24G millimeter wave radar multi-target tracking method and system based on dynamic and static point cloud separation and self-adaption

Abstract

The invention discloses a 24G millimeter wave radar multi-target tracking method and system based on dynamic and static point cloud separation and self-adaption. The method comprises the following steps: firstly, obtaining dynamic and static point cloud data of a millimeter wave radar through a radio frequency signal received by the radar; setting a distance threshold and an energy threshold, carrying out pre-filtering on invalid point clouds, and removing dynamic and static interference points which exceed a radar action range and are too low in reflection energy; then matching the preprocessed dynamic and static point cloud data with an existing target list, wherein the matching process is based on dual criteria of spatial distance and speed consistency; for isolated point clouds which cannot be matched with any existing target, starting a target new establishment judgment mechanism; and finally, after the target is successfully established and enters an activated state, starting a tracking process. According to the method, the precision, robustness and system efficiency of target tracking are remarkably improved, the problem of tracking loss of edge scenes such as crossing pedestrians is solved, precise adaptation to complex states such as static states, crossing states and conventional motion states is achieved, and the state estimation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar tracking technology, and in particular to a 24G millimeter-wave radar multi-target tracking method and system based on dynamic and static point cloud separation and adaptive tracking. Background Technology

[0002] With the widespread application of radar in fields such as autonomous driving and intelligent transportation, target tracking technology based on radar point clouds has become a key link in environmental perception. Radar obtains the distance, velocity, and azimuth information of a target through echo signals, but its output point cloud data is often affected by factors such as multipath effects, environmental clutter, and target occlusion, resulting in problems such as sparsity, fragmentation, and noise, which pose challenges to target identification and continuous tracking.

[0003] Existing tracking systems commonly suffer from false target detection, output jitter, and redundant target residue when processing this type of data. Due to the lack of a comprehensive discrimination mechanism considering velocity, point cloud density, and distance, the system easily identifies transient noise as real targets. Smoothing strategies often employ fixed parameters, making it difficult to balance near-field response speed with far-field stability. Furthermore, traditional methods lack effective identification and removal mechanisms for redundant targets that are spatially close and exhibit harmonic velocities. In addition, target activation and resource management strategies are relatively fixed and lack adaptability.

[0004] Therefore, there is an urgent need for a target tracking method that can dynamically suppress interference, adaptively process output, and improve tracking reliability. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a 24G millimeter-wave radar multi-target tracking method and system based on dynamic and static point cloud separation and adaptation. This method solves the problems of false target interference, poor stability, redundancy residue and low resource utilization in the target tracking system of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The multi-target tracking method for 24G millimeter-wave radar based on dynamic and static point cloud separation and adaptation provided by this invention includes the following steps: The radio frequency signals received by the radar are processed to generate dynamic and static point cloud data of the millimeter-wave radar. By setting distance and energy thresholds, invalid point clouds are pre-filtered to remove dynamic and static interference points that are outside the radar's range or have too low reflected energy. The preprocessed static and dynamic point cloud data are matched with the existing list of tracking targets. The matching process is based on the dual criteria of spatial distance and velocity consistency. For isolated point clouds that fail to match any existing targets, initiate a target creation determination mechanism; Once the target is successfully established and enters the active state, the tracking process is initiated.

[0007] Furthermore, the step of generating dynamic and static point cloud data includes: Perform a range-dimensional Fast Fourier Transform on the signal to calculate the target distance information; The signal is divided into a dynamic point cloud generation branch and a static point cloud generation branch; In the dynamic point cloud branch, after filtering out static targets through the moving target display MTI, a velocity dimension FFT is performed to conduct two-dimensional CFAR detection. In the static point cloud branch, the velocity dimension FFT is directly performed to perform two-dimensional CFAR detection on the zero-frequency region; The angle of arrival of the target is estimated by phase detection to obtain angle information.

[0008] Furthermore, in the pre-filtering step, the Euclidean distance, angle, and energy between the measuring point and the starting origin are calculated to determine whether the point is within the set range of distance, angle, and energy, and invalid point clouds outside the range are filtered out.

[0009] Furthermore, in the matching step, the Euclidean distance and velocity difference between the point cloud and the estimated position of the target are calculated. If the distance is less than a preset spatial threshold and the velocity difference is within the allowable range, the matching is determined to be successful.

[0010] Furthermore, the target creation determination mechanism includes: Determine whether the point cloud meets the false target suppression conditions: the velocity is greater than the minimum velocity threshold, the number of matched point clouds is less than the minimum point cloud requirement, and the distance is greater than the distance threshold; If the conditions are met, skip the new process; If the conditions are not met, a new target is allocated in an idle tracking slot, and the target parameters are initialized. The activation cycle is dynamically set based on the target distance and the number of matching point clouds.

