Multi-level label management method for dense multi-target tracking scene

Abstract

The invention belongs to the field of radar signal processing and multi-target tracking, and relates to a multi-level label management method for a dense multi-target tracking scene, which comprises the steps of target detection data preprocessing, target tracker construction, dense target layering strategy and label layering step-by-step updating. According to the method, the problems of tracking association errors and insufficient tracking precision caused by intersection of dense multi-target tracks and target shielding are solved, a filtering updating parameter set is constructed by combining micro-motion features and target state estimation, and hierarchical step-by-step updating is carried out to improve the robustness of dense multi-target tracking.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing and multi-target tracking, and relates to a multi-level label management method for dense multi-target tracking scenarios. More specifically, it is a method based on radar signal feature extraction and radar detection information fusion, and constructing a multi-level label management method for tracking dense multi-targets. Background Art

[0002] Dense target tracking technology is an important branch of multi-target tracking and is widely used in security, transportation and other fields. The tracking needs for ballistic missile targets and drone groups are particularly significant. These two types of targets have the significant characteristics of high-speed flight and multi-target coordination, which put forward strict requirements on target recognition, data association and trajectory prediction.

[0003] Ballistic target tracking is mainly used in air defense early warning and missile defense systems. Its core task is to capture and predict the real-time trajectory of high-speed moving, multi-warhead targets. Ballistic targets fly at extremely high speeds, with complex trajectories and nonlinear characteristics. At the same time, such targets are usually accompanied by decoy releases and jamming signals, which makes it easy for the real target signal in the sensor data to be covered or tampered with, increasing the complexity of the tracking task. Drone swarm target tracking is widely used in scenarios such as logistics distribution and airspace management. In these scenarios, the targets are highly dense, similar in appearance, and have variable motion trajectories, which makes traditional multi-target tracking algorithms face difficulties such as occlusion processing, target association, and identity maintenance. In addition, the environmental adaptability requirements of drones also put forward higher requirements on the real-time and robustness of tracking algorithms.

[0004] In response to the above possible problems, existing multi-target tracking algorithms, such as Joint Probabilistic Data Association (JPDA), consider the correlation between observation points and targets and the influence between observation points, effectively dealing with target occlusion in dense target scenes; Multiple Hypothesis Tracking (MHT) methods effectively deal with interference and target occlusion by maintaining multiple hypotheses. The method of joint data association by maintaining multiple hypothesis probabilities has high computational complexity in dense multi-target scenes, and the cost of maintaining target tracks and evaluation requires a lot of storage.

[0005] In this work, dense target tracking requires comprehensive consideration of target density, environmental complexity and real-time processing performance. A multi-level label management method is established by combining multi-hypothesis data association and multi-hypothesis event methods to evaluate the trade-offs in trajectory prediction, target recognition and interference suppression to meet the actual needs of complex scenarios. Summary of the invention

[0006] In order to overcome the high difficulty of target separation and association in the prior art, the defect of easy association errors caused by feature confusion in dense environments, and to achieve continuous and stable tracking of dense multi-targets in complex scenarios, the present invention proposes a dense multi-target tracking method based on radar target detection and micro-motion feature fusion, including a filtering update technical solution based on the DeepSORT mechanism, and combines micro-motion feature extraction, target state modeling and label hierarchical step-by-step update strategy to optimize the state prediction and existence probability estimation of the target. On this basis, the micro-motion signal is further separated and parameter estimated by sinusoidal frequency modulated Fourier-Bessel transform (SFMFBT), and a filtering update parameter set is established for dense group targets.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A multi-level label management method for dense multi-target tracking scenarios includes the following parts:

[0009] (I) Construction of filter update parameter set

[0010] In radar-based multi-target tracking tasks, in order to deal with the target confusion problem caused by insufficient radar resolution in dense target scenes, the sinusoidal frequency modulated Fourier-Bessel transform (SFMFBT) technology is introduced to enhance the ability to distinguish dense targets. Through SFMFBT, the phase of the discrete signal is decomposed into Bessel function bases of different orders, and combined with the kernel function resolution parameter k, the target signal frequency is separated and estimated.

