Multi-target dynamic marking method and system based on multi-dimensional trajectory fingerprint and intelligent arbitration

Abstract

The application discloses a multi-target dynamic marking method and system of multi-dimensional trajectory fingerprint and intelligent arbitration, relates to the technical field of radar signal processing and target identification, and comprises the following steps: acquiring radar original trajectory data and preprocessing to obtain a multi-dimensional motion feature vector; inputting the multi-dimensional motion feature vector into a local sensitive hash module, performing trajectory fingerprint matching by using preset composite hash functions in L independent hash tables, and obtaining an association ID; monitoring distances between all targets in real time, and when the distance between two targets is less than a trajectory conflict determination distance threshold, determining whether the trajectories are independent based on trajectory decoupling verification, if yes, maintaining the association ID, and if not, starting a conflict arbitration mechanism; and decoupling the conflict targets by adopting a progressive three-level arbitration strategy, so as to maintain the association ID; the method improves the distinguishing ability of similar trajectories, takes into account real-time performance and accuracy, and reduces the frequency of manual intervention.

Description

Multi-dimensional trajectory fingerprinting and intelligent arbitration multi-target dynamic marking method and system Technical Field

[0001] This invention relates to the field of radar signal processing and target recognition technology, and in particular to a multi-target dynamic marking method and system using multi-dimensional trajectory fingerprinting and intelligent arbitration. Background Technology

[0002] In fields such as drone management and intelligent security, it is often necessary to continuously track and uniquely identify multiple high-speed maneuvering targets (such as target drones and drone swarms) within the radar detection range. This task is known as "multi-target tracking and identification," and its core challenge lies in how to correctly distinguish and stably mark each individual target among a massive, dense, and interwoven array of radar points.

[0003] Current multi-target tracking and identification technologies mainly suffer from the following technical bottlenecks:

[0004] Trajectory intersection confusion problem: Traditional algorithms (such as K-means and nearest neighbor association) are not good at distinguishing multiple targets with highly similar or intersecting trajectories, which can easily lead to incorrect switching of target identity (ID).

[0005] Problems with the singularity of feature representation and environmental sensitivity: Methods that rely on static physical features (such as radar cross-section RCS) are susceptible to environmental factors (such as rain and fog) and changes in target attitude, resulting in low reliability.

[0006] Real-time processing performance bottleneck: Deep learning-based recognition models have high computational complexity and large processing latency when dealing with hundreds of targets, making it difficult to meet the real-time labeling requirements of highly maneuverable targets. Summary of the Invention

[0007] Based on the technical problems existing in the background technology, this invention proposes a multi-target dynamic marking method and system for multi-dimensional trajectory fingerprinting and intelligent arbitration, which improves the ability to distinguish similar trajectories, balances real-time performance and accuracy, and reduces the frequency of manual intervention.

[0008] The multi-target machine dynamic labeling method for multi-dimensional trajectory fingerprinting and intelligent arbitration proposed in this invention includes:

[0009] Acquire raw radar trajectory data and preprocess it to obtain multidimensional motion feature vectors;

[0010] The multidimensional motion feature vector is input into the local sensitive hash module. By querying the hash buckets determined by the preset composite hash function in L independent hash tables, a fast retrieval and matching is performed to assign or associate an associated ID to the current trajectory.

[0011] Real-time monitoring of the distance between all targets. When the distance between two targets is less than the distance threshold for trajectory conflict determination, the trajectory decoupling verification is used to determine whether the trajectories are independent. If they are independent, the associated ID is maintained; otherwise, the conflict arbitration mechanism is initiated.

[0012] A progressive three-tier arbitration strategy is adopted to adjudicate conflicting objectives, thereby maintaining or determining the correct association ID.

[0013] Furthermore, the acquisition of raw radar trajectory data and preprocessing to obtain a multi-dimensional motion feature vector specifically involves:

[0014] The original radar trajectory data under the current sliding window is obtained using a configurable sliding window, and the trajectory segments are formed by sorting them by time.

[0015] Feature extraction is performed on the trajectory segment to obtain a multidimensional feature vector;

[0016] The multidimensional feature vector is normalized using Z-score with pre-stored feature mean and standard deviation to obtain a multidimensional motion feature vector, which is used to characterize the motion pattern.

[0017] Furthermore, the process for generating the configurable duration sliding window is as follows:

[0018] The system maintains a first-in-first-out data cache in memory to store each point of the original radar trajectory data in chronological order. Each point contains at least the target ID, timestamp, and three-dimensional spatial coordinates (x, y, z), where x, y, and z are the horizontal direction data, vertical direction data, and height or depth data, respectively.

[0019] The system checks the data cache at fixed time intervals;

[0020] During each check, the latest point in the data cache is used as the time endpoint, and all stored points within the configured time period are selected in the historical direction to form a sliding window for processing.

[0021] Furthermore, the multidimensional feature vector includes: standard deviation of acceleration change rate, turning radius fluctuation, velocity entropy value, fluctuation of the angle between neighborhood normal vectors, root mean square of jerk, maximum angular velocity, centroid offset of spectrum, trajectory fractal dimension, peak value of autocorrelation coefficient, maneuver intensity index, motion persistence, trajectory asymmetry, average curvature, average torsion, motion mobility, and motion regularity.