[0011] Furthermore, the trajectory prediction step in the continuous tracking process includes: Determine the target's movement pattern based on its historical state; For stationary targets, the current state is used as the predicted value; An acceleration compensation model is used for targets that cross the target. A uniform acceleration model is used for conventional moving targets.

[0012] Furthermore, the point cloud matching step includes: Perform intersection scene prediction and pre-screen based on the consistency of motion between point cloud and target; The matching threshold is dynamically adjusted based on the target state and environment. Adaptive matching range expansion is implemented for distant or boundary targets; The target with the smallest distance is selected as the matching result.

[0013] Furthermore, the state update step employs an improved Alpha-Beta-Gamma filter, selecting different update strategies based on the target state and the number of matching point clouds.

[0014] Furthermore, the target disappearance determination step includes: Based on the target speed, it can be classified as static, slightly moving, or in motion. Combined with a hysteresis mechanism to prevent state jumps; Set the release counter according to the differences in target status and location; Reduce release delay for close-range backward movement scenarios.

[0015] The present invention provides a millimeter-wave radar multi-target tracking system based on dynamic and static point cloud separation and adaptation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method. The beneficial effects of this invention are as follows: This invention proposes a multi-target tracking method and system for 24G millimeter-wave radar based on static / dynamic point cloud separation and adaptation. Through innovative technical means, it significantly improves the accuracy, robustness, and system efficiency of target tracking. First, by combining pre-filtering (distance / energy threshold) with false target suppression conditions (velocity, point cloud quantity, distance threshold), it effectively eliminates high-speed, low-point-cloud noise and invalid interference points, reducing the false alarm rate. Combined with a matching strategy based on the consistency of spatial distance and velocity, supplemented by adaptive matching range expansion (for long-range / boundary targets), it solves the tracking loss problem in edge scenarios such as pedestrians crossing. Second, by employing multi-model prediction (uniform acceleration model, acceleration compensation model, etc.) and improved Alpha-Beta-Gamma filtering, it achieves accurate adaptation to complex states such as stationary, crossing, and normal motion, improving the accuracy of state estimation.

[0016] In terms of resource management, a dynamic activation cycle setting (3 frames for near-field / 2 frames for far-field) balances response speed and resource consumption. Combined with a heartbeat counter, differentiated release thresholds (for moving and static targets), and a hysteresis anti-shake mechanism, invalid trajectories are recovered, reducing redundancy and improving system stability. For complex environments, a novel approach separates dynamic and static point cloud processing (MTI filtering + zero-frequency CFAR detection), combined with phase detection and angle measurement technology to accurately acquire target 3D information. By statistically analyzing the number of point cloud matches and determining the validity of trajectory initialization, occlusion and fragmentation issues in high-density target scenarios are optimized. Furthermore, the algorithm employs a lightweight design, avoiding complex Kalman filtering, adapting to embedded hardware deployment, meeting the low-power requirements of millimeter-wave radar, and enhancing environmental adaptability through threshold adjustments driven by dynamic parameters such as distance and velocity (e.g., matching range, activation cycle). Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0017] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0018] Figure 1 This is a flowchart of the 24G radar multi-target tracking process in this embodiment; Figure 2 This is a flowchart of the dynamic and static point cloud generation process in this embodiment; Figure 3 This is a flowchart of the target creation determination mechanism in this embodiment; Figure 4 This is a flowchart of the multi-target continuous tracking system in this embodiment; Figure 5 This is a diagram showing the tracking results of the cross motion of moving and stationary targets in this embodiment; Figure 6 This is a diagram showing the tracking results of the lateral cross motion of two people in this embodiment; Figure 7 This is a diagram showing the results of two-person radial cross motion tracking in this embodiment. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Example 1 like Figure 1As shown, the 24G millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation provided in this embodiment includes the following steps: Step 1: Process the radio frequency signals received by the radar to generate millimeter-wave radar dynamic and static point cloud data; Step 2: By setting distance and energy thresholds, invalid point clouds are pre-filtered to effectively remove dynamic and static interference points that are outside the radar's range or have too low reflection energy, thereby improving the data quality for subsequent processing.

[0021] Step 3: Match the preprocessed static and dynamic point cloud data with the existing list of tracking targets. The matching process is based on the dual criteria of spatial distance and velocity consistency.

[0022] Step 4: For isolated point clouds that failed to match any existing targets in Step 2, initiate the target creation determination mechanism.

[0023] Step 5: After the target is successfully established and enters the active state, the system starts the continuous tracking process, including trajectory prediction, point cloud matching, status update and target disappearance determination.

