that, In step S3, the minimum number of samples The following rules apply to dynamic adjustments: If the total number of valid points in the current frame... < 50, then =3; if 50≤ < 200, then =5; if ≥ 200, then =7; if the target number was confirmed in the previous frame >5, then Add 2 extra.

4. The method according to claim 1, characterized in that, In step S4, the Hungarian matching also has a maximum matching distance threshold. When the Euclidean distance between the historical target and the current cluster is greater than In this case, the corresponding position in the cost matrix is ​​set to infinity to prohibit matching. After matching is completed, if the distance of a matching pair is still greater than a certain value, the matching will be performed. Then reject the match.

5. The method according to claim 1, characterized in that, In step S4, the radius of noise matching =1.0 m, noise point threshold =2, number of frames for continued life =5 frames.

6. The method according to claim 1, characterized in that, In step S5, the state vector of the Kalman filter is = State transition matrix Observation matrix ,in , These are the velocities in the x and y directions, respectively. This represents the time interval between two adjacent frames.

7. The method according to claim 1, characterized in that, In step S5, each target independently maintains its own observation noise covariance matrix. The initial value is =diag(0.05,0.05), amplification factor for high abrupt changes. =10.

8. The method according to claim 1, characterized in that, In step S6, the initial survival period of the candidate target. =10 frames, target lifespan confirmed. =100 frames, maximum survival time limit =100 frames, target confirmation threshold =10 frames.

9. The method according to claim 1, characterized in that, The threshold for the number of consecutive unmatched frames of a target within the prohibited zone. =10 frames.

10. A top-mounted human trajectory tracking system based on millimeter-wave 4D point clouds, characterized in that, include: Millimeter-wave radar module, used to transmit and receive electromagnetic waves and generate raw 4D point cloud data; An MCU processor is configured to perform the method according to any one of claims 1 to 9; The communication module is used to receive configuration parameters and upload tracking results.

Images