[0022] Furthermore, the step of inputting the multidimensional motion feature vector into the locality-sensitive hash module, and using the composite hash function preset in L independent hash tables to perform trajectory fingerprint matching, thereby obtaining the associated ID, specifically involves:

[0023] The output of the composite hash function is used as the hash bucket number, and the composite hash function is obtained by randomly combining K hash functions.

[0024] Construct L independent hash tables, each using a different composite hash function;

[0025] Calculate the hash key of the multidimensional motion feature vector in each hash table, retrieve and merge all historical feature vectors in all corresponding hash buckets, and form a candidate set after deduplication.

[0026] Calculate the similarity between the multidimensional motion feature vector and each vector in the candidate set. If the highest similarity exceeds the fingerprint matching similarity threshold, the match is successful. Mark the trajectory corresponding to the current sliding window as the historical ID. Otherwise, assign a new ID to the trajectory corresponding to the current sliding window and update the local sensitive hash index.

[0027] Furthermore, the determination of whether trajectories are independent based on trajectory decoupling verification specifically involves:

[0028] Calculate the Hausdorff distance of the target's trajectory over a given time period. ;

[0029] like If the distance is greater than or equal to the Hausdorff distance independence threshold, the two targets are considered independent.

[0030] Otherwise, the conflict arbitration mechanism will be initiated.

[0031] Furthermore, the adoption of a progressive three-tier arbitration strategy to adjudicate conflicting objectives specifically includes:

[0032] The first level is physical feature arbitration: comparing the micro-Doppler spectra of two targets. If the spectral correlation coefficient is less than the preset spectral threshold, the arbitration is successful. The targets are distinguished based on the spectral differences and their associated IDs are maintained. Otherwise, the process proceeds to the second level.

[0033] The second level is behavior prediction arbitration: short-term trajectory prediction is performed using a partially observable Monte Carlo planning predictor. If the predicted distance between the two targets in the future is greater than the preset trajectory conflict judgment distance threshold, the arbitration is successful and the associated ID is maintained to wait for natural differentiation; otherwise, it proceeds to the third level.

[0034] The third level is manual arbitration: freezing the current tags of the two targets in conflict and issuing an alarm; receiving an operator's identity verification instruction based on the alarm to unfreeze and update the tags of the two targets.

[0035] A multi-target dynamic marking system for multi-dimensional trajectory fingerprinting and intelligent arbitration, including:

[0036] The trajectory fingerprint feature extraction module is used to acquire the original radar trajectory data and preprocess it to obtain a multi-dimensional motion feature vector;

[0037] The Locality Sensitive Hash (LSH) module is used to perform trajectory fingerprint matching on the multidimensional motion feature vector using pre-set composite hash functions in L independent hash tables, thereby obtaining the associated ID.

[0038] The multi-target trajectory decoupling module is used to monitor the distance between all targets in real time. When the distance between two targets is less than the trajectory conflict judgment distance threshold, the trajectory decoupling verification is used to determine whether the trajectories are independent. If they are independent, the association ID is maintained; otherwise, the conflict arbitration mechanism is initiated.

[0039] The conflict arbitration module employs a progressive three-level arbitration strategy to decouple conflicting targets, thereby maintaining the associated ID.

[0040] Furthermore, the trajectory fingerprint feature extraction module specifically comprises:

[0041] The original radar trajectory data under the current sliding window is obtained using a configurable sliding window, and the trajectory segments are formed by sorting them by time.

[0042] Feature extraction is performed on the trajectory segment to obtain a multidimensional feature vector;

[0043] The multidimensional feature vector is standardized using Z-score with pre-stored feature mean and standard deviation to obtain a multidimensional motion feature vector, which is used to characterize the motion pattern.

[0044] The multidimensional feature vectors include: standard deviation of acceleration change rate, turning radius fluctuation, velocity entropy value, fluctuation of the angle between neighborhood normal vectors, root mean square of jerk, maximum angular velocity, centroid offset of spectrum, trajectory fractal dimension, peak value of autocorrelation coefficient, maneuver intensity index, motion persistence, trajectory asymmetry, average curvature, average torsion, motion mobility, and motion regularity.

[0045] Furthermore, the locality-sensitive hash module includes:

[0046] The hash function and index building module uses K hash functions to randomly combine to obtain a composite hash function, and uses the output of the composite hash function as the hash bucket number.

[0047] The hash table building module is used to build L independent hash tables, each using a different composite hash function;

[0048] The candidate set construction module is used to calculate the hash key of the multidimensional motion feature vector in each hash table, retrieve and merge all historical feature vectors in all corresponding hash buckets, and form a candidate set after deduplication.

[0049] The matching and decision module is used to calculate the similarity between the multidimensional motion feature vector and each vector in the candidate set. If the highest similarity exceeds the threshold, it means that the match is successful and the corresponding historical ID is associated. Otherwise, a new ID is assigned and the local sensitive hash index is updated.

