Multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching

Abstract

The application provides a multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching, first, real-time collection of heterogeneous data of different sensors is carried out; second, dynamic space-time registration processing is carried out on the heterogeneous data, and multi-sensor measurement data of space-time alignment is obtained; then, noise filtering and target preliminary screening are carried out on the measurement data, multi-modal data fusion is realized by combining sensor confidence weight, and multi-target position information after multi-sensor fusion is obtained; thereafter, a forward space-time enhanced cost matrix and a reverse space-time enhanced cost matrix are constructed and solved, bidirectional verification is carried out on the target point and the track, the obtained multi-target position information is matched with the existing track; finally, the target state is predicted through the extended Kalman filter, and the above process is dynamically optimized through the closed-loop feedback mechanism, stable tracks are obtained, and multi-target track tracking is completed. The application can obviously improve the continuous stable tracking capability of multi-target and provide reliable technical support for sea area situation awareness.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information processing, and relates to a multi-target track tracking method, in particular to a multi-target track tracking method based on dynamic noise-resistant clustering and bidirectional verification matching. BACKGROUND

[0002] With the growing demand for ocean resource development and maritime traffic supervision, high-precision monitoring technology for maritime targets has become a research hotspot for various countries. However, in a complex marine environment, a single sensor (such as a radar or sonar) is difficult to achieve stable detection due to the limitations of physical characteristics and working environment. Although multi-sensor cooperative networking can make up for the limitations of a single sensor, it still faces three major challenges: spatiotemporal asynchrony of heterogeneous data, complex noise interference, and uncertainty of track association.

[0003] On the one hand, different sensors (such as radars and sonars) have differences in sampling frequency, coordinate system, etc., resulting in non-uniform spatiotemporal reference, and traditional static registration methods are difficult to cope with dynamic scenarios such as target maneuvering and sensor clock drift. On the other hand, the superposition of marine environmental noise, sensor self-noise and multipath effect forms heterogeneous noise coupling, which makes traditional clustering algorithms prone to false detection or missed detection due to fixed parameter settings. In addition, track crossing and temporary occlusion of dense targets can easily cause ID jumping, and traditional one-way matching algorithms lack spatiotemporal continuity constraints, often leading to track fragmentation or identity confusion. Most existing technologies focus on single-sensor optimization or offline batch fusion, and it is difficult to balance real-time performance, robustness and adaptability, which cannot meet the needs of stable tracking of multiple targets in complex maritime environments. SUMMARY

[0004] To solve the above technical problems in the background art, the present application provides a multi-target track tracking method based on dynamic noise-resistant clustering and bidirectional verification matching, which can significantly improve the continuous and stable tracking capability of multiple targets and provide reliable technical support for maritime situation awareness.

[0005] The technical solution of the present application is as follows:

[0006] A multi-target track tracking method based on dynamic noise-resistant clustering and bidirectional verification matching, comprising the following steps:

[0007] Step 1: Real-time acquisition of heterogeneous data from different sensors;

[0008] Step 2: Dynamic spatiotemporal registration processing of the heterogeneous data obtained in step 1 to obtain spatiotemporally aligned multi-sensor measurement data;

[0009] Step 3: Noise filtering and target preliminary screening of the measurement data obtained in step 2, combined with sensor confidence weight to realize multi-modal data fusion, to obtain multi-target position information after multi-sensor fusion;

[0010] Step 4: By constructing and solving the forward spatiotemporal enhancement cost matrix and the inverse spatiotemporal enhancement cost matrix, the target point and the track are verified bidirectionally, and the multi-target position information obtained in Step 3 is matched with the existing track.

[0011] Step 5: Predict the target state by using extended Kalman filtering, and dynamically optimize steps 2 to 4 through a closed-loop feedback mechanism to obtain a stable track and complete multi-target track tracking.

[0012] In a further preferred embodiment, in step 1, heterogeneous data from different sensors are collected in real time via the UDP protocol; the sensors include at least radar sensors and sonar sensors; the heterogeneous data includes radar sensor measurement data and sonar sensor measurement data; the radar sensor measurement data is the latitude and longitude coordinates of the target with noise; the sonar sensor measurement data is the azimuth angle of the target and the distance information between the target and the sonar sensor.

[0013] A further optimized solution, the specific implementation process of step 2 is as follows:

[0014] Step 2.1: Convert the coordinates of the radar sensor measurement data and the sonar sensor measurement data obtained in Step 1 to the UTM coordinate system to complete the spatial transformation of the radar sensor measurement data and the sonar sensor measurement data.

[0015] Step 2.2: The sliding window dynamic time warping algorithm is used to calculate the time offset compensation of the sonar sensor measurement data, eliminate the clock asynchronous error between the sonar sensor measurement data and the radar sensor measurement data, realize the time alignment of the radar sensor measurement data and the sonar sensor measurement data, and obtain spatiotemporally aligned multi-sensor measurement data.

[0016] A further optimized solution, the specific implementation process of step 3 is as follows:

[0017] Step 3.1: Dynamically adjust the neighborhood radius based on local density:

[0018] For each data point entering this step Calculate its relationship with neighboring The Euclidean distance between each neighboring point is used to generate data points. adaptive radius :

[0019]

[0020] in for The point in the middle, It's point P. Nearest neighbor set, determined by distance from point The recent data points consist of, a preset number of neighbors;

[0021] Adaptive radius Set the upper limit of the radius , get the amplitude limited adaptive radius:

[0022]

[0023] wherein is a preset maximum radius threshold;

[0024] Step 3.2: Based on the adaptive radius obtained in step 3.1 , the dynamic anti-noise DBSCAN clustering algorithm is used to remove noise points from the spatio-temporal aligned multi-sensor measurement data obtained in step 2, extract the target cluster with consistent space-time in the measurement data, and complete the noise filtering and target preliminary screening;

[0025] Step 3.3: For the target cluster processed in step 3.2, a weighted assignment fusion weight is used to generate multi-target position information after multi-sensor fusion: for each target cluster data, the target position after weighted fusion is calculated :

[0026]

[0027]

[0028] wherein is the number of data points in the target cluster, is the measurement value of the th sensor in the target cluster, is the sensor measurement weight.

