A method for tracking maritime targets based on multiple marine radars

Abstract

This invention relates to a method for tracking maritime targets based on multiple maritime radars. The method includes the following steps: Step 1: Using a clustering algorithm to identify targets from point cloud data acquired by the radar, obtaining the position and size information of the identified observations; Step 2: Matching the current observation with existing historical tracking trajectories based on a joint probabilistic data interconnection method, obtaining the state information of the matched trajectories; Step 3: Estimating the state of the track being tracked based on the state information of the matched trajectories, obtaining the state estimation results, including the target's state information and the corresponding estimated variance value; Step 4: Correcting the parameters of the clustering algorithm based on the state estimation results; Step 5: Fusion of the tracking results from multiple radars, that is, fusing the state estimation results from each radar individually. Compared with existing technologies, this invention has advantages such as alleviating the difficulty of setting target identification algorithm parameters and increasing system robustness.

Description

Technical Field

[0001] This invention relates to the field of radar target tracking technology, and in particular to a method for tracking maritime targets based on multiple marine radars. Background Technology

[0002] In recent years, autonomous surface vessels have received increasing attention because they are able to independently complete complex and dangerous tasks at sea. As a shipborne environmental perception sensor, the target tracking technology of marine radar has also become a hot topic in this field.

[0003] Most existing radar tracking technologies separate identification and tracking sequentially. Target identification often incorporates deep learning techniques, using deep neural networks to identify and classify targets within radar point clouds. This method relies heavily on large amounts of surface radar data, and the current lack of sufficient data compromises its generalization ability. Another target identification technique uses clustering algorithms, but this requires setting algorithm parameters, which are difficult to determine when tracking moving targets on the water surface. Therefore, for practical radar tracking systems, environmental adaptability and technical implementation difficulty must be considered. Regarding tracking algorithms, the core issue is filtered tracking in complex multi-target environments. In multi-target environments, correctly allocating observation points to each target is crucial, which involves data interconnection. Current research on data interconnection methods mainly falls into two categories: maximum likelihood algorithms and Bayesian algorithms. Maximum likelihood algorithms have not seen further development due to their complex engineering implementation. Bayesian algorithms, on the other hand, have been widely used because their concepts are easily combined with Kalman filter forms. For multi-radar data fusion, there are two main structures: centralized and distributed. Compared to centralized structures, distributed fusion offers advantages such as less susceptibility to environmental influences on data transmission, higher system stability, and higher information utilization. Therefore, distributed fusion methods are currently the optimal approach.

[0004] However, radar target tracking technology still has shortcomings in recognition. Separating recognition and tracking will lead to the inability to fully utilize the time information between multiple frames of data. Therefore, even though deep learning methods for recognition and tracking have been widely studied, their environmental adaptability is obviously not as good as clustering methods due to insufficient data. And using clustering methods requires solving the problem of online parameter updates. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for tracking maritime targets based on multiple marine radars.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for tracking maritime targets based on multiple marine radars, comprising the following steps:

[0008] Step 1: Use clustering algorithms to identify targets in the point cloud data acquired by radar, and obtain the location and size information of the identified targets;

[0009] Step 2: Based on the joint probabilistic data interconnection method, match the current observation with the existing historical tracking trajectory to obtain the status information of the matched trajectory;

[0010] Step 3: Estimate the state of the track being tracked based on the state information of the matched trajectory, and obtain the state estimation results, including the target's state information and the corresponding estimated variance value;

[0011] Step 4: Adjust the parameters of the clustering algorithm based on the state estimation results;

[0012] Step 5: Fusion of tracking results from multiple radars, that is, merging the state estimation results of each radar individually.

[0013] In step 1, the clustering algorithm identifies targets by determining the minimum distance between two categories. The process of using the clustering algorithm to identify targets from the point cloud data acquired by radar specifically includes the following steps:

[0014] Step 101: Transform the point cloud data acquired by the radar in the polar coordinate system into the rectangular coordinate system using coordinate transformation. The formula for coordinate transformation is:

[0015] x = r n cosθ n +X n

[0016] y = r n sinθ n +Y n

[0017] Where, (θ n ,r n Let (x, y) be the position of the nth point cloud data in the polar coordinate system, and (x, y) be the position of the nth point cloud data in the world coordinate system. n ,Y n () represents the radar's position in the world coordinate system when observing the nth point cloud data.