[0050] The advantages of the multi-dimensional trajectory fingerprinting and intelligent arbitration multi-target dynamic labeling method and system provided by this invention are as follows: It introduces multi-dimensional motion feature vectors, integrating kinematic differential, geometric, and statistical features to comprehensively represent behavior, thus improving the ability to distinguish similar trajectories. Furthermore, it utilizes Local Sensitive Hashing (LSH) to achieve rapid trajectory fingerprint matching, and combines a partially observable Monte Carlo Programming (POMCP) predictor with Hausdorff distance to verify trajectory independence, balancing real-time performance and accuracy. Finally, it sets up a progressive three-level arbitration strategy, from physical features to behavior prediction, and finally to the progressive three-level arbitration strategy process, effectively reducing the frequency of manual intervention while ensuring final accuracy. Attached Figure Description

[0051] Figure 1 is a schematic diagram of the process of the present invention;

[0052] Figure 2 is a schematic diagram of the progressive three-level arbitration strategy process. Detailed Implementation

[0053] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0054] As shown in Figures 1 and 2, the multi-target machine dynamic labeling method for multi-dimensional trajectory fingerprinting and intelligent arbitration proposed in this invention includes:

[0055] Step 1: Acquire raw radar trajectory data and preprocess it to obtain multidimensional motion feature vectors;

[0056] Step 2: Input the multidimensional motion feature vector into the local sensitive hash module. Using L independent hash tables, perform fast retrieval and matching by querying the hash buckets determined by the preset composite hash function in each hash table. Based on the matching result, assign or associate an associated ID to the current trajectory.

[0057] Step 3: Monitor the distance between all targets in real time. When the distance between two targets is less than the distance threshold for trajectory conflict determination, determine whether the trajectories are independent based on trajectory decoupling verification. If they are independent, maintain the association ID; otherwise, start the conflict arbitration mechanism.

[0058] Step 4: Adopt a progressive three-tier arbitration strategy to adjudicate the conflicting objectives, thereby maintaining or determining the correct association ID.

[0059] This embodiment introduces a multi-dimensional motion feature vector, integrating kinematic differential, geometric, and statistical features to comprehensively represent behavior, thereby improving the ability to distinguish similar trajectories. Furthermore, it utilizes Local Sensitive Hashing (LSH) to achieve rapid trajectory fingerprint matching, and combines a Partially Observable Monte Carlo Programming (POMCP) predictor with Hausdorff distance to verify trajectory independence, balancing real-time performance and accuracy. Finally, a progressive three-level arbitration strategy is implemented, from physical features to behavior prediction, and finally to the progressive three-level arbitration strategy process, effectively reducing the frequency of human intervention while ensuring final accuracy.

[0060] In this embodiment, the multi-target dynamic labeling method is implemented through the following multi-target dynamic labeling system, which includes:

[0061] The trajectory fingerprint feature extraction module is used to acquire the original radar trajectory data and preprocess it to obtain a multi-dimensional motion feature vector;

[0062] The Local Sensitive Hash module performs fast retrieval and matching by querying the hash buckets determined by the pre-defined composite hash function in L independent hash tables. Based on the matching results, it assigns or associates an associated ID (i.e., a unique identifier ID) to the current trajectory.

[0063] The multi-target trajectory decoupling module is used to monitor the distance between all targets in real time. When the distance between two targets is less than the trajectory conflict judgment distance threshold, the trajectory decoupling verification is used to determine whether the trajectories are independent. If they are independent, the association ID is maintained; otherwise, the conflict arbitration mechanism is initiated.

[0064] The conflict arbitration module employs a progressive three-tier arbitration strategy to adjudicate conflicting objectives, thereby maintaining or determining the correct association ID.

[0065] Understandably, the functions of the multi-target trajectory decoupling module and the conflict arbitration module are: 1) to check whether the currently maintained association ID is correct (whether there is confusion); 2) to initiate a correction mechanism (arbitration) when confusion occurs, with the ultimate goal of correctly maintaining the association ID.

[0066] In one embodiment, step one, acquiring the original radar trajectory data and preprocessing it to obtain a multi-dimensional motion feature vector, specifically involves the trajectory fingerprint feature extraction module:

[0067] The trajectory fingerprint feature extraction module transforms a piece of raw radar trajectory data into a 16-dimensional feature vector V that can characterize its unique motion pattern. It can be understood that the 16-dimensional feature vector V can be set with other dimensions of feature vectors according to actual needs. This embodiment uses the 16-dimensional feature vector V as an example for explanation.

[0068] The 16-dimensional feature vector V is extracted from trajectory data within a configurable sliding window, specifically including:

[0069] Standard deviation of the rate of change of acceleration ( ): Quantifying the stability of the target's maneuvering behavior;

[0070] Turning radius fluctuation ( ): Reflects the stability of target-shifting habits;

[0071] Velocity entropy value ( ): Characterizes the randomness of motion patterns;

[0072] Fluctuation of the angle between neighborhood normal vectors ( ): Quantify the local continuity and abrupt changes in the curvature of the trajectory;

[0073] Root mean square of accelerometer (RMS) ): This characterizes the intensity of the target's "rapid movement"; it should be noted that jerk refers to the rate of change of acceleration, also known as "rapidity".