[0011] The SFMFBT and ISFMFBT of a discrete sinusoidal frequency modulated signal s[n] of length N are defined as follows:

[0012]

[0013] Among them, S α [v] is the Fourier-Bessel transform of the sinusoidal frequency modulated SFM signal s[n], n is the number of terms representing the SFM signal s[n], J α (*) is the α-order Bessel function, v is the SFMFB domain transformation, N is the length of the signal s[n], k is the kernel function resolution, λ v represents the positive root of the Bessel function, f s Represents the sampling rate of the discrete signal, and j represents a pure imaginary number.

[0014] SFMFBT exhibits excellent frequency resolution capability in the signal separation of micro-motion targets, especially under high resolution parameter k, it can effectively distinguish the aliased signals of multiple targets and further extract key micro-motion features.

[0015] The observation vector z obtained by transforming the point trace data of the i-th target at time t under radar observation i,t and the state vector x i,k It is expressed as:

[0016]

[0017] Where: x, y, z represent the three-dimensional coordinates of space, x′ t Represents x t The first derivative of x″ t Represents x t The second derivative of y t ,z t Similarly, R represents the radial distance of the target relative to the seeker radar, V represents the radial velocity of the target relative to the seeker radar, α represents the estimated radial acceleration of the target, θ represents the azimuth angle of the target in the radar coordinate system, and β represents the pitch angle of the target in the radar coordinate system.

[0018] In order to improve the filtering accuracy, the target state information such as distance, speed, angle and so on are combined with the micro-motion characteristics through state vector definition, motion modeling and noise modeling, and a filter update parameter set is established to optimize the robustness of the filtering process. Through the above model, the state of the target can be estimated and predicted more accurately, thereby improving the stability of multi-target tracking.

[0019] 2. Dense Target Layering Strategy

[0020] In order to achieve accurate tracking and management of dense targets, the present invention proposes a technical solution based on a dense target hierarchical strategy. This strategy introduces a hierarchical step-by-step update mechanism, a Murty association algorithm, and dynamic prediction and update of target existence probability, combined with the filtering update mechanism of DeepSORT technology, to perform multi-level correction and optimization of target states in complex dense scenes. The Murty algorithm is used to perform data association between the tracked target and the measurement, and k optimal or suboptimal matching hypotheses are generated to ensure the rationality of target matching in complex scenes.

[0021] For each hypothetical event, the system calculates its posterior probability P(θ h |Z) This posterior probability represents the hypothetical event θ given the measurement h The calculation of the posterior probability combines the prior probability of the hypothesized event with the likelihood function and is obtained through normalization.

[0022] The prediction of the target existence probability is based on the state estimation at the previous moment, and the target existence probability at time t is predicted. i pred (t) The formula is as follows:

[0023] r i pred (t) = p 11 ·r i (t-1)+p 12 ·(1-r i (t-1)) (5)

[0024] Among them, p 11 and p 12 are the probability of the target remaining and transitioning from non-existence to existence, r i (t-1) is the probability of the target existing at time t-1. This step predicts the possibility of the target existing at the current moment.

[0025] The best match usually has a higher posterior probability, while the suboptimal match has a relatively lower posterior probability. Through the weighted summation method, the system can combine the results of the best and suboptimal matches to update the probability of the target's existence.

[0026] The probability of a suboptimal match event r for a target that exists but no measurement is generated i 0 (t):

[0027]

[0028] The probability of the best matching event r for the target to exist and be associated with the measurement i l (t):

[0029]

[0030] The target existence probability update calculation formula is:

[0031]

[0032] Among them, P d and P w is the detection probability and threshold probability, ζ(i,j) indicates that the suboptimal measurement j comes from the hypothetical event of target i, ζ(i,l) indicates that the optimal measurement l comes from the hypothetical event of target i, represents the set of measurements within the tracking gate of target i; θ h is a hypothetical event, P(θ h ) is the probability of the hypothetical event.

[0033] Based on the matching hypothesis generated by Murty, the target existence probability is dynamically predicted and updated by combining the optimal matching and suboptimal matching results, thereby performing label hierarchical management based on the target existence probability and tracker status.