[0029] Further preferred scheme, the specific implementation process of step 4 is:

[0030] Step 4.1: Generate a forward spatio-temporal enhanced cost matrix and a reverse spatio-temporal enhanced cost matrix by combining the Euclidean distance between the target point and the predicted position, the motion consistency and the time window.

[0031] Step 4.2: Adopt the Hungarian algorithm to match and solve the forward spatio-temporal enhanced cost matrix and the reverse spatio-temporal enhanced cost matrix respectively, remove the matching pairs with spatio-temporal continuity conflict, and realize the matching of target position information and existing track.

[0032] Further preferred scheme, in step 4, the forward spatio-temporal enhanced cost matrix is:

[0033]

[0034]

[0035]

[0036]

[0037] in The first in the positive spatiotemporal enhancement cost matrix element, The cost is the positive Euclidean distance. For the sake of consistency in positive motion, Time window constraint; For the first The location of the target point to be matched. For the first track The position is predicted by extended Kalman filtering; These are the weighting coefficients. To predict track speed With the The velocity of the target point to be matched Cosine similarity; For the first The timestamp of the target point to be matched For the flight path The timestamp of the last frame, This is the maximum permissible time difference;

[0038] The inverse spatiotemporal enhancement cost matrix for:

[0039]

[0040]

[0041]

[0042]

[0043] in The first in the inverse spatiotemporal enhancement cost matrix element, For the cost of reverse Euclidean distance, The cost of consistency in reverse motion; track The position of the last frame, track The position of the second to last frame, This represents the time difference between adjacent points on the flight path.

[0044] The further optimized solution, the specific process of step 4.2 is as follows:

[0045] The Hungarian algorithm is used to solve the positive spatiotemporal enhancement cost matrix. Find the set of target points and forward-aligned tracks that minimize the total cost. The Hungarian algorithm is used to solve the inverse spatiotemporal augmentation cost matrix. Find the set of target points and reverse track matching pairs that minimize the total cost. If there is a discrepancy between the forward matching result and the reverse matching result, a conflict resolution strategy is adopted to select the pair with the smaller sum of overall costs.

[0046] For target point matching track Then update the track. state;

[0047] For unmatched target points, a new track is initialized and assigned a unique ID;

[0048] For unmatched tracks, their positions are continuously predicted. If multiple consecutive frames fail to match, the track is marked as lost or terminated.

[0049] A further optimized solution, step 5, is as follows:

[0050] Step 5.1: For a track that successfully matches the target point, use the observation data of the matched target point as the observation information of the track, combine it with the original track state, perform state update on the track through extended Kalman filter, and obtain the tracking residual; after the update, based on the latest estimated state, further predict the state of each track for the next time step.

[0051] Step 5.2: Pass the tracking residual to the spatiotemporal registration step in Step 2, dynamically update the time offset compensation amount and coordinate deviation weight; optimize the neighborhood radius of the clustering algorithm in Step 3 in reverse according to the target loss rate and false detection rate; adaptively adjust the weight coefficient of motion consistency cost in Step 4 based on the track survival period, form a perception-tracking-optimization closed loop, obtain a stable track, and complete multi-target track tracking.

[0052] A further optimized solution, the specific process in step 5.1 is as follows:

[0053] According to the formula

[0054]

[0055]

[0056]

[0057] updating the track state vector wherein is the track Kalman gain matrix at time k, is the track predicted state covariance matrix at time k, is the Jacobian matrix of the observation model, is the transpose of is the observation noise covariance, is the track updated state covariance matrix at time k, I is the identity matrix, is the observation model, is the track final state vector estimate at time k, is the track predicted state vector at time k, is the observation data of the target point matched to the track at time k; and according to the formula

[0058]

[0059]

[0060]

[0061] performing a state prediction for the track for the next time, wherein is the track predicted state covariance matrix at time k+1, is the Jacobian matrix of the nonlinear state transition function is the transpose of is the process noise covariance matrix at time k, is the track predicted state vector at time k+1; and dynamically adjusting the process noise covariance matrix based on the tracking residual

[0062]

[0063]

[0064] wherein , is an adaptive coefficient.

[0065] In a further preferred embodiment, the specific procedure in step 5.2 is as follows:​​​​

[0066] spatio-temporal registration compensation: the tracking residual is fed back to the spatio-temporal registration step in step 2 to update the time offset compensation

[0067]

[0068] wherein is the time offset compensation at time k, is the time offset compensation at time k+1, is the learning rate;

[0069] cluster parameter optimization: adjusting the neighborhood radius of the clustering algorithm in step 3 according to the target loss rate and the false detection rate

[0070]

[0071]

[0072]

[0073] wherein is the adaptive radius of the data point at time k, is the adaptive radius of the data point at time k+1, is the target loss rate balance coefficient, is the false detection rate balance coefficient; is the current number of active tracks, is the number of tracks determined to be lost per unit time; is the number of all tracks generated by the system per unit time, is the number of tracks marked as false;

[0074] motion consistency cost weight adaptive adjustment: adjusting the motion consistency cost weight coefficient according to the track survival period

[0075]

[0076] wherein is the motion consistency cost weight coefficient corresponding to the track is the preset initial motion consistency weight, and τ is a scale parameter, is the track survival period corresponding to the track

[0077] ​​​​​​Beneficial effects:

[0078] The application provides a multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching.