[0074] Maximum angular velocity ( ): Reflects the target's ability to perform extreme maneuvers;

[0075] Spectral centroid shift ( ): Principal components that characterize motion from the frequency domain perspective;

[0076] Trajectory fractal dimension ( ): Measures the complexity of the trajectory path;

[0077] Peak value of autocorrelation coefficient ( ): Used to detect the periodicity of motion;

[0078] Mobility Intensity Index ( ): The overall maneuver intensity of the comprehensive quantitative target;

[0079] Exercise duration ( ): Assess the target's tendency to maintain its current direction of motion;

[0080] Trajectory asymmetry ( ): Identify the target's turning preferences;

[0081] Mean curvature ( ): Describes the overall curvature of the path;

[0082] Mean torsion ( ): Represents the degree of twisting of the trajectory in space;

[0083] Motion mobility ( ): Measures the linearity and efficiency of target motion;

[0084] Regularity of motion ( ): Assess the predictability of the target motion pattern.

[0085] The 16-dimensional feature vectors mentioned above are Z-score standardized using the pre-stored feature mean and standard deviation to eliminate the influence of dimensions, resulting in 16-dimensional motion feature vectors used to characterize motion patterns.

[0086] In this embodiment, the trajectory fingerprint feature extraction module aims to transform continuous raw radar trajectory data into trajectory segments that can be used for subsequent analysis, specifically:

[0087] (a1) Data caching;

[0088] The system maintains a first-in-first-out (FIFO) data buffer in memory to store each target point reported by the radar in chronological order. Each point point contains at least a target ID (initially unknown), a timestamp, and three-dimensional spatial coordinates (x, y, z).

[0089] It should be noted that,

[0090] Input layer ID: This refers to the target ID in the data buffer area, which is the radar's original, local, and potentially unstable observation identifier.

[0091] Output layer ID: This is the associated ID recorded in this embodiment. It is a stable trajectory identity identifier that is newly generated and continuously maintained through this embodiment.

[0092] This embodiment aims to replace and associate the target ID with the associated ID, thereby solving the problem of the target ID's instability in complex scenarios.

[0093] (a2) Sliding window generation;

[0094] The system checks the cache at fixed time intervals.

[0095] During each check, the latest data point in the data cache is used as the time endpoint, and all stored data points within the configured time period are traced back in the historical direction (i.e. forward). All data points within this time period are formed into a sliding window.

[0096] (a3) Trajectory segment extraction;

[0097] For each unique target within the data buffer (identified by a temporary ID or initial track), extract all points belonging to the current window from its point set, sort them by time, and form a trajectory segment. ,in , This represents the number of dots within the sliding window. The first A dot mark, For dots The three-dimensional coordinates For dots Timestamp.

[0098] The temporary ID is the identification result obtained by the method in this embodiment, which will eventually evolve into an associated ID. That is, the system extracts a set of target points from the data cache based on the target ID that comes with the point, forming a trajectory segment; then, the trajectory segment is processed to assign or match a temporary ID to it.

[0099] (a4) Trajectory fingerprint feature extraction;

[0100] Each trajectory segment The calculations are performed to generate the corresponding 16-dimensional feature vectors. The following is a detailed explanation of the calculation methods for some core features:

[0101] (b1) Calculate primary kinematic quantities;

[0102] Velocity sequence V: for each point in the trajectory ( The instantaneous velocity vector is calculated using the central difference method. :

[0103] ;

[0104] in, For time intervals, For dots The three-dimensional coordinates For dots The three-dimensional coordinates They are dots timestamp, For the first A dot mark.

[0105] Acceleration sequence A: Based on the velocity sequence V, the instantaneous acceleration vector is also calculated using the central difference method. .

[0106] Instantaneous velocity sequence s: Calculate the magnitude of each velocity vector, the first... The magnitude of the velocity vector of each point .

[0107] (b2) Calculate 16-dimensional features:

[0108] (c1) Standard deviation of the rate of change of acceleration ( ):

[0109] First, calculate the numerical difference of the acceleration sequence A to obtain the acceleration sequence. ;

[0110] Calculate sequence The standard deviation of is This value reflects the target's maneuverability.

[0111] (c2) Fluctuation of the angle between the neighborhood normal vectors ( ):

[0112] For each point in the trajectory segment Take the k points before and after it to form a local point set. For example, when k=2, the local point set formed is... ,in, The first A dot mark.

[0113] right Perform three-dimensional principal component analysis (PCA) to extract the eigenvector corresponding to the smallest eigenvalue, which is then used as the trace of that point. Approximate normal vector at point The normal vector is perpendicular to the local trajectory plane and can characterize the direction of curvature.

[0114] Calculate the sequence of angles between the normal vectors of consecutive points: ,in, This is represented as the dot product of vectors. They are dots The approximate normal vector at that point.

[0115] Calculate the included angle sequence The standard deviation of is The standard deviation quantifies the degree of fluctuation in the direction of trajectory curvature.

[0116] (c3) Velocity entropy value ( );

[0117] Divide the range of values ​​of the instantaneous rate sequence s into M intervals (e.g., M=10).

[0118] Statistically calculate the frequency of rate values ​​falling within each interval and then calculate the probability distribution. , This indicates that, within the entire instantaneous rate sequence s, its rate value falls on the th . The statistical frequency of each interval For the entire instantaneous rate sequence s, its rate value falls within the th... Statistical frequency of each interval;

[0119] According to this probability distribution Calculate the Shannon entropy to obtain the velocity entropy value: The velocity entropy value characterizes the randomness of velocity changes.