[0024] Furthermore, step one specifically includes the following sub-steps: Step 1.1: Perform a Fast Fourier Transform on the signal in the distance dimension, and process the echo signal in the time domain using pulse compression technology to calculate the target's distance information; Step 1.2: The signal is processed by dividing it into two branches: dynamic point cloud generation and static point cloud generation. Step 1.3: For dynamic point clouds, static targets are removed by moving target display MTI filtering technology, and then a velocity dimension fast Fourier transform is performed to obtain the range-Doppler spectrum. Two-dimensional constant false alarm rate (CFAR) detection is performed to filter out target points and their intensity information. Step 1.4: For static point clouds, directly perform velocity-dimensional fast Fourier transform to obtain the range-Doppler spectrum, perform two-dimensional constant false alarm rate (CFAR) detection on zero frequency, and filter out static target points and their intensity information. Step 1.5: Perform phase detection to estimate the target's angle of arrival and obtain the target angle information; Furthermore, in step two, by setting distance and energy thresholds, the invalid point cloud is pre-filtered to effectively remove dynamic and static interference points that are outside the radar's effective range or have too low reflected energy. Specifically, the starting origin, angle measurement range, distance measurement range, and energy range are set. Based on the calculated Euclidean distance, angle, and energy between the measurement point and the starting origin, it is determined whether the point is within the range. That is, when the distance, angle, and energy are all within the preset range, the processed dynamic and static point cloud is obtained.

[0025] Furthermore, in step three, the preprocessed static and dynamic point cloud data is matched with the existing list of tracking targets. The matching process is based on dual criteria of spatial distance and velocity consistency. Specifically, the Euclidean distance between each point cloud in the current frame and the estimated position of each existing target is calculated, and their velocity differences are compared. If the distance is less than a preset spatial threshold and the velocity difference is within the allowable range (for example, the distance difference between the point cloud and the target is less than 1m and the velocity difference is less than 0.5m / s, or specify a specific numerical range), then the point cloud is determined to belong to the corresponding target, which is used as the observation input for the target in the current frame, and its matching point cloud quantity and status information are updated.

[0026] Furthermore, in step four, the target creation determination mechanism is as follows: the system first determines whether the point cloud meets the false target suppression conditions, namely: its speed is greater than the minimum speed threshold, the number of matching point clouds is less than the minimum point cloud requirement, and the distance is greater than the set distance threshold; if the conditions are met, it is considered to be high-speed low point cloud noise, and the target creation process is skipped to suppress false alarms.

[0027] For isolated point clouds identified through false target suppression, the system iterates through all tracking slots within a preset maximum tracking range, searching for slots that are currently idle. Once an available slot is found, it is allocated to construct a new target and its target identifier, state vector, number of matching point clouds, heartbeat count, and initial activation period are initialized.

[0028] The activation period is dynamically set based on the distance to the new target and the current number of matching point clouds. Specifically: if the target distance is less than 250 cm, the activation period is set to 3 frames to achieve rapid response to near-field targets; if the distance is greater than or equal to 250 cm, the number of matching point clouds is further determined: if there are no less than 8, the activation period is also set to 3 frames; otherwise, it is set to 2 frames to enhance the stability of the establishment process in sparse point cloud scenarios.

[0029] After a new target is successfully assigned a slot and initialized, it enters the transient tracking phase. The system continuously monitors its matching status in subsequent frames, and only when the target is continuously detected and successfully matched for the set number of activation period frames will its status be changed to active state.

[0030] Furthermore, in step five, after the target is successfully established and enters the active state, the system initiates a continuous tracking process, which specifically includes the following steps: Step 5.1: Predict the trajectory of the activated target. The system first determines the current motion pattern of the target based on its historical state sequence, and then uses the corresponding prediction model to generate the state estimate for the next moment.

[0031] Specifically, set goals In the At frame rate, its horizontal position in the past 10 frames is stored in an array. In, among them, For the latest value, the system calculates the lateral displacement change between the three most recent consecutive frames: ; ; ; If satisfied If so, the target is determined to be in a traversing state, and the traversing direction is determined according to the sign of the displacement: ; If the conditions are not met, and , , Meanwhile, the current radial distance If the number of currently matched point clouds is not less than 3 and comes from static point cloud matching, then the target is determined to be in a static state.

[0032] in, This represents the difference between the target's distance in frame 9 and the distance in frame 8; This represents the difference between the target's distance in frame 7 and the distance in frame 6; This represents the difference between the target's distance in frame 5 and the distance in frame 4; Indicates the direction of the target's movement; Step 5.2: Use different prediction models for targets in different states: If the state is static, the predicted state is equal to the current smoothed state: ; If the object is traversing the lateral direction, an acceleration compensation model is used for prediction. The lateral position prediction considers a fixed offset direction (±0.5 m / s) and an acceleration term: ; The vertical position remains unchanged: Radial distance and velocity are updated using kinematic equations: ; ; in Where is the frame interval. , This is an estimate of historical acceleration.