[0079] The spatiotemporal reference uniformity is improved: by using dynamic time warping (DTW) based multi-sensor clock asynchronous compensation precision improvement, the spatial registration residual in the ship maneuvering scene is reduced, and the fusion positioning drift caused by coordinate deviation is effectively avoided.

[0080] Anti-interference and robustness optimization: by using an adaptive neighborhood radius calculation method based on a dynamic anti-noise clustering algorithm, combined with sensor confidence weight, the clustering density threshold is dynamically adjusted, the real target and noise point can be effectively distinguished in a complex noise environment, and the outlier misjudgment rate is reduced.

[0081] Enhance the reliability of target association in complex scenes: by using an improved bidirectional verification track matching algorithm, the misassociation rate in the target dense intersection scene is greatly reduced, and the ID jump frequency is reduced.

[0082] It has full-process closed-loop adaptive ability: through the feedback of multi-dimensional indexes such as tracking residual and target loss rate, the dynamic optimization of spatiotemporal registration, clustering parameters and association threshold is realized, which greatly enhances the adaptability of the system in the high maneuvering environment, and the overall positioning precision fluctuation range is controlled within ±5 meters.

[0083] The application is compatible with radar, sonar and other types of sensors, and realizes real-time sea area situation awareness through unique ID generation and satellite map mapping, and meets the high-concurrency scene demand of maritime supervision, search and rescue and the like.

[0084] The dynamic closed-loop multi-source heterogeneous data fusion and track management framework proposed in the application fuses dynamic time warping (DTW), weighted anti-noise clustering and bidirectional spatiotemporal enhanced matching algorithms, and introduces a closed-loop feedback optimization mechanism, realizes the spatiotemporal joint registration of multi-sensor data, heteroscedastic noise suppression and track identity uniqueness guarantee, significantly improves the multi-target continuous stable tracking ability in complex sea conditions, and provides theoretical innovation and technical support for maritime situation awareness and collaborative decision-making.

[0085] The application realizes real-time optimization of the whole process from multi-source data access, preprocessing to track association and tracking by adopting space-time registration technology, anti-noise clustering and multi-modal data fusion, bidirectional verification track matching based on space-time constraints, and closed-loop feedback mechanism based on extended Kalman filtering, effectively solves the problems of clock asynchrony, coordinate deviation, noise interference, and cross-media target misassociation and error accumulation in multi-platform (radar, sonar) detection in complex sea environment, and significantly improves the accuracy, robustness and real-time performance of multi-target tracking on water, in the air and underwater, thereby providing reliable technical support for sea situation awareness.

[0086] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0087] The above and / or additional aspects and advantages of the application will become apparent and be readily understood by considering the following detailed description, from which the aspects and advantages of the application will be readily understood, taken in conjunction with the accompanying drawings.

[0088] Figure 1 is the overall flowchart of the multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching provided by the application;

[0089] Figure 2 is a schematic diagram of the dynamic anti-noise clustering and data fusion process;

[0090] Figure 3 is a bidirectional verification matching strategy diagram used by the application;

[0091] Figure 4 is the target S-type motion real track and multi-radar, multi-sonar observation data;

[0092] Figure 5 is a comparison of several baseline algorithm tracking tracks and target real tracks;

[0093] Figure 6 is a comparison of the positioning error of the algorithm tracking track;

[0094] Figure 7 is a schematic diagram of the intersection scene of two targets moving at a constant speed;

[0095] Figure 8 is a tracking result diagram of various algorithms in the intersection scene. DETAILED DESCRIPTION

[0096] The embodiments of the application are described in detail below, which are exemplary and intended to explain the application, and cannot be understood as a limitation of the application.

[0097] Reference Figure 1The application provides a multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching, and shows the whole process of multi-sensor data from collection, registration, clustering, fusion, track matching, track prediction, feedback optimization and track output. Table 1 shows the key data form evolution of data in different steps.

[0098] Table 1 Method key data form evolution

[0099]

[0100] The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching provided by the application has the following core ideas: first, the heterogeneous data of multiple radars and multiple sonar sensors are collected in real time by using the UDP protocol, including the latitude and longitude coordinates, the azimuth angle and the distance information of the target; second, the radar and sonar coordinates are converted to the UTM coordinate system by dynamic space-time registration, and the asynchronous sampling error is eliminated by using the sliding window dynamic time warping (DTW) algorithm, so that the space-time alignment is realized; third, the dynamic anti-noise DBSCAN clustering algorithm is used to filter noise and preliminarily screen targets based on multi-source data, and the multi-modal data fusion is realized by combining the sensor confidence weight; then, the bidirectional verification Hungarian algorithm is used for track matching by constructing a forward and reverse space-time enhanced cost matrix, so as to ensure the uniqueness of the target ID and the stability of the track; further, the target state is predicted by using the extended Kalman filter, and the space-time registration, clustering parameters and fusion weight are dynamically optimized by using the closed-loop feedback mechanism, so as to improve the tracking robustness of the high-maneuvering target; finally, the stable track is projected to the satellite map in real time, and the multi-dimensional situation information such as target position and speed is output, so as to support multi-platform cooperative monitoring and command decision, and significantly improve the sea target monitoring capability.