[0120] (c4) Trajectory fractal dimension ( );

[0121] Using box counting, with a side length of A cubic mesh covers the entire trajectory segment. Count the number of cubes containing at least one trajectory point, denoted as . In this context, a trajectory point is a point within a trajectory segment, with one point corresponding to one trajectory point.

[0122] Gradually decrease A series of data points were obtained. ,in For the first The side length, for The corresponding number of cubes.

[0123] Trajectory fractal dimension Defined by the following relation:

[0124] .

[0125] Therefore, in a double logarithmic coordinate system, with As the independent variable, Using the data points as the dependent variable, a linear fit is performed, and the slope of the resulting line is the fractal dimension. The estimated value.

[0126] (c5) In the 16-dimensional feature vector of this embodiment, the remaining features are all obtained based on the aforementioned calculated sequence of basic kinematic quantities (position, velocity, and acceleration sequences) using mathematical and signal processing methods known in the art. Specifically:

[0127] Root mean square of jerk: Calculate the numerical difference of the acceleration sequence to obtain the jerk sequence, and then calculate the root mean square value of the jerk sequence.

[0128] Maximum angular velocity: Take the maximum value from the angular velocity sequence.

[0129] Peak value of autocorrelation coefficient: The peak value of the autocorrelation function at the main period is taken when calculating the autocorrelation function of the velocity sequence.

[0130] Mobility intensity index: defined as the integral mean of the magnitude of the acceleration vector within the sliding window.

[0131] Motion persistence: The trend of directional persistence is measured by linearly fitting the historical velocity direction sequence and using the fitting residual.

[0132] Trajectory asymmetry: Calculate the ratio of the cumulative curvature of the trajectory when turning left to the cumulative curvature when turning right.

[0133] Mean curvature and mean torsion: Based on the principles of differential geometry, the mean curvature and mean torsion of the trajectory segment are calculated respectively.

[0134] Motion mobility: defined as the ratio of net displacement to total path length.

[0135] Movement regularity: This is assessed by calculating the variance of the similarity between the current window features and its own historical window features.

[0136] Furthermore, turning radius fluctuation refers to the standard deviation of the instantaneous turning radius within the sliding window, where the instantaneous turning radius can be calculated based on vehicle dynamics or geometric models. Spectral centroid shift is the weighted average frequency of the signal spectrum, used to characterize the dominant frequency distribution of kinetic energy; its calculation method can be found in well-known techniques in the field of signal processing.

[0137] (a5) Feature normalization;

[0138] The calculated multidimensional feature vector (e.g., 16-dimensional feature vector) ), using pre-stored feature mean and standard deviation Z-score standardization is performed to obtain multidimensional motion feature vectors. Among them, the mean and standard deviation It comes from offline analysis of a large amount of historical trajectory data.

[0139] The calculation formula is as follows:

[0140] ;

[0141] in, For the dimension index of the multidimensional motion feature vector, It is the first 3D motion feature vector, For the first 3D feature vectors For example, when the total number of dimensions is 1, the total number of dimensions is 2. =16, It is a 16-dimensional feature vector.

[0142] In one embodiment, step two involves inputting the multidimensional motion feature vector into a locality-sensitive hashing module. This module performs rapid retrieval and matching by querying the hash buckets determined by a preset composite hash function in L independent hash tables, assigning or associating an associated ID to the current trajectory. Specifically, the locality-sensitive hashing module is:

[0143] The Locality Sensitive Hash (LSH) module employs a customized LSH technique to achieve rapid matching of trajectory fingerprints.

[0144] (d1) Hash functions and index construction;

[0145] A single hash function ,in, For random vectors, For random offset, For the width of the bucket, The multidimensional motion feature vector, in this embodiment, refers to the vector after feature normalization of the 16-dimensional feature vector.

[0146] In this embodiment, random vector A random projection direction is defined to project the multidimensional motion feature vector. Projecting onto this direction yields a scalar value. Random offset Translate this projection axis, then widen the bucket. The projection axis is divided into a series of equally wide intervals. The value of the hash function is... The index of the interval to which it belongs. Crucially, the vector... and The closer the Euclidean distances are in the original space, the greater their scalar values. and The closer the hashes are, the higher the probability that they will be mapped to the same hash bucket (i.e., the same interval). This is precisely the principle behind Locality Sensitive Hashing (LSH)'s ability to achieve fast, near-nearest neighbor retrieval. and These represent any two multidimensional motion feature vectors from the same feature space that are to be compared.

[0147] Combine K independently generated hash functions of this type into a composite hash function. The output of this composite hash function is the multidimensional motion feature vector. The hash bucket number. Wherein, , Let K be the Kth hash function.

[0148] Construct L independent hash tables, each using a different composite hash function.

[0149] (d2) Online matching process;

[0150] Generate candidate sets: for multidimensional motion feature vectors Calculate its hash key in L hash tables, collect all historical vectors in the corresponding hash buckets, merge and deduplicate them to form a candidate set C.

[0151] Fine matching: If the candidate set C is not empty, calculate Calculate the similarity (e.g., cosine similarity) with each vector in the candidate set C. Find the highest similarity max_sim and its corresponding historical ID.