[0033] If the motion is in a normal state, then a uniform acceleration model is used for prediction: ; ; ; ; This predicted state will be used for point cloud matching and state updates in subsequent frames.

[0034] in, Indicates the first The first goal in The state prediction vector of the frame; Indicates the first The target of frame prediction is The position coordinates of the direction; Indicates the first The actual target observed in the frame is The position of direction; Indicates the first The target of frame prediction is The position coordinates of the direction; Indicates the first The actual target observed in the frame is The position of direction; Indicates the first Target distance predicted in the frame; Indicates the first The actual target distance observed in the frame; Indicates the first Target velocity predicted in frames; Indicates the first The actual target velocity observed in the frame; Step 5.3: Matching process based on predicted state and current point cloud. The specific steps are as follows: Step 5.3.1: First, perform intersection scene prediction to resolve matching ambiguity caused by close-range interactions between multiple targets. Let the current point cloud... The state is The system iterates through all activated target pairs. Calculate the predicted distance difference between it and the point cloud. , If satisfied If the radial distance difference between the two targets is less than 0.5m, it is determined to be in an "interactive state". At this time, the system performs pre-screening based on the motion consistency between the point cloud and the target: if the point cloud is static ( ), then exclude all dynamic targets ( If the point cloud is dynamic ( ), then exclude all static targets ( For target pairs that are both traversing each other, the candidate target with the closer lateral distance is prioritized for retention. This preprocessing mechanism effectively avoids erroneous association when dynamic and static targets intersect.

[0035] in, Indicates the current state of the point cloud; Indicate target The predicted distance; Representing point clouds and targets The distance difference; Representing point clouds and targets The distance difference; Indicate target The predicted distance; This represents the difference in distance between two targets. Indicate target speed; Indicate target speed; Step 5.3.2: After completing the pre-screening of the intersection scene, the system performs primary matching on the remaining candidate targets. The matching threshold is dynamically adjusted according to the target state and environment: if the point cloud distance is less than 2m, the spatial matching threshold is set to 1m; otherwise, it is 0.5m; if the target is in a "crossing + convergence" state, the threshold is relaxed to 1.2m. The system calculates the point cloud. With each target Euclidean distance: like Then the target Add to the candidate set. In particular, for matching static targets with static point clouds, it is further required that the difference between the lateral and longitudinal positions is less than 1m to compensate for the problem of inaccurate angle estimation of static targets.

[0036] in, Representing point clouds and targets The distance difference; Indicates the first A point cloud coordinate; Indicates the first A point cloud coordinate; Indicate target of coordinate; Indicate target of coordinate; Indicates the distance threshold; Step 5.3.3: For targets at long distances or in boundary areas, the system implements an adaptive matching range expansion strategy. If the radial distance to the target is greater than 4m and the absolute value of the azimuth angle is greater than 40°, the matching threshold is widened to twice the original value; if the distance is greater than 6m and the angle is greater than 30°, it is widened to 1.5 times. This strategy effectively improves the tracking continuity of targets in edge areas, and is especially suitable for pedestrian tracking scenarios that cross the radar field of view boundary.

[0037] Step 5.3.4: Among all targets that meet the distance threshold, the system selects the one with the smallest distance as the matching result. Let the minimum distance be... The corresponding target index is If there exists at least one candidate target, that is... If the match is successful, the point cloud will be determined. Assigned to target It updates its matching point cloud list and counter. Simultaneously, if both the point cloud and the target are static, it increments the target's "static point matching counter" for subsequent weighted smoothing strategy optimization in state updates.

[0038] Step 5.3.5: After matching is completed, the system counts the number of matching point clouds for each target and updates the target status in conjunction with the heartbeat mechanism.

[0039] Step 5.4: After a successful match, update the state vector and perform adaptive smoothing. The specific steps are as follows: Step 5.4.1: Trajectory initialization and matching validity determination; Step 5.4.2: Classification of dynamic and static point clouds and construction of observation vectors; Step 5.4.3: Residual calculation; Step 5.4.4: Improved Alpha-Beta-Gamma filter state update; Step 5.4.5: Handling unmatched targets; Step 5.4.5: Target state determination and disappearance mechanism; Furthermore, in step 5.4.1, for each target trajectory already established in the current frame... (in , (To the maximum number of tracks), perform the following operations: If the trajectory is inactive, skip the subsequent processing; Obtain the current trajectory of the activated target. Matching point cloud set ,in This represents the number of matching points; like If the match is not found, the process will proceed to the unmatch processing step. Execute multi-level spurious target suppression logic: if and If the target is not in the display area and is not crossing the line, it is judged as potential clutter and enters the unmatched processing flow. like If the target is not in the display area and is in a horizontal state, then suppress the update; like or and If it is not in the display area, then suppress the update; If the target is determined to be static and only matched by static points, and and Then, updates will be suppressed.