[0101] Specifically, the multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching provided by the application includes the following steps:

[0102] Step 1: Collecting multi-source heterogeneous data. In this step, the application acquires the measurement data from multiple heterogeneous sensors (including multiple radars and multiple sonars) in real time by using the UDP transmission protocol, and extracts the time stamp and target position information in the data. Among them, the multi-radar sensor can provide the target latitude and longitude coordinates disturbed by noise, and the multi-sonar sensor can provide the target azimuth angle and the distance information between the target and the sensor.

[0103] Step 2: Dynamic space-time registration. Next, the global coordinate system of radar measurement (WGS84 coordinate system) and the polar coordinate system of sonar measurement are converted to the same Cartesian coordinate system (UTM coordinate system), completing the spatial conversion. To solve the problem of asynchronous sampling of multiple radars and multiple sonars, the sliding window dynamic time warping (DTW) algorithm is used to calculate the time offset compensation of sonar data based on high-precision radar data, eliminate clock asynchronous error, and realize time alignment. The steps of spatial alignment and time alignment will be introduced in detail below.

[0104] Step 2-1: Spatial alignment. In the data fusion process, it is necessary to convert the data provided by different sensors to the same coordinate system for easy calculation and analysis. Cartesian coordinate system (UTM coordinate system) is a global projection coordinate system used to convert the spherical latitude and longitude of the earth (WGS84) to a plane rectangular coordinate for surveying, navigation and geographic information analysis. The global coordinate system of radar measurement (WGS84 coordinate system) and the polar coordinate system of sonar measurement are converted to the same Cartesian coordinate system, completing the spatial alignment.

[0105] WGS84 geographic coordinate system to UTM coordinate system: use WGS84 ellipsoid parameters to calculate UTM projection coordinates:

[0106]

[0107]

[0108] Where: , The east and north coordinates of the target point obtained by radar are respectively, , The latitude and longitude of the target point (WGS84 coordinate system) are respectively, The central meridian longitude of the UTM zone, =0.9996 is the central meridian scale factor, used to scale the coordinate values after projection to reduce distortion error, is the length of the WGS84 ellipsoid major axis, is the first eccentricity.

[0109] Sonar polar coordinate system to Cartesian coordinate system step: sonar output is usually polar coordinates ( , , ), where is the distance, is the azimuth, is the elevation, converted to Cartesian coordinates:

[0110]

[0111] wherein , are the east and north coordinates of the target point obtained by the sonar, respectively.

[0112] Step 2-2: Time synchronization. To solve the problem of asynchronous sampling of multi-radar and multi-sonar sensors, a sliding window dynamic time warping (DTW) algorithm is used. A high-precision radar data is selected as the reference, and the time offset compensation of the sonar data is calculated to eliminate the clock asynchronous error and achieve time alignment.

[0113] Define window and sequence: Let the radar data window be , and the sonar data window be , , where is the number of data points in the radar data window, and is the number of data points in the sonar data window. The preset window length is , which is adaptively adjusted according to the sampling rate. When the buffer area of the radar and the sonar is filled with the preset window size (e.g. 300ms data), the alignment calculation is triggered.

[0114] Calculate the time average offset in the window:

[0115]

[0116] is the time offset compensation of the sonar data relative to the radar data. If >0, it means that the sonar data lags and needs to be moved forward; if <0, it means that the sonar data leads and needs to be moved backward. is the preset number of data points in the window.

[0117] Smooth the time offset by exponential weighted moving average (EWMA) to suppress the random fluctuations of single-window calculation and improve the stability of compensation. The smooth formula for time offset compensation is:

[0118]

[0119] where is the smoothed time offset compensation, representing the global offset accumulated over time; is the smoothing coefficient, used to control the weight of historical data, with a value range of γ∈[0.8,0.95] ;

[0120] Subtract the global lag from the original timestamp of the sonar to achieve time axis alignment:

[0121]

[0122] wherein is the aligned sonar timestamp.

[0123] For each radar timestamp , find the adjacent sonar data points , compute the interpolated coordinates:

[0124]

[0125] Similarly compute , i.e. the spatio-temporally aligned sonar data. Subsequently, input the computed , and radar data to step 3 for data fusion.

[0126] Step 3: Anti-noise clustering and data fusion. Firstly, based on the dynamic anti-noise DBSCAN clustering algorithm, the multi-source sensor data after spatio-temporal alignment is subjected to target preliminary screening and noise filtering. The clustering search radius is dynamically adjusted through adaptive neighborhood radius calculation, and the radar data and sonar data are subjected to differential processing in combination with the sensor confidence weight, so as to reduce the outlier misjudgment probability of low signal-to-noise ratio sensors, and output a set of effective target clusters consistent in space and time. For the effective target cluster set after preliminary screening, a fusion weight is assigned and a high-precision target position estimate is generated, so as to realize multi-modal data fusion. This step screens effective targets through dynamic anti-noise clustering, and improves the multi-source data positioning accuracy based on weighted fusion. The specific process is as follows:

[0127] Step 3-1: Dynamic anti-noise DBSCAN clustering, realizing target preliminary screening and noise filtering, so as to eliminate noise points from multi-sensor data and extract target clusters consistent in space and time. The traditional DBSCAN clustering algorithm is a commonly used density-based clustering algorithm, including a fixed neighborhood radius (Eps), but in the multi-source heterogeneous sensor data scene, the parameter solidification will cause the following problems: sensor accuracy difference leads to noise misjudgment (such as low confidence sensor data being easily misclustered), and fixed radius cannot adapt to the dynamic changes of spatial density (leading to over-segmentation or missed detection). This method realizes dynamic anti-noise clustering through adaptive neighborhood search, and the specific process is as follows:

[0128] The neighborhood radius is dynamically adjusted according to the local density to avoid parameter solidification. In dense areas, the search range is reduced to avoid over-segmentation, and in sparse areas, the range is expanded to prevent missed detection. Firstly, for each data point entering this step, the Euclidean distance between it and the adjacent neighborhood points is calculated, and the adaptive radius of the data point is generated:

[0129]

[0130] wherein: is a point in is the neighborhood set of point p, consisting of the closest data points, is a preset number of neighbors, determined by the number of sensors, for dense areas is reduced, while for sparse areas is enlarged.