[0034] (III) Label layering and step-by-step updating

[0035] In multi-target tracking (MOT), the label hierarchical update strategy aims to improve the accuracy and stability of target labels, especially in scenes with dense targets, occlusions and complex motions, which can effectively reduce the risk of label errors and losses. According to the dense target hierarchical strategy, the filter states of targets at different levels are updated to achieve multi-level label management of the current frame. The core idea of ​​the label hierarchical update is to update the label information of the target layer by layer, through multi-level estimation optimization and update, hierarchical optimization, dynamic update and gradual convergence, to ensure that the correctness and stability of the label can be effectively maintained in complex scenes. Combined with the dynamic adjustment of the existence probability, this scheme significantly improves the accuracy and robustness of the multi-target tracking system in complex scenes.

[0036] Beneficial effects of the present invention:

[0037] This method solves the problems of tracking association errors and insufficient tracking accuracy caused by the intersection of dense multi-target tracks and target occlusion, and constructs a filter update parameter set by combining micro-motion features with target state estimation, and performs hierarchical and step-by-step updates to improve the robustness of dense multi-target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the overall structure of a radar multi-target tracking algorithm provided by an embodiment of the present invention;

[0039] Figure 2 A multi-level tag management flow chart provided by an embodiment of the present invention;

[0040] Figure 3 The algorithm flowchart is updated step by step for stratification.

[0041] Figure 4 Framework diagram of dense target tracking algorithm with multi-level label management. DETAILED DESCRIPTION

[0042] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0043] The present invention provides a multi-level label management method for dense multi-target tracking scenarios. First, the input radar dense target detection data is aligned with the corresponding extracted micro-motion features, and the consistency and confidence of the data are confirmed, and the false alarm rate and detection rate are extracted from the detection part. Then, the data of each frame is input into the dense multi-target tracker in time series to ensure efficient and accurate tracking of the targets at different time points. The multi-target tracking algorithm in the multi-target tracker is as follows: Figure 1 As shown in the figure, the relationship between the multi-level label management method and the multi-target tracking algorithm is as follows Figure 4As shown in the figure, step 1 corresponds to the input data part, which represents the preprocessing of the target detection data; step 2 is the construction of the target tracker, including the construction of the filter update parameter set, which is then added to the life cycle management and passed to the subsequent data association; step 3 is the target stratification strategy and step 4 is the label stratification step-by-step update part, which is incorporated into the label stratification management; step 5 corresponds to the output data part, including the output of the data and the feedback of historical information.

[0044] like Figure 2 As shown, the specific steps are as follows:

[0045] Step 1: Object detection data preprocessing

[0046] Perform point trace preprocessing and calibration on the input radar dense target detection data. Preprocess the extracted M frames of detection data, each frame contains N m Detection data, a total of M×N m Each detection data corresponds to a point trace, and each point trace is composed of four parameters: radial distance, radial velocity, azimuth angle and pitch angle. Through the sinusoidal frequency modulated Fourier-Bessel transform (SFMFBT), the phase of the discrete signal is decomposed into Bessel function bases of different orders, combined with the kernel function resolution parameter k, to achieve the separation and estimation of the target signal frequency and extract the key micro-motion features;

[0047] The SFMFBT and ISFMFBT of a discrete sinusoidal frequency modulated (SFM) signal s[n] of length N are defined as follows:

[0048]

[0049] Among them, S α [v] is the Fourier-Bessel transform of the sinusoidal frequency modulated SFM signal s[n], n is the number of terms representing the SFM signal s[n], J α (*) is the α-order Bessel function, v is the SFMFB domain transformation, N is the length of the signal s[n], k is the kernel function resolution, λ v represents the positive root of the Bessel function, f s Represents the sampling rate of the discrete signal, and j represents a pure imaginary number.