[0152] Decision: If the highest similarity max_sim exceeds the fingerprint matching similarity threshold, the match is successful, and the trajectory of the current sliding window is marked with that historical ID; otherwise, the system assigns a completely new unique ID (i.e., the new ID) to the trajectory of the current sliding window, and inserts the new trajectory fingerprint and the new ID as new entries into the Local Sensitive Hash Index, thereby updating the Local Sensitive Hash Index. The historical ID and the new ID serve as the association ID.

[0153] In one embodiment, step three involves real-time monitoring of the distance between all targets. When the distance between two targets is less than the trajectory conflict determination distance threshold, the trajectory decoupling verification is used to determine whether the trajectories are independent. If they are independent, the associated ID is maintained; otherwise, a conflict arbitration mechanism is initiated. Specifically, the multi-target trajectory decoupling module is as follows:

[0154] In this embodiment, when the spatial distance between targets is too close, the multi-target trajectory decoupling module is responsible for verifying the trajectory independence:

[0155] Independent Monte Carlo Planning (POMCP) Predictor: Maintain an independent partially observable Monte Carlo planning predictor (i.e., POMCP predictor) for each target to predict its future movement path, thereby monitoring the distance between all targets in real time.

[0156] Trajectory independence verification: When the distance between two targets is less than the trajectory conflict determination distance threshold (distance_threshold), calculate the Hausdorff distance of their nearest trajectories. Its definition is:

[0157] ;

[0158] Where A and B are the sets of recent trajectory points of target A and target B, respectively. and It refers to the point trace in the set (i.e., the spatial coordinates of the target at a specific time). For dots with dots The Euclidean distance between them For dots to set The minimum distance between all points. For set All dots to set The maximum value among these minimum distances, another term It is a symmetrical term.

[0159] like If the distance is greater than or equal to the Hausdorff distance independence threshold (i.e., haussdorff_threshold), the two targets are considered independent; otherwise... If the distance is less than the Hausdorff distance independence threshold (i.e., haussdorff_threshold), the trajectories are considered to be highly similar, posing a risk of confusion, and the conflict arbitration mechanism is immediately initiated.

[0160] This embodiment uses Hausdorff distance to verify trajectory independence, which can more effectively capture the overall shape differences and maximum separation trends between trajectories and improve the discrimination accuracy.

[0161] In one embodiment, step four involves employing a progressive three-level arbitration strategy to adjudicate conflicting targets, thereby maintaining or determining the correct association ID; as shown in Figure 2, the conflict arbitration module specifically comprises:

[0162] The conflict arbitration module employs a progressive three-tier arbitration strategy to resolve conflicts between highly similar objectives, specifically:

[0163] The first level is physical feature arbitration (L1 arbitration): compare the micro-Doppler spectra of the two targets. If the spectral correlation coefficient is less than the preset spectral threshold (e.g., the spectral threshold is 0.8), the arbitration is successful. The targets are distinguished based on the spectral differences and their associated ID (i.e., unique identity) is maintained. Otherwise, proceed to the second level.

[0164] The second level is behavior prediction arbitration (L2 arbitration): short-term trajectory prediction is performed using a partially observable Monte Carlo Programming (POMCP) predictor. If the predicted distance between the two targets in the future is greater than the preset trajectory conflict judgment distance threshold, the arbitration is successful and the associated ID is maintained to wait for natural differentiation; otherwise, it proceeds to the third level.

[0165] The third level is manual arbitration (L3 arbitration): freezes the current tags of two targets in conflict and issues an alarm; receives an operator's identification confirmation instruction based on the alarm to unfreeze and update the tags of the two targets.

[0166] Example 1

[0167] The multi-target dynamic marking method for multi-dimensional trajectory fingerprinting and intelligent arbitration specifically includes steps S1 to S5.

[0168] S1: System initialization and parameter configuration;

[0169] When the system starts up, a series of operating parameters need to be configured according to the specific requirements of the application scenario (such as target density, mobility, radar accuracy, etc.). To clearly illustrate the specific implementation of this embodiment, a set of typical configuration parameters verified by simulation will be used as an example below. Those skilled in the art should understand that these parameters can be adjusted within a certain range without departing from the protection scope of this embodiment.

[0170] (e1) Example configuration of trajectory fingerprint parameters:

[0171] The sliding window's duration is configured as window_size = 1.0 s;

[0172] Window overlap ratio = 0.7;

[0173] Radar data sampling rate: sampling_rate = 100 Hz.

[0174] (e2) Example configuration of Local Sensitive Hash Index (LSH index) parameters:

[0175] Hash bucket width w = 4.0;

[0176] The number of each composite hash function is K = 10;

[0177] The number of hash tables is L = 20.

[0178] (e3) Example configuration of arbitration mechanism threshold;

[0179] The distance threshold for trajectory conflict determination is distance_threshold = 100.0 m;

[0180] The Hausdorff distance independence threshold is set to haussdorff_threshold = 15.0 m.

[0181] The fingerprint matching similarity threshold is 0.90.

[0182] Short-term trajectory prediction duration: prediction_window = 0.05 s;

[0183] S2: Data preprocessing and windowing;

[0184] The original radar trajectory data was acquired and preprocessed to obtain a 16-dimensional motion feature vector.

[0185] The sliding window is generated as follows:

[0186] The system checks the data buffer at fixed time intervals (e.g., every window_size×(1 - overlap_ratio) = 1.0×0.3 = 0.3 seconds).