[0131] To avoid misclassification of irrelevant points into a cluster due to an excessively large radius in sparse areas, a further upper limit on the radius is set , thus obtaining the amplitude-limited adaptive radius:

[0132]

[0133] wherein, is a preset maximum radius threshold; if at least MinPts points are contained in the neighborhood of point p with a radius , it is marked as a core point, wherein MinPts is a preset threshold.

[0134] Step 3-2: Data fusion. For the set of valid target clusters after preliminary screening, a weighted distribution fusion weight is used to generate a high-precision target position estimate. For each target cluster data, the target position after weighted fusion is calculated :

[0135]

[0136]

[0137] wherein is the number of data points in the target cluster, is the measurement value of the i-th sensor in the target cluster, is the sensor measurement weight, and the higher the measurement accuracy, the greater the weight. Figure 2 The process of noise-resistant clustering and multi-modal data fusion is shown. Through this step, multi-target position information after multi-sensor fusion is obtained.

[0138] Step 4: Two-way verification track matching with space-time constraints. In this step, the fused target points are matched with existing tracks to ensure the uniqueness of target ID by constructing forward and reverse space-time enhanced cost matrices. First, the Euclidean distance between target points and track predicted positions, motion consistency, and time window are considered to generate forward and reverse space-time enhanced cost matrices. Hungarian algorithm is used to match the forward and reverse space-time enhanced cost matrices to achieve two-way verification and eliminate matching pairs with space-time continuity conflicts. Finally, the two-way verification mechanism ensures the space-time continuity of the matching results. Unmatched target points are initialized as new tracks, and unmatched tracks are marked as lost or terminated, ensuring the uniqueness of target ID and track stability in complex scenarios. By constructing a two-way verification track matching with space-time constraints, the fused target points are matched with existing tracks to ensure the uniqueness of target ID. The specific process is as follows:

[0139] Step 4-1: To achieve accurate association between fused target points and existing tracks, a two-way space-time enhanced cost matrix is proposed to quantify the matching possibility of target points and tracks by considering spatial proximity, motion consistency, and time continuity. Assume that the existing track set is where the th track is obtained from the last frame of the track p j k-1 =[ x j k-1 , y j k-1 ] T and the position information is obtained by extending Kalman filtering p ̂ j k =[ x j k , y j k ] T and velocity information . The fused target point set is where the th target point to be matched is the target position after weighted fusion in step 3 [ x ̂ i , y ̂ i ] T .

[0140] Forward space-time enhanced cost matrix is composed of the following three parts:

[0141]

[0142] ​wherein is the element in the forward spatio-temporal cost matrix, is the forward Euclidean distance cost, is the forward motion consistency cost, is the time window constraint.

[0143] Forward Euclidean distance cost: the Euclidean distance between the target point and the predicted position of the track is calculated as:

[0144]

[0145] The smaller the distance, the closer the spatial position of the target point and the track.

[0146] Forward motion consistency cost: for the track , its predicted velocity is calculated by the extended Kalman filter; while the velocity of the target point is estimated by the difference between its position and the last frame position of the track:

[0147]

[0148] wherein is the time difference between adjacent points in the track.

[0149] The cosine similarity between the predicted track velocity and the velocity of the target point is calculated as:

[0150]

[0151] The motion consistency cost is then calculated as:

[0152]

[0153] wherein is the weight coefficient, cos θ ij ∈[-1 ， 1] The more consistent the direction, the smaller the motion consistency cost.

[0154] Time window constraint: used to limit the time difference between the target point and the track. The time window constraint ensures that the new target will not be incorrectly associated with the old track that has disappeared:

[0155] ​​​​​

[0156] in For target point timestamp, For the flight path The last frame timestamp, To determine the maximum permissible time difference, 1.5 seconds is used in this embodiment.

[0157] Similarly, to verify the spatiotemporal continuity of track matching, it is also necessary to construct an inverse spatiotemporal enhancement cost matrix based on the historical state of the track (rather than the predicted state). Its structure is similar to the forward spatiotemporal enhancement cost matrix, but the calculation reference point is different. At this point:

[0158]

[0159] in The first in the inverse spatiotemporal enhancement cost matrix element, For the cost of reverse Euclidean distance, This comes at the cost of consistency in reverse motion.

[0160] Inverse Euclidean distance cost: Calculates the geometric distance between the target point and the historical position of the track (rather than the future position predicted by the extended Kalman filter):

[0161]

[0162] Inverse motion consistency cost: Checking the consistency between the target point and the historical motion direction of the track (rather than the future motion direction predicted by the extended Kalman filter):

[0163]

[0164]

[0165] in For the flight path The position of the second to last frame in the middle.

[0166] Step 4-2, bidirectional verification of Hungarian match:

[0167] Forward Matching: Solving the Forward Spatiotemporal Augmentation Cost Matrix Using the Hungarian Algorithm Find the set of target points and forward track matching pairs that minimize the total cost. This ensures that each target point can be matched with at most one track.