[0040] in, Indicate target Current distance; Indicates the distance threshold; Indicate target Current horizontal angle; Furthermore, in step 5.4.2, the matching point cloud set is... Each point in Based on their motion properties, they are marked as dynamic points or static points. Definition: in For the current frame number The number of dynamic points matched by each target. For the current frame number The number of static points matched by each target.

[0041] Construct observation vectors The components are calculated using a weighted average: , , , , Furthermore, the state vector is defined in step 5.4.3: ; Calculate the observation residual vector ; in, Represents the observation vector; Represents the observation distance vector; Represents the observed horizontal angle vector; Represents the observed velocity vector; Indicates observation Directional velocity vector; Indicates observation Directional velocity vector; Target The average distance of the point cloud to all matched moving points; Target The number of matched moving points; dynamic; Point cloud The distance; Target The average angle of all moving point clouds matched Point cloud Angle Target The average angle of all moving point clouds matched Point cloud speed ; ; ; ; ; Target exist Predicted distance in the direction; Target exist Predicted distance in the direction; Target exist Predicted acceleration in the direction; Target exist Predicted acceleration in the direction; Furthermore, in step 5.4.4, an update strategy is selected based on the number of matching points and the target state, as detailed below: Scenario 1: Dynamic target update, requirements The estimated state is as follows: , , , , , , , Scenario 2: Static target update, only distance and velocity are updated, other states inherit the predicted values. Furthermore, in step 5.4.5, if the target If no match is found in the current frame, then execute: ; Simultaneously update the heartbeat count and activation flag for both static and dynamic targets.

[0042] in, Updated target of coordinate; Target Predicted coordinate; Target exist The difference between the predicted and observed distances in the directional direction; Target Predicted coordinate; Target Predicted coordinate; Target exist The difference between the predicted and observed distances in the directional direction; Updated target exist Velocity in the direction; Target exist Predicted position in the direction; ; Target exist The difference between the predicted and observed directional velocity values; Updated target exist Velocity in the direction; Target exist Predicted velocity in the direction; Target exist The difference between the predicted and observed directional velocity values; Updated target exist Acceleration in the direction of; Target exist Predicted acceleration in the direction; Updated target exist Acceleration in the direction of; Target exist Predicted acceleration in the direction; ; Indicate target The difference between the predicted and observed acceleration values; Updated target Horizontal angle; Updated target The distance; Updated target speed; Target Predicted speed; Target The difference between the predicted and observed velocity values; ; ; Further, in step 5.4.5, the system determines the motion state based on the updated target speed, classifying it into static, slightly moving, or moving states, and incorporates a hysteresis mechanism to prevent frequent state transitions. The target's activation state is determined by the "appearToActive" counter; when its value exceeds a preset threshold by ten times, the target is marked as active. For unmatched targets, this counter gradually decreases. The trajectory disappearance determination is handled differently based on the target's state and position: moving targets are released when the "threMovingToFree" counter exceeds its limit or the activation count reaches zero; static targets are released based on the "threStaticToFree" counter if they are located within the static detection area, otherwise they are treated as moving targets. For special scenarios such as close-range retreat, the system shortens the release delay and accelerates resource recovery. Target trajectories that meet the release conditions will be cleared, and their state will be set to idle.

[0043] Among them, appearToActive represents the target activation counter; threMovingToFree represents the moving target release counter; and threStaticToFree represents the static target release counter.

[0044] Example 2 This embodiment provides a further detailed description of the method.

[0045] like Figure 1 As shown, this embodiment discloses a 24G millimeter-wave radar multi-target tracking method and system based on dynamic and static point cloud separation and adaptation, including the following steps: Step 1: Process the radio frequency signals received by the radar to generate millimeter-wave radar dynamic and static point cloud data; like Figure 2 As shown, the system first performs Fast Fourier Transform (FFT) processing on the two received signals in the range dimension. Based on the working principle of 24GHz FMCW radar, it uses a linear frequency modulated continuous wave signal to mix with the received echo signal to form an intermediate frequency signal. Then, pulse compression technology is used to accurately process the echo signal in the time domain, enabling accurate calculation of target range information. Specifically, the system transmits a linear frequency modulated signal with a frequency range of 24.0-24.25GHz and a bandwidth of 250MHz.

[0046] Subsequently, the system intelligently divides the processed signal into two independent branches: dynamic point cloud generation and static point cloud generation. In the dynamic point cloud branch, the system first applies a first-order recursive MTI filtering technique to suppress static clutter and significantly highlight dynamic target signals by comparing signals from two consecutive frames. Next, the system performs velocity-dimensional FFT processing on the filtered signal to generate a range-Doppler spectrum, accurately extracting the target's Doppler frequencies to obtain the target's radial velocity information. Finally, the system employs a cell-averaged CFAR algorithm to adaptively filter target points and their intensity information from the range-Doppler two-dimensional data.