[0050] The steps of micro-motion feature extraction are as follows:

[0051] Step a: Initialization: Set the kernel function resolution k of SFMFBT, the signal energy threshold μ, and calculate the maximum number of valid items λ1 is the positive root of the first Bessel function, and the frequency resolution is Δf k =f s / 2πkN;

[0052] Step b: Calculate the positive root of the α-order Bessel function of the current length N signal, and use the positive root of the Bessel function to calculate the SFMFBT coefficient L v =SFMFBT[s[n]]. Find the maximum entry v in the SFMFB domain of the current loop. max = argmax v {|Re(S α [v])|}, Re represents the real part, calculate the corresponding frequency

[0053] Step c: L v Take the SFMFBT coefficient interval with a length of 2w as the center Calculating Energy Ratio If the ratio is greater than the threshold, go to step d; if the ratio is less than the threshold, stop the iteration;

[0054] Step d: The center is discretized into Calculate the center error frequency f ce , that is, the mean of the maximum error and minimum error of each frequency in the interval and the corrected frequency The corrected frequency is obtained as the micro-motion feature;

[0055] Step e: Let \ indicates from L v Culling calculate s i [n] Repeat steps a to e for the target echo signal reconstructed in this round of iterations.

[0056] By registering the radar target detection data with the micro-motion feature data, the consistency of the data in time and space is ensured. Then, by comparing the output data of radar echo signal processing and radar dense target detection, the existence of the target and the accuracy of positioning are confirmed, and the false alarm rate and detection rate are extracted to ensure the reliability of the data. Thus, the detection rate of the target is determined, and the registration result is passed to the subsequent target tracking. The registration result includes the micro-motion feature data of the point trace data.

[0057] Step 2: Construction of the object tracker

[0058] For the point data in the registration result obtained in step 1, the state vector is constructed through the constant acceleration model (CA), the point data of the 0th frame is transferred through the constant acceleration model, and the radial distance, radial velocity, azimuth, and pitch angle in the point data are converted into a three-dimensional constant acceleration model state vector, which constitutes a filtering update parameter set with the micro-motion feature; the details are as follows:

[0059] The observation vector z obtained by transforming the point trace data of the i-th target at time t under radar observationi,t and the state vector x i,k It is expressed as:

[0060]

[0061] Where: x, y, z represent the three-dimensional coordinates of space, x′ t Represents x t The first derivative of x″ t Represents x t The second derivative of y t ,z t Similarly, R represents the radial distance of the target relative to the seeker radar, V represents the radial velocity of the target relative to the seeker radar, α represents the estimated radial acceleration of the target, θ represents the azimuth angle of the target in the radar coordinate system, and β represents the pitch angle of the target in the radar coordinate system.

[0062] Create a target tracker for the initial state vector. Based on the state estimation at the previous moment, each target tracker initializes the state of its target through Kalman filtering, generates a set of filter update parameter sets, and establishes an identification matrix based on the filter update parameter sets of the target tracker to prepare for the next step of target stratification.

[0063] Step 3: Dense Target Layering Strategy

[0064] First, for each target tracker in step 2, the target state vector x i,t Perform state estimation prediction and target existence probability prediction; the process of target existence probability prediction is as follows:

[0065] The prediction of the target existence probability is based on the state estimation at the previous moment, and the target existence probability at time t is predicted. i pred (t) The formula is as follows:

[0066] r i pred (t) = p 11 ·r i (t-1)+p 12 ·(1-r i (t-1)) (5)

[0067] Among them, p 11 and p 12 are the probability of the target remaining and transitioning from non-existence to existence, r i (t-1) is the probability of the target existing at time t-1. This step predicts the possibility of the target existing at the current moment.

[0068] Then, the Murty association algorithm generates n sets of optimal or suboptimal matching results with the lowest cost on the given cost matrix to form hypothetical events. Each hypothetical event corresponds to a possible matching relationship between the target and the measurement. For each hypothetical event, the target tracker calculates its posterior probability P(θ h |Z) This posterior probability represents the hypothetical event θ given the measurement h The probability of occurrence. The optimal match usually has a higher posterior probability, while the posterior probability of the suboptimal match is relatively low.