[0168] Reverse verification: Solving the inverse spatiotemporal augmentation cost matrix using the Hungarian algorithm Find the set of target points and reverse track matching pairs that minimize the total cost. , verify the rationality of matching pairs.

[0169] After solving the forward matching pair set and the reverse matching pair set by the Hungarian algorithm respectively, if the forward and reverse matching results are inconsistent, such as the target point Forward matching track Reverse matching track If the conflict resolution strategy is adopted, the pair with smaller comprehensive cost is selected, that is:

[0170]

[0171] For the target point Comprehensive matching track.

[0172] For the target point matching track , update the track State;

[0173] For the unmatched target point, initialize a new track and assign a unique ID;

[0174] For the unmatched track, continuously predict its position, and if it is not matched for several consecutive frames, mark the track as lost or terminated.

[0175] As Figure 3 The bidirectional verification matching strategy diagram is shown.

[0176] Step 5: Track prediction and feedback optimization. For each track, perform extended Kalman filter update independently, predict the target state at the next time, and dynamically adjust the process noise covariance matrix according to the residual error between the new observation data and the predicted value, to enhance the tracking robustness of high maneuvering targets; At the same time, design a full-process closed-loop feedback mechanism: pass the tracking residual to the time and space registration step to dynamically update the time offset compensation amount; According to the target loss rate and false detection rate, optimize the neighborhood radius of the clustering algorithm in reverse; Based on the survival period of the track, adaptively adjust the weight of the time and space enhancement cost matrix, forming a "perception-tracking-optimization" closed loop. Through track prediction and feedback optimization, robust tracking of high maneuvering targets is realized and the full-process adaptability is improved. Specifically:

[0177] Step 5-1, track prediction: for the track successfully matched to the target point, the observation data of the matched target point is used as the observation information of the track, combined with the original track state, the track is executed state update through extended Kalman filter, and the tracking residual is obtained; after updating, based on the latest estimated state, further state prediction of each track at next time is carried out; the extended Kalman filter (EKF) linearizes the nonlinear motion model through Jacobian matrix, and combines the adaptive mechanism of dynamically adjusting process noise covariance, realizes the robust state estimation and tracking of high maneuvering target. The present application adopts extended Kalman filter (EKF) to carry out independent state update and prediction for each track, and the specific process is as follows:

[0178] Suppose the target point is matched with the track , the track is in the state vector at k-1 time X j k-1 =[ x j k-1 , y j k-1 , v j,x k-1 , v j,y k-1 ] T , is the position information of the track at k-1 time, is the speed information of the track at k-1 time; is the predicted state vector of the track at k time, containing position information and speed information, and the observation data of the target point matched with the track at the current k time is .

[0179] (1) update the track state vector by using EKF :

[0180]

[0181]

[0182]

[0183] Among them, is the Kalman gain matrix of the track at k time, used to balance the weight of the predicted state and the observation value; is the predicted state covariance matrix of the track at k time, indicating the prediction uncertainty from time to k time; Jk= ∂hk∂xk HkT HkT Rk Pk Updated state covariance matrix at time k, representing the uncertainty after fusing observations; I is the identity matrix. hk(xk) Pk Final state vector estimate at time k.

[0184] (2) Predict state:

[0185]

[0186]

[0187] where Pk Predicted state covariance matrix at time k+1, Jk+1 HkT HkT Process noise covariance matrix at time k, Pk+1 Predicted state vector at time k+1. (3) Adaptive adjustment of process noise: To cope with target sudden maneuvers, the tracking residual

[0188] is used to dynamically adjust the process noise covariance matrix:

[0189] where

[0190] is the adaptive coefficient, and the residual increases to increase the process noise covariance to enhance the tracking sensitivity to target sudden maneuvers. After track update and prediction, the system will output the dynamic target position, velocity and uncertainty quantification fused with multi-sensor observations, as well as the next time target position and error range to support real-time decision-making. In addition, for high-maneuvering targets, the system will adaptively adjust the process noise covariance based on the observation residual to enhance the adaptability to sudden maneuvers.

[0191]

[0192] ​Step 5-2, the tracking residual is passed to the space-time registration step in step 2 to dynamically update the time offset compensation and coordinate deviation weight; the target loss rate and false detection rate are used to optimize the neighborhood radius of the clustering algorithm in step 3 in reverse; the weight coefficient of the motion consistency cost in step 4 is adaptively adjusted based on the survival period of the track, forming a perception-tracking-optimization closed loop to obtain stable tracks and complete multi-target track tracking.

[0193] The specific process is as follows:

[0194] (1) Space-time registration compensation: the tracking residual is fed back to the space-time registration step in step 2 to update the time offset compensation :

[0195]

[0196] wherein is the time offset compensation at time k, is the time offset compensation at time k+1, is the learning rate, which is calculated by back propagation to calculate the sensitivity of time offset to residual, and gradually eliminates the cumulative error caused by sensor clock drift or network delay.

[0197] (2) Clustering parameter optimization: according to the target loss rate and the false detection rate , the neighborhood radius of the clustering algorithm in step 3 is adjusted in reverse :

[0198]

[0199]

[0200]

[0201] wherein is the adaptive radius of the data point at time k, is the adaptive radius of the data point at time k+1, is the target loss rate balance coefficient, is the false detection rate balance coefficient, when is high, it may mean that the clustering is too tight, causing some targets to be missed, so the neighborhood radius is increased to capture more points, when is high, it may mean that the clustering is too loose, causing noise to be mistaken for targets, so the neighborhood radius is reduced to reduce false detection. In the target loss rate formula, is the current number of active tracks, is the number of tracks determined to be lost in unit time, and the condition for determining loss in the embodiment is continuous confirmed target point; in the false detection rate formula, is the number of all tracks generated by the system in unit time, is the number of tracks marked as false, and the condition for determining false in the embodiment is motion contradiction.