[0047] In the static point cloud branch, the system directly performs velocity-dimensional FFT processing on the original signal to generate a range-Doppler spectrum, focusing only on regions where the velocity is zero in the range-Doppler spectrum. Subsequently, the system performs two-dimensional CFAR detection on the zero-frequency region, and adaptively adjusts the detection threshold by statistically analyzing the background noise in the zero-frequency region to accurately identify static target points and their intensity information.

[0048] Finally, the system uses the phase difference between the two received signals to perform phase detection estimation, calculates the target's angle of arrival, and obtains the azimuth information. Specifically, the system calculates the phase difference between the two received signals, Rx1 and Rx2. Combined with antenna spacing and signal wavelength Using the formula Calculate the azimuth of the target.

[0049] By following the steps described above, relatively accurate dynamic and static point cloud information can be obtained.

[0050] Step 2: By setting distance and energy thresholds, invalid point clouds are pre-filtered to effectively remove dynamic and static interference points that are outside the radar's range or have too low reflection energy, thereby improving the data quality for subsequent processing.

[0051] Specifically, the starting origin, angle measurement range, distance measurement range, and energy range are set. Based on the calculated Euclidean distance, angle, and energy between the measurement point and the starting origin, it is determined whether the data is within the range, thus obtaining the processed dynamic and static point cloud.

[0052] Step 3: Match the preprocessed static and dynamic point cloud data with the existing list of tracking targets. The matching process is based on the dual criteria of spatial distance and velocity consistency.

[0053] Specifically, the Euclidean distance between each point cloud in the current frame and the estimated position of each existing target is calculated, and their velocity differences are compared. If the distance is less than a preset spatial threshold and the velocity difference is within the allowable range, the point cloud is determined to belong to the corresponding target, which is used as the observation input of the target in the current frame, and its matching point cloud quantity and status information are updated.

[0054] Step 4: For isolated point clouds that failed to match any existing targets in Step 3, initiate the target creation determination mechanism.

[0055] like Figure 3 As shown, the system first determines whether the point cloud meets the false target suppression conditions, namely: its speed is greater than the minimum speed threshold, the number of matched point clouds is less than the minimum point cloud requirement, and the distance is greater than the set distance threshold. If the conditions are met, it is considered to be high-speed low point cloud noise, and the target creation process is skipped to suppress false alarms.

[0056] For isolated point clouds identified through false target suppression, the system iterates through all tracking slots within a preset maximum tracking range, searching for slots that are currently idle. Once an available slot is found, it is allocated to construct a new target and its target identifier, state vector, number of matching point clouds, heartbeat count, and initial activation period are initialized.

[0057] The activation period is dynamically set based on the distance to the new target and the current number of matching point clouds. Specifically, if the target distance is less than 250 cm, the activation period is set to 3 frames to achieve rapid response to near-field targets; if the distance is greater than or equal to 250 cm, the number of matching point clouds is further determined: if there are no less than 8, the activation period is also set to 3 frames; otherwise, it is set to 2 frames to enhance the stability of the establishment process in sparse point cloud scenarios.

[0058] After a new target is successfully assigned a slot and initialized, it enters the transient tracking phase. The system continuously monitors its matching status in subsequent frames, and only when the target is continuously detected and successfully matched for the set number of activation period frames will its status be changed to active state.

[0059] Step 5: After the target is successfully established and enters the activation state, the system starts the tracking process.

[0060] like Figure 4 As shown, after the target is successfully established and enters the active state, the system starts the continuous tracking process, which specifically includes the following steps: Step 5.1: Predict the trajectory of the activated target. The system first determines the target's current motion pattern based on its historical state sequence, and then uses a corresponding prediction model to generate a state estimate for the next moment. Specifically, let the target... In the At frame rate, its horizontal position in the past 10 frames is stored in an array. Among them This is the latest value. The system calculates the lateral displacement change between the three most recent consecutive frames: , , If satisfied If so, the target is determined to be in a traversing state, and the traversing direction is determined according to the sign of the displacement: If the conditions are not met, and , , Meanwhile, the current radial distance If the number of currently matched point clouds is not less than 3 and comes from static point cloud matching, then the target is determined to be in a static state.

[0061] Step 5.2: Use different prediction models for targets in different states: If the state is static, the predicted state is equal to the current smoothed state: If the object is traversing the lateral direction, an acceleration compensation model is used for prediction. The lateral position prediction considers a fixed offset direction (±0.5 m / s) and an acceleration term: The vertical position remains unchanged: Radial distance and velocity are updated using kinematic equations: , in Where is the frame interval. , This is an estimate of historical acceleration.

[0062] If the motion is in a normal state, then a uniform acceleration model is used for prediction: This predicted state will be used for point cloud matching and state updates in subsequent frames.