[0069] Finally, the event probability r is calculated by the optimal matching corresponding to multiple hypothetical events i l (t) and several suboptimal matches to calculate the event probability r i 0 (t), the target tracker can combine the results of the optimal and suboptimal matching to update the probability of the target existence. The details are as follows:

[0070] The probability of a suboptimal match event r for a target that exists but no measurement is generated i 0 (t):

[0071]

[0072] The probability of the best matching event r for the target to exist and be associated with the measurement i l (t):

[0073]

[0074] The target existence probability update calculation formula is:

[0075]

[0076] Among them, P d and P w is the detection probability and threshold probability, ζ(i,j) indicates that the suboptimal measurement j comes from the hypothetical event of target i, ζ(i,l) indicates that the optimal measurement l comes from the hypothetical event of target i, represents the set of measurements within the tracking gate of target i; θ h is a hypothetical event, P(θ h ) is the probability of the hypothetical event.

[0077] For each target tracker in the target tracking system, the target tracking system calculates the target existence probability to ensure whether the target state is valid at the current moment. The calculation of the target existence probability will be dynamically adjusted along with the prediction and update steps of each tracker.

[0078] When the new observation vector z is obtainedi,k After that, the target existence probability will be updated through the matching results and the posterior probability. If the confidence of the target is high, the target existence probability will increase, otherwise it will decrease. The updated target existence probability helps the system to accurately identify valid targets in multi-target tracking and avoid misjudgment and missed detection.

[0079] Combined with the DeepSORT filtering mechanism, the prediction update and the distinction between micro-motion features are used. In the Murty algorithm to generate hypothesis events, for high-confidence (Confirmed) targets, when the cost is less than 1.2 times the matching threshold, the target to which the tracker belongs is successfully associated with the observation point trace; and for low-confidence (Tentative) targets, when the cost is less than 0.8 times the matching threshold, the target to which the tracker belongs is successfully associated with the observation point trace. Based on Murty multi-hypothesis events, the probability of target existence is dynamically adjusted, and the probability of existence of a certain threshold is used as the basis for distinguishing the state of the target tracker, and the number of consecutive associated frames and the number of skipped frames of the target tracker are used as multi-frame representations. When the target tracker meets the state conditions, the target tracker of each target is divided into high-confidence confirmed state (Confirmed), low-confidence transient state (Tentative), and death state (death) for target stratification. The state of the tracker depends on historical information and feedback. After multiple iterative updates, the state inspection and transformation of the target are gradually completed to ensure its accuracy.

[0080] Step 4: Update the label layer by layer

[0081] First, initialize the set of identification matrices obtained in step 2, and transfer the parameter set to be updated into the next level hierarchical step-by-step update strategy. Combined with the target state estimation and target existence probability update mechanism in step 3, the hierarchical step-by-step update strategy is used to carefully optimize the target. Figure 3 As shown in the figure, in this step, the targets are divided into different levels by calculating the probability of target existence according to factors such as their confidence and trajectory stability, and then the deviation between the track and the current update is calculated for continuous screening, and the identification matrix and the filter parameter set are gradually updated in a step-by-step updating manner.

[0082] The target tracker updates its state estimate: The target tracker updates the state estimate in a hierarchical manner, using the observation vector z corresponding to the best matching event obtained in step 3 i,t Perform filter update to obtain an updated identification matrix and filter parameter set.

[0083] Dynamic update of target tracker labels: The target tracker labels will change dynamically according to the state update and probability adjustment of each level. For high-confidence targets, the label stability is higher; for low-confidence targets, the label may be marked as pending and corrected through multiple updates.

[0084] Repeat steps 2-4, and continuously predict and update these target parameters through the filter to ensure real-time correction of the target state. After each frame of data is connected, the filter updates the filter parameters based on the Kalman gain and the current measurement data to optimize the state estimation and tracking accuracy of the target.

[0085] Step 5: Output and Feedback

[0086] Finally, the tracking results will be fed back to the system control module in real time for further target behavior analysis, path planning or other processing tasks. At each moment, the target tracking system will output the target's position, speed, micro-motion characteristics and other state parameters, as well as the target's existence probability and label information. The output data can be used for further trajectory analysis, prediction and path planning. The accuracy of the tracking results is fed back to the system so that the update strategy can be adjusted according to the actual situation to optimize the target tracking effect.