[0187] During each check, the latest data point in the data cache is used as the time endpoint, and all stored data points within a configured duration window_size (e.g., 1.0 seconds) are selected in the historical direction to form a sliding window for processing. For example, at T=10.0 seconds, the generated sliding window will contain all data points from T=9.0 seconds to T=10.0 seconds.

[0188] The next window starts at T=9.3 seconds and ends at T=10.3 seconds, and so on, to achieve sliding window analysis and ensure the continuity of motion.

[0189] Finally, extract the trajectory fragment of the current sliding window. .

[0190] S3: Trajectory fingerprint feature extraction;

[0191] For each trajectory segment Calculations are performed to generate the corresponding 16-dimensional feature vector. This is then normalized to obtain the 16-dimensional motion feature vector. .

[0192] S4: Real-time fingerprint matching;

[0193] Locality Sensitive Hash (LSH) retrieval: Input the Locality Sensitive Hash (LSH) module. Using the pre-defined composite hash function in L hash tables, calculate the hash key for each table, and extract all historical fingerprint vectors from the corresponding hash bucket numbers to form a candidate set C.

[0194] If the candidate set C is not empty, then calculate The cosine similarity with each candidate vector in C. If the highest similarity max_sim ≥ similarity_threshold, the current window trajectory is marked with that historical ID. Otherwise, the system assigns a completely new unique ID to the current trajectory and inserts the new fingerprint and ID as a new entry into the LSH index (i.e., Local Sensitive Hash Index).

[0195] S5: Continuous tracking and trajectory decoupling verification;

[0196] Real-time monitoring of the distance between all targets; when the distance between two targets is less than the trajectory conflict determination distance_threshold, the Hausdorff distance of their nearest trajectories is calculated. .like If <haussdorff_threshold, then the conflict arbitration mechanism is activated.

[0197] S6: Conflict Arbitration;

[0198] L1 arbitration: Analyze the micro-Doppler spectrum of the conflicting targets. If there are significant differences in spectral characteristics (e.g., spectral correlation coefficient <0.8), the arbitration is successful.

[0199] L2 arbitration: Perform short-term trajectory prediction. If the prediction results show that the two will be clearly separated (e.g., the predicted distance between the two targets > distance_threshold), the arbitration is successful, and the ID is maintained to wait for natural differentiation.

[0200] L3 Arbitration: Freezes the status of conflict target markers, issues an alert, and awaits manual operator adjudication.

[0201] This embodiment has the following advantages:

[0202] 1. Full algorithm layer implementation: No special hardware is required; it only utilizes existing radar data, resulting in low cost and easy integration.

[0203] 2. High robustness and accuracy: Multi-dimensional dynamic features and multiple arbitration mechanisms can effectively cope with the challenges of trajectory intersection, environmental interference and high target maneuverability.

[0204] 3. Excellent real-time performance: Thanks to the compact design of the 16-dimensional feature vector, the fast fingerprint matching mechanism based on Locality Sensitive Hash (LSH), and the lightweight POMCP trajectory prediction, this system can still maintain a low labeling latency in high-concurrency target scenarios.

[0205] 4. Strong interpretability: The physical meaning of the features is clear, the arbitration logic is clear, and it is easy to debug and trust.

[0206] Based on the above description of the embodiments, those skilled in the art will understand that the multi-target dynamic marking method and system for multi-dimensional trajectory fingerprinting and intelligent arbitration described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0207] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-target machine dynamic marking method for multi-dimensional trajectory fingerprinting and intelligent arbitration, characterized in that, include: The system acquires raw radar trajectory data and preprocesses it to obtain a multidimensional motion feature vector. This multidimensional motion feature vector is then input into a local sensitive hash module. By querying the hash buckets determined by a pre-defined composite hash function in L independent hash tables, the system performs rapid retrieval and matching, assigning or associating an association ID to the current trajectory. The system also monitors the distance between all targets in real time. When the distance between two targets is less than the trajectory conflict determination distance threshold, the system determines whether the trajectories are independent based on trajectory decoupling verification. If they are independent, the association ID is maintained; otherwise, a conflict arbitration mechanism is initiated. A progressive three-tier arbitration strategy is adopted to adjudicate conflicting objectives, thereby maintaining or determining the correct association ID.

2. The method according to claim 1, characterized in that, The process of acquiring and preprocessing the raw radar trajectory data to obtain a multidimensional motion feature vector involves: acquiring the raw radar trajectory data under the current sliding window using a configurable duration sliding window, and sorting the data by time to form trajectory segments; and extracting features from the trajectory segments to obtain a multidimensional feature vector. The multidimensional feature vector is normalized using Z-score with pre-stored feature mean and standard deviation to obtain a multidimensional motion feature vector, which is used to characterize the motion pattern.

3. The method according to claim 2, characterized in that, The process of generating the configurable duration sliding window is as follows: The system maintains a first-in-first-out data buffer in memory to store each point of the original radar trajectory data in chronological order. Each point contains at least the target ID, timestamp, and three-dimensional spatial coordinates. The system checks the data buffer at fixed time intervals. During each check, the latest point in the data buffer is used as the time endpoint, and all stored points within the configured duration are selected in the historical direction to form a sliding window for processing.