[0202] (3) Motion consistency cost weight self-adaption: adjusting the motion consistency cost weight coefficient according to the track survival period :

[0203]

[0204] wherein is the track corresponding motion consistency cost weight coefficient, is the preset initial motion consistency weight, and tau is a scale parameter, used for controlling the sensitivity of the weight to the track survival period ; is the track corresponding track survival period; the long survival track will increase the motion consistency weight (the weight increases with ).

[0205] The prediction result is combined with a closed-loop feedback mechanism to further optimize the sensor registration parameters, clustering threshold and cost matrix weight, form a dynamically adjusted "prediction-tracking-calibration" closed loop, and finally support stable association of target ID and robust tracking in complex scenes. Subsequently, the stable track is projected to the satellite map in real time, and multi-dimensional situation information such as target position and speed is output, supporting multi-platform cooperative monitoring, situation awareness and command decision, and improving the sea target monitoring capability.

[0206] In order to verify the effectiveness of the algorithm proposed in the application, the ablation experiment settings in Table 2 are used to test the algorithm performance in terms of positioning accuracy and ID tracking stability.

[0207] Table 2 Ablation experiment algorithm design

[0208]

[0209] Figure 4 , Figure 5 and Figure 6are positioning accuracy performance analysis respectively. The application performs best on the RMSE (Root Mean Square Error) index, and its dynamic noise resistance and closed-loop feedback mechanism effectively suppresses the error accumulation at the curved part of the track. The simple Hungarian matching produces a lag error at the turning point of the maneuver because it does not consider the motion model; the standard KF filter (RMSE=10.4m) fails to effectively handle nonlinear maneuvers and has low overall accuracy; the fixed parameter EKF has low overall error, but it exposes the defect of poor adaptability to sudden changes in the track due to parameter solidification; the DBSCAN clustering algorithm produces a large number of outliers because the fixed neighborhood radius cannot adapt to the changes in the spatial distribution of the maneuvering target. The experiments verify the high positioning accuracy of the method in complex maneuvering target tracking scenarios.

[0210] Figure 7 and Figure 8 are ID tracking stability analysis respectively. In the target track intersection scene, the ID maintenance ability of each algorithm shows significant differences. The method of the application achieves zero ID switching near the track intersection point through the synergistic optimization of dynamic noise resistance clustering and bidirectional verification matching, verifying the correlation robustness in complex scenarios. The simple Hungarian matching produces track confusion near the intersection point, resulting in frequent ID jumps, because it only relies on position similarity. The standard KF algorithm and the fixed parameter EKF algorithm only change the track update and prediction algorithms, and the dynamic noise resistance DBSCAN clustering and bidirectional verification matching modules still work, and both can track the target well without ID switching. The DBSCAN clustering produces 7 false associations because the fixed neighborhood radius cannot adapt to the density changes in the intersection area. Experiments show that the dynamic noise resistance clustering and bidirectional verification matching modules proposed by the application can effectively improve the ID stability in complex scenarios.

[0211] Although the embodiments of the application have been shown and described above, it should be understood that the above-described embodiments are exemplary and should not be construed as limiting the application, and those of ordinary skill in the art can make changes, modifications, replacements and variations to the above-described embodiments without departing from the principles and purposes of the application within the scope of the application.

Claims

1. A multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching, characterized by: The method comprises the following steps: Step 1: real-time acquisition of heterogeneous data of different sensors; Step 2: dynamic space-time registration processing of the heterogeneous data obtained in step 1 to obtain multi-sensor measurement data aligned in space-time; Step 3: noise filtering and target preliminary screening of the measurement data obtained in step 2, combined with sensor confidence weight to realize multi-modal data fusion, to obtain multi-target position information after multi-sensor fusion; Step 4: constructing and solving forward and reverse space-time enhanced cost matrices to bidirectionally verify target points and tracks, and matching the multi-target position information obtained in step 3 with existing tracks; Step 5: predicting target state through extended Kalman filtering and dynamically optimizing steps 2 to 4 through a closed-loop feedback mechanism to obtain stable tracks and complete multi-target track tracking; The specific implementation process of step 3 is as follows: Step 3.1: dynamically adjusting the neighborhood radius according to the local density; For each data point entering this step , calculate its Euclidean distance with its adjacent neighbor points, generate the adaptive radius of the data point : wherein is a point in is a neighbor set of the point p, consisting of the closest data points to the point p, is a preset number of neighbors; Adaptive radius Setting upper limit on radius , obtaining the amplitude-limited adaptive radius: wherein is a preset maximum radius threshold value; Step 3.2: adaptive radius based on the result of step 3.1 Step 2: Based on the multi-sensor measurement data obtained in step 1, the dynamic anti-noise DBSCAN clustering algorithm is used to remove noise points, extract target clusters with consistent space-time in the measurement data, and complete noise filtering and target preliminary screening. Step 3.3: For the target cluster data after step 3.2, the multi-sensor fusion target position information after multi-sensor fusion is generated by using weighted distribution fusion weight: for each target cluster data, the target position after weighted fusion is calculated : wherein is the number of data points in the target cluster, is the measurement of the sensor in the target cluster, is the sensor measurement weight; The specific implementation process of step 4 is as follows: Step 4.1: Generate the forward spatio-temporal augmented cost matrix by combining the Euclidean distance between the target point and the predicted position of the track, the motion consistency and the time window and the backward spatio-temporal augmented cost matrix ; Step 4.2: using the Hungarian algorithm to match and solve the forward and reverse space-time enhanced cost matrices respectively, eliminating matching pairs with space-time continuity conflicts, and realizing matching of target position information and existing tracks.

2. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 1, characterized in that: In step 1, heterogeneous data of different sensors is collected in real time through the UDP protocol; the sensors at least include a radar sensor and a sonar sensor; the heterogeneous data includes radar sensor measurement data and sonar sensor measurement data; the radar sensor measurement data is the latitude and longitude coordinates of the target with noise; and the sonar sensor measurement data is the azimuth angle of the target to be measured and the distance information between the target to be measured and the sonar sensor.

3. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 2, characterized in that: The specific implementation process of step 2 is as follows: Step 2.1: uniformly converting the coordinates of the radar sensor measurement data and the coordinates of the sonar sensor measurement data obtained in step 1 to the UTM coordinate system to complete the spatial conversion of the radar sensor measurement data and the sonar sensor measurement data; Step 2.2: calculating the time offset compensation amount of the sonar sensor measurement data by using the sliding window dynamic time warping algorithm to eliminate the clock asynchronous error between the sonar sensor measurement data and the radar sensor measurement data, realizing time alignment of the radar sensor measurement data and the sonar sensor measurement data, and obtaining multi-sensor measurement data aligned in space-time.

4. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 3, characterized in that: In step 4, the forward space-time gain cost matrix is: in The first in the positive spatiotemporal enhancement cost matrix element, The cost is the positive Euclidean distance. The cost of positive motion consistency, Time window constraint; For the first The location of the target point to be matched. For the first track The position is predicted by extended Kalman filtering; These are the weighting coefficients. To predict track speed With the The velocity of the target point to be matched Cosine similarity; For the first The timestamp of the target point to be matched For the track The timestamp of the last frame, This is the maximum permissible time difference; the inverse spatiotemporal enhancement cost matrix is: wherein is the element in the inverse spatio-temporal augmented cost matrix, is the inverse Euclidean distance cost, is the inverse motion consistency cost; track position of the last frame, track position of the second last frame, is the time difference between adjacent points in the track.​ 5. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 3, characterized in that: The specific process of step 4.2 is as follows: Solving the forward spatio-temporal enhanced cost matrix using the Hungarian algorithm Finding the set of forward matching pairs of target points and tracks with the minimum total cost Solving the inverse spatio-temporal enhanced cost matrix using the Hungarian algorithm Finding the set of inverse matching pairs of target points and tracks with the minimum total cost If the forward matching result and the inverse matching result are inconsistent, a conflict resolution strategy is adopted to select the pair with a smaller total cost For target point matching track Then update the track State; For unmatched target points, a new track is initialized and a unique ID is assigned; For unmatched tracks, continuously predict the position, and if it is not matched for consecutive multiple frames, mark the track as lost or terminated.

6. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 4, characterized in that: The specific process of step 5 is as follows: Step 5.1: for the tracks successfully matched to the target points, using the observation data of the matched target points as the observation information of the track, combining the original track state, performing state update on the track through extended Kalman filtering, and obtaining tracking residuals; after updating, further predicting the state of each track at the next time based on the latest estimated state; Step 5.2: The tracking residual is passed to the space-time registration step in step 2, dynamically updating the time offset compensation and coordinate deviation weight; the target loss rate and false detection rate are used to optimize the neighborhood radius of the clustering algorithm in step 3 in reverse; the weight coefficient of the motion consistency cost in step 4 is adaptively adjusted based on the survival period of the track, forming a perception-tracking-optimization closed loop to obtain stable tracks and complete multi-target track tracking.

7. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 6, characterized in that: The specific process in step 5.1 is: According to the formula updated track state vector wherein is a track Kalman gain matrix at time k, is a track predicted state covariance matrix at time k, is a Jacobian matrix of the observation model, is the transpose of observation noise covariance, is a track updated state covariance matrix at time k, I is an identity matrix, is an observation model, is a track final state vector estimate at time k, is a track predicted state vector at time k of is observation data of a target point matched to track at time k; And according to the formula track Predict the state at the next time step, where For the track The predicted state covariance matrix at time k+1, It is a nonlinear state transition function Jacobian matrix, for transpose, Let k be the process noise covariance matrix at time k. For the track The predicted state vector at time k+1; and based on the tracking residual Dynamic adjustment of process noise covariance matrix: wherein , is an adaptive coefficient.

8. The multi-target track tracking method based on dynamic anti-noise clustering and bidirectional verification matching according to claim 7, characterized in that: The specific process in step 5.2 is: spatial-temporal registration compensation: the tracking residual is fed back to the spatial-temporal registration step in step 2, updating the temporal offset compensation amount : wherein is the time offset compensation quantity at k time instant, is the time offset compensation quantity at k+1 time instant, is the learning rate; Cluster parameter optimization: according to target loss rate and false detection rate , inversely adjust neighborhood radius of cluster algorithm in step 3 : wherein is the data point is the adaptive radius at k time, is the data point is the adaptive radius at k+1 time, is the target loss rate balance coefficient, is the false detection rate balance coefficient; is the current number of active tracks, is the number of tracks determined to be lost per unit time; is the total number of tracks generated by the system per unit time, is the number of tracks marked as false. Motion consistency cost weight adaptive adjustment: adjust motion consistency cost weight coefficient according to track survival period : wherein is a track is a corresponding motion consistency cost weight coefficient, is a preset initial motion consistency weight, and τ is a scale parameter, is a track is a corresponding track survival period.

Images