Claims

1. A multi-level tag management method for dense multi-target tracking scenarios, characterized in that: Here are the steps: Step 1: Object detection data preprocessing Perform point trace preprocessing and calibration on the input radar dense target detection data; preprocess the extracted M frames of detection data, each frame contains N m Detection data, a total of M×N m detection data, each detection data corresponds to a point trace, and each point trace is composed of four parameters: radial distance, radial velocity, azimuth angle and pitch angle; through the sinusoidal frequency modulated Fourier-Bessel transform SFMFBT, the phase of the discrete signal is decomposed into Bessel function basis of different orders, combined with the kernel function resolution parameter k, the separation and estimation of the target signal frequency is realized, and the key micro-motion features are extracted; By registering radar target detection data with micro-motion feature data, the consistency of data in time and space is ensured; then, by comparing the output data of radar echo signal processing and radar dense target detection, the existence of the target and the accuracy of positioning are confirmed, and the false alarm rate and detection rate are extracted to ensure the reliability of the data; Thus, the detection rate of the target is determined, and the registration result is passed to the subsequent target tracking. The registration result includes the micro-motion feature data of the point trace data; Step 2: Construction of the object tracker For the point data in the registration result obtained in step 1, the state vector is constructed through the constant acceleration model CA, the point data of the 0th frame is transferred through the constant acceleration model, and the radial distance, radial velocity, azimuth, and pitch angle in the point data are converted into a three-dimensional constant acceleration model state vector, which constitutes a filtering update parameter set with the micro-motion feature; the details are as follows: The observation vector z obtained by transforming the point trace data of the i-th target at time t under radar observation i,t and the state vector x i,k It is expressed as: Where: x, y, z represent the three-dimensional coordinates of space, x' t Represents x t The first derivative of x” t Represents x t The second derivative of y t ,z t Similarly; R represents the radial distance of the target relative to the seeker radar, V represents the radial velocity of the target relative to the seeker radar, α represents the estimated radial acceleration of the target; θ represents the azimuth angle of the target in the radar coordinate system, and β represents the pitch angle of the target in the radar coordinate system; Create a target tracker for the initial state vector. Based on the state estimation at the previous moment, each target tracker initializes the state of its target through Kalman filtering, generates a set of filter update parameter sets, and establishes an identification matrix based on the filter update parameter sets of the target tracker. Step 3: Dense Target Layering Strategy First, for each target tracker in step 2, the target state vector x i,t Perform state estimation prediction and target existence probability prediction; the process of target existence probability prediction is as follows: The prediction of the target existence probability is based on the state estimation at the previous moment, and the target existence probability at time t is predicted; the target existence probability prediction The formula is as follows: Among them, p 11 and p 12 are the probability of the target remaining and transitioning from non-existence to existence, r i (t-1) is the probability of the target existing at time t-1; this step predicts the possibility of the target existing at the current moment; Then, the Murty association algorithm generates n groups of optimal or suboptimal matching results with the lowest cost on the given cost matrix to form hypothetical events. Each hypothetical event corresponds to a possible matching relationship between the target and the measurement. For each hypothetical event, the target tracker calculates its posterior probability P(θ h |Z) This posterior probability represents the hypothetical event θ given the measurement h Likelihood of occurrence; Finally, the event probability is calculated by the optimal matching corresponding to multiple hypothetical events Calculate event probability with several suboptimal matches The target tracker can combine the results of the optimal and suboptimal matches to update the target existence probability; specifically: The probability of a suboptimal match event r for a target that exists but no measurement is generated i 0 (t): The probability of the best matching event r for the target to exist and be associated with the measurement i l (t): The target existence probability update calculation formula is: Among them, P d and P w is the detection probability and threshold probability, ζ(i,j) indicates that the suboptimal measurement j comes from the hypothetical event of target i, ζ(i,l) indicates that the optimal measurement l comes from the hypothetical event of target i, represents the set of measurements within the tracking gate of target i; θ h is a hypothetical event, P(θ h ) is the probability of the hypothetical event; For each target tracker in the target tracking system, the target tracking system calculates the target existence probability to ensure whether the target state is valid at the current moment; the calculation of the target existence probability will be dynamically adjusted along with the prediction and update steps of each tracker; When the new observation vector z is obtained i,k After that, the existence probability of the target will be updated through the matching results and the posterior probability; Combined with the DeepSORT filtering mechanism, the prediction update and micro-motion feature distinction are used. In the Murty algorithm to generate hypothesis events, for high-confidence targets, when the cost is less than 1.2 times the matching threshold, the target to which the tracker belongs is successfully associated with the observed point trace; and for low-confidence targets, when the cost is less than 0.8 times the matching threshold, the target to which the tracker belongs is successfully associated with the observed point trace; based on Murty multi-hypothesis events, the target existence probability is dynamically adjusted, and the existence probability of a certain threshold is used as the basis for distinguishing the state of the target tracker, and the number of consecutive associated frames and skipped frames of the target tracker are used as multi-frame representations; when the target tracker meets the state conditions, the target tracker of each target is divided into high-confidence confirmation state, low-confidence transient state, and death state for target stratification. The state of the tracker depends on historical information and feedback. After multiple iterative updates, the state inspection and transformation of the target are gradually completed to ensure its accuracy; Step 4: Update the label layer by layer First, initialize the set of identification matrices obtained in step 2, and transfer the parameter set to be updated into the next level of hierarchical step-by-step update strategy; combine the target state estimation and target existence probability update mechanism in step 3, and use the hierarchical step-by-step update strategy to optimize the target in detail; the details are as follows: The target tracker updates its state estimate: The target tracker updates the state estimate in a hierarchical manner, using the observation vector z corresponding to the best matching event obtained in step 3 i,t Perform filter update to obtain an updated identification matrix and filter parameter set; Dynamic update of target tracker labels: The target tracker labels will change dynamically according to the state update and probability adjustment of each layer; for high-confidence targets, the label stability is higher; for low-confidence targets, the label may be marked as pending and corrected through multiple updates; Repeat steps 2-4, and continuously predict and update these target parameters through the filter to ensure real-time correction of the target state. After each frame of data is received, the filter updates the filter parameters based on the Kalman gain and the current measurement data to optimize the state estimation and tracking accuracy of the target. Step 5: Output and Feedback Finally, the tracking results will be fed back to the system control module in real time for further target behavior analysis, path planning or other processing tasks; at each moment, the target tracking system will output the state parameters of the target's position, speed, micro-motion characteristics, as well as the target's existence probability and label information.