4. The method according to claim 2, characterized in that, The multidimensional feature vectors include: standard deviation of acceleration change rate, turning radius fluctuation, velocity entropy value, fluctuation of the angle between neighborhood normal vectors, root mean square of jerk, maximum angular velocity, centroid offset of spectrum, trajectory fractal dimension, peak value of autocorrelation coefficient, maneuver intensity index, motion persistence, trajectory asymmetry, average curvature, average torsion, motion mobility, and motion regularity.

5. The method according to claim 1, characterized in that, The process involves inputting the multidimensional motion feature vector into a locality-sensitive hashing module, and using pre-defined composite hash functions in L independent hash tables to perform trajectory fingerprint matching to obtain the associated ID. Specifically, the output of the composite hash function is used as the hash bucket number, and the composite hash function is obtained by randomly combining K hash functions. L independent hash tables are constructed, with each hash table using a different composite hash function. The hash key of the multidimensional motion feature vector in each hash table is calculated, and all historical feature vectors in all corresponding hash buckets are retrieved and merged. After deduplication, a candidate set is formed. Calculate the similarity between the multidimensional motion feature vector and each vector in the candidate set. If the highest similarity exceeds the fingerprint matching similarity threshold, the match is successful. Mark the trajectory corresponding to the current sliding window as the historical ID. Otherwise, assign a new ID to the trajectory corresponding to the current sliding window and update the local sensitive hash index.

6. The method according to claim 1, characterized in that, The determination of whether trajectories are independent based on trajectory decoupling verification specifically involves: calculating the Hausdorff distance of the target's trajectory over a given time period. ； like If the distance is greater than or equal to the Hausdorff distance independence threshold, the two targets are determined to be independent; otherwise, the conflict arbitration mechanism is initiated.

7. The method according to claim 1, characterized in that, The proposed progressive three-tier arbitration strategy for adjudicating conflicting objectives specifically comprises: the first tier being physical characteristic arbitration; Compare the micro-Doppler spectra of the two targets. If the spectral correlation coefficient is less than the preset spectral threshold, the arbitration is successful. The targets are distinguished based on the spectral differences and their associated IDs are maintained. Otherwise, proceed to the second level. The second level is behavior prediction arbitration: short-term trajectory prediction is performed using a partially observable Monte Carlo planning predictor. If the predicted distance between the two targets in the future is greater than the preset trajectory conflict judgment distance threshold, the arbitration is successful and the associated ID is maintained to wait for natural differentiation; otherwise, it proceeds to the third level. The third level is manual arbitration: freezing the current markers of the two targets in conflict and issuing an alert; The operator receives the identity verification command provided by the alarm to unfreeze and update the tags of the two targets.

8. A multi-target dynamic marking system for multi-dimensional trajectory fingerprinting and intelligent arbitration, characterized in that, include: The trajectory fingerprint feature extraction module is used to acquire the original radar trajectory data and preprocess it to obtain a multi-dimensional motion feature vector; The Local Sensitive Hash (LSH) module is used to perform trajectory fingerprint matching on the multidimensional motion feature vector using pre-set composite hash functions in L independent hash tables, thereby obtaining the associated ID. The multi-target trajectory decoupling module is used to monitor the distance between all targets in real time. When the distance between two targets is less than the trajectory conflict determination distance threshold, the trajectory decoupling verification is used to determine whether the trajectories are independent. If they are independent, the associated ID is maintained; otherwise, the conflict arbitration mechanism is initiated. The conflict arbitration module employs a progressive three-level arbitration strategy to decouple conflicting targets, thereby maintaining the associated ID.

9. The system according to claim 8, characterized in that, The trajectory fingerprint feature extraction module specifically involves: acquiring the original radar trajectory data under the current sliding window using a configurable duration sliding window, and sorting it by time to form trajectory segments; and extracting features from the trajectory segments to obtain multi-dimensional feature vectors. The multidimensional feature vector is Z-score standardized using pre-stored feature mean and standard deviation to obtain a multidimensional motion feature vector used to characterize the motion pattern. The multidimensional feature vector includes: standard deviation of acceleration rate of change, turning radius fluctuation, velocity entropy value, fluctuation of neighborhood normal vector angle, root mean square of jerk, maximum angular velocity, spectral centroid offset, trajectory fractal dimension, peak value of autocorrelation coefficient, maneuver intensity index, motion persistence, trajectory asymmetry, average curvature, average torsion, motion mobility, and motion regularity.

10. The system according to claim 8, characterized in that, The Locality Sensitive Hash (LSH) module includes: a hash function and index construction module, which uses K hash functions to randomly combine to obtain a composite hash function, and uses the output of the composite hash function as the hash bucket number; a hash table construction module, which constructs L independent hash tables, each hash table using a different composite hash function; a candidate set construction module, which calculates the hash key of the multidimensional motion feature vector in each hash table, retrieves and merges all historical feature vectors in all corresponding hash buckets, and forms a candidate set after deduplication; and a matching and decision module, which calculates the similarity between the multidimensional motion feature vector and each vector in the candidate set. If the highest similarity exceeds a threshold, it indicates a successful match, and the corresponding historical ID is associated; otherwise, a new ID is assigned and the LSH index is updated.

Images