[0063] Step 5.3: Matching process based on predicted state and current point cloud. The specific steps are as follows: Step 5.3.1: First, perform intersection scene prediction to resolve matching ambiguity caused by close-range interactions between multiple targets. Let the current point cloud... The state is The system iterates through all activated target pairs. Calculate the predicted distance difference between it and the point cloud. , If satisfied If the radial distance difference between the two targets is less than 0.5m, it is determined to be in an "interactive state". At this time, the system performs pre-screening based on the motion consistency between the point cloud and the target: if the point cloud is static ( ), then exclude all dynamic targets ( If the point cloud is dynamic ( ), then exclude all static targets ( For target pairs that are both traversing each other, the candidate target with the closer lateral distance is prioritized for retention. This preprocessing mechanism effectively avoids erroneous association when dynamic and static targets intersect.

[0064] Step 5.3.2: After completing the pre-screening of the intersection scene, the system performs primary matching on the remaining candidate targets. The matching threshold is dynamically adjusted according to the target state and environment: if the point cloud distance is less than 2m, the spatial matching threshold is set to 1m; otherwise, it is 0.5m; if the target is in a "crossing + convergence" state, the threshold is relaxed to 1.2m. The system calculates the point cloud. With each target Euclidean distance: like Then the target Add to the candidate set. In particular, for matching static targets with static point clouds, it is further required that the difference between the lateral and longitudinal positions is less than 1m to compensate for the problem of inaccurate angle estimation of static targets.

[0065] Step 5.3.3: For targets at long distances or in boundary areas, the system implements an adaptive matching range expansion strategy. If the radial distance to the target is greater than 4m and the absolute value of the azimuth angle is greater than 40°, the matching threshold is widened to twice the original value; if the distance is greater than 6m and the angle is greater than 30°, it is widened to 1.5 times. This strategy effectively improves the tracking continuity of targets in edge areas, and is especially suitable for pedestrian tracking scenarios that cross the radar field of view boundary.

[0066] Step 5.3.4: Among all targets that meet the distance threshold, the system selects the one with the smallest distance as the matching result. Let the minimum distance be... The corresponding target index is If there exists at least one candidate target, that is... If the match is successful, the point cloud will be determined. Assigned to target It updates its matching point cloud list and counter. Simultaneously, if both the point cloud and the target are static, it increments the target's "static point matching counter" for subsequent weighted smoothing strategy optimization in state updates.

[0067] Step 5.3.5: After matching is completed, the system counts the number of matching point clouds for each target and updates the target status in conjunction with the heartbeat mechanism.

[0068] Step 5.4: After a successful match, update the state vector and perform adaptive smoothing. The specific steps are as follows: Step 5.4.1: Trajectory Initialization and Matching Validity Determination For each target trajectory already established in the current frame (in , (To the maximum number of tracks), perform the following operations: If the trajectory is inactive, skip the subsequent processing; Obtain the current trajectory of the activated target. Matching point cloud set ,in This represents the number of matching points; like If the match is not found, the process will proceed to the unmatch processing step. Execute multi-level spurious target suppression logic: if and If the target is not in the display area and is not crossing the line, it is judged as potential clutter and enters the unmatched processing flow. like If the target is not in the display area and is in a horizontal state, then suppress the update; like or and If it is not in the display area, then suppress the update; If the target is determined to be static and only matched by static points, and and Then, updates will be suppressed.

[0069] Step 5.4.2: Classification of Dynamic and Static Point Clouds and Construction of Observation Vectors Matching point cloud set Each point in Based on their motion properties, they are marked as dynamic points or static points. Definition: in For the current frame number The number of dynamic points matched by each target. For the current frame number The number of static points matched by each target.

[0070] Construct observation vectors The components are calculated using a weighted average: , , , Step 5.4.3: Residual Calculation Define state vector Calculate the observation residual vector Step 5.4.4: Improved Alpha-Beta-Gamma Filter State Update The update strategy is selected based on the number of matching points and the target state, as detailed below: Scenario 1: Dynamic target update, requirements The estimated state is as follows: , , , , , , Scenario 2: Static target update, only distance and velocity are updated, other states inherit the predicted values. Step 5.4.5: Handling Unmatched Targets If the target If no match is found in the current frame, then execute: Simultaneously update the heartbeat count and activation flag for both static and dynamic targets.

[0071] Step 5.4.5: Target State Determination and Disappearance Mechanism The system determines the motion state of the target based on the updated target speed, classifying it as static, slightly moving, or moving, and incorporates a hysteresis mechanism to prevent frequent state transitions. The target's activation state is determined by the "appearToActive" counter; when its value exceeds a preset threshold by ten times, the target is marked as active. For unmatched targets, this counter gradually decreases. The disappearance of a trajectory is determined based on the target's state and position: moving targets are released when the "threMovingToFree" counter exceeds its limit or the activation count reaches zero; static targets are released based on the "threStaticToFree" counter if they are within the static detection zone, otherwise they are treated as moving targets. For special scenarios such as close-range retreat, the system shortens the release delay and accelerates resource recovery. Target trajectories that meet the release conditions are cleared, and their state is set to idle.