2. A multi-level tag management method for dense multi-target tracking scenarios according to claim 1, characterized in that: In step 1, The SFMFBT and ISFMFBT of a discrete sinusoidal frequency modulated SFM signal s[n] of length N are defined as follows: Among them, S α [v] is the Fourier-Bessel transform of the sinusoidal frequency modulated SFM signal s[n], n is the number of terms representing the SFM signal s[n], J α (*) is the α-order Bessel function, v is the SFMFB domain transformation, N is the length of the signal s[n], k is the kernel function resolution, λ v represents the positive root of the Bessel function, f s represents the sampling rate of discrete signals, j represents a pure imaginary number; The steps of micro-motion feature extraction are as follows: Step a: Initialization: Set the kernel function resolution k of SFMFBT, the signal energy threshold μ, and calculate the maximum number of valid items λ1 is the positive root of the first Bessel function, and the frequency resolution is Δf k =f s / 2πkN; Step b: Calculate the positive root of the α-order Bessel function of the current length N signal, and use the positive root of the Bessel function to calculate the SFMFBT coefficient L v =SFMFBT[s[n]]; find the maximum item v in the SFMFB domain of the current loop max = argmax v {|Re(S α [v])|}, Re represents the real part, calculate the corresponding frequency Step c: L v Take the SFMFBT coefficient interval L with a length of 2w as the center vmax,2w =[L vmax-w ,L vmax+w ], calculate the energy ratio ratio = ∑(L vmax,2w ) 2 / ∑(L v ) 2 ; If the ratio is greater than the threshold, go to step d; if the ratio is less than the threshold, stop the iteration; Step d: f vmax The center is discretized into Calculate the center error frequency f ce , that is, the mean of the maximum error and minimum error of each frequency in the interval and the corrected frequency The corrected frequency is obtained as the micro-motion feature; Step e: Let \ indicates from L v Culling calculate s i [n] Repeat steps a to e for the target echo signal reconstructed in this round of iterations.

Images