[0072] Test data and algorithm verification: This invention uses a 24G millimeter-wave radar, some of whose parameters are as follows: Sweep bandwidth 220MHz Starting frequency 24.01GHz Send and receive quantity 1 send 2 receive Number of sampling points 64 Chirp Quantity 32 Maximum distance 8.8m Frame period 50ms Depend on Figure 5 As shown, the scenario is an indoor environment, with a moving target and a stationary target moving in opposite directions. The result image shows that when the moving target passes by the stationary target, the stationary target is not carried away, and the flight path does not repeat its initial start. Figure 6 As shown, the scene is an indoor environment, with two moving targets performing lateral intersecting movements. The result image shows that when the two targets intersect, their trajectories are complete, and there are no issues with ID swapping or target mismatch due to occlusion. Figure 7 As shown in the figure, the scenario is an indoor environment, with two moving targets undergoing longitudinal crisscrossing motion. The results demonstrate that when the two targets intersect, the algorithm can accurately distinguish and continuously track each target, avoiding trajectories from intersecting or becoming confused. These results prove the stability and reliability of this invention in tracking multiple targets under different motion modes.

[0073] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A multi-target tracking method for millimeter-wave radar based on dynamic and static point cloud separation and adaptive methods, characterized in that: Includes the following steps: The radio frequency signals received by the radar are processed to generate dynamic and static point cloud data of the millimeter-wave radar. By setting distance and energy thresholds, invalid point clouds are pre-filtered to remove dynamic and static interference points that are outside the radar's range or have too low reflected energy. The preprocessed static and dynamic point cloud data are matched with the existing list of tracking targets. The matching process is based on the dual criteria of spatial distance and velocity consistency. For isolated point clouds that fail to match any existing targets, initiate a target creation determination mechanism; Once the target is successfully established and enters the active state, the tracking process is initiated.

2. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The steps for generating dynamic and static point cloud data include: Perform a range-dimensional Fast Fourier Transform on the signal to calculate the target distance information; The signal is divided into a dynamic point cloud generation branch and a static point cloud generation branch. In the dynamic point cloud branch, after filtering out static targets through the moving target display MTI, a velocity dimension FFT is performed to conduct two-dimensional CFAR detection. In the static point cloud branch, the velocity dimension FFT is directly performed to perform two-dimensional CFAR detection on the zero-frequency region; The angle of arrival of the target is estimated by phase detection to obtain angle information.

3. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, In the pre-filtering step, the Euclidean distance, angle, and energy between the measuring point and the starting origin are calculated to determine whether the point is within the set range of distance, angle, and energy, and invalid point clouds that are not within the range are filtered out.

4. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, In the matching step, the Euclidean distance and velocity difference between the point cloud and the estimated position of the target are calculated. If the distance is less than a preset spatial threshold and the velocity difference is within the allowable range, the matching is considered successful.

5. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The target creation determination mechanism includes: Determine whether the point cloud meets the false target suppression conditions: the velocity is greater than the minimum velocity threshold, the number of matched point clouds is less than the minimum point cloud requirement, and the distance is greater than the distance threshold. If the conditions are met, skip the new process; If the conditions are not met, a new target is allocated in an idle tracking slot, and the target parameters are initialized. The activation cycle is dynamically set based on the target distance and the number of matching point clouds.

6. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The trajectory prediction step in the continuous tracking process includes: Determine the target's movement pattern based on its historical state; For stationary targets, the current state is used as the predicted value; An acceleration compensation model is used for targets that cross the target. A uniform acceleration model is used for conventional moving targets.

7. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The point cloud matching step includes: Perform intersection scene prediction and pre-screen based on the consistency of motion between point cloud and target; The matching threshold is dynamically adjusted based on the target state and environment. Adaptive matching range expansion is implemented for distant or boundary targets; The target with the smallest distance is selected as the matching result.

8. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The state update step employs an improved Alpha-Beta-Gamma filter, selecting different update strategies based on the target state and the number of matching point clouds.

9. The millimeter-wave radar multi-target tracking method based on dynamic and static point cloud separation and adaptation as described in claim 1, characterized in that, The target disappearance determination step includes: Based on the target speed, it can be classified as static, slightly moving, or in motion. Combined with a hysteresis mechanism to prevent state jumps; Set the release counter according to the differences in target status and location; Reduce release delay for close-range backward movement scenarios.

10. A multi-target tracking system for millimeter-wave radar based on dynamic and static point cloud separation and adaptation, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 9.

Images