[0018] Step 102: While processing the coordinate system transformation, compress the point cloud data. Compress the point cloud data using the average value along the equal angle and equal distance directions observed by the radar, and store it in a radar_data matrix. The first four rows of the radar_data matrix store the point cloud data in polar coordinates and the world coordinate system. Calculate the measurement error matrix R in the rectangular coordinate system. The formula for calculating the measurement error matrix R is:

[0019] R = TΛT T

[0020] Where T is the transformation matrix and Λ is the radar measurement error covariance matrix;

[0021] The expression for the transformation matrix T is:

[0022]

[0023] Where (θ,r) represents the position in the polar coordinate system;

[0024] The expression for the radar measurement error covariance matrix Λ is:

[0025]

[0026] in, For radar radial observation variance, Variance of radar angle observations;

[0027] Step 103: Initialize the parameters of the target recognition algorithm based on the prior environmental information. The prior environmental information is the nearest distance threshold parameter set according to the density of the target. When the nearest distance between two point clusters is lower than the nearest distance threshold parameter, they are considered to belong to the same observation.

[0028] Step 104: Point clustering is achieved by iteratively calculating the similarity between point clusters to realize target recognition. A point cluster is a set of point cloud data that is considered to be of the same class during the clustering process. The similarity between point clusters is measured by the nearest distance between point clusters. The formula for calculating the nearest distance between point clusters is:

[0029]

[0030] Among them, A i and B j Let a and b represent the set of the i-th point cluster and the set of the j-th point cluster, respectively, where a and b are any points in the i-th point cluster and the j-th point cluster, respectively.

[0031] Step 105: For each point cloud data, the last row of the radar_data matrix stores the category to which the point cloud data belongs. For point clusters belonging to the same observation, the position of the centroid is used as the observation position, that is, the average X and Y coordinates of all point cloud data in a group are used as the position of the point cluster. By traversing the radar_data matrix and accumulating the point cloud data of each category, the centroid of the point cloud data of each category is calculated to obtain the position of the observation, and the size of the observation is calculated. The formula for calculating the size is:

[0032]

[0033] Among them, S k (t) represents the area at time t for the k-th observation, δ θ δ is the angular resolution of the radar data. R C represents the radial resolution of radar data. k (t) represents the radial distance of the l-th point cloud data belonging to the k-th observation.

[0034] Step 2, which involves matching the current observation with existing historical tracking trajectories based on the joint probabilistic data interconnection method, specifically includes the following steps:

[0035] Step 201: Calculate the distance between the current observed state information and the state information of each trajectory in the previous frame of data. Each trajectory in the previous frame of data is the historical tracking trajectory. The distance v between the current observed state information and the state information of each trajectory in the previous frame of data is... km The calculation formula is:

[0036]

[0037] Among them, [x k ,y k [x] represents the position of the k-th observation. m ,y m [S] represents the predicted location of the m-th historical tracking trajectory. k S is the area of ​​the kth observation. m S is the predicted area of ​​the m-th historical tracking trajectory. m (t) represents the area of ​​the m-th historical tracking trajectory at time t, λ is the weight constant, and N is the transformation matrix;

[0038] Step 202: Determine if the distance is less than the nearest distance threshold parameter. If so, there is a correlation probability between the observation and the tracking trajectory. Record the result using an acknowledgment matrix. The predicted state of the trajectory in the previous frame is obtained by second-order extrapolation.

[0039] X m (t|t-1)=Φm (t-1)X m (t-1|t-1)

[0040] Among them, X m (t|t-1) represents the prediction of the state at time t for the m-th historical tracking trajectory based on the state estimate at time t-1, Φ m (t-1) is the state transition matrix of the m-th historical tracking trajectory at time t-1, using a second-order linear model, X. m (t-1|t-1) represents the state estimate of the m-th historical tracking trajectory at time t-1;

[0041] The expression for the confirmation matrix is:

[0042] R = [r] km ] q×w ,(k=1,2,...,q),(m=1,2,...,w)

[0043]

[0044] Where, r km To confirm the elements in the matrix, when r km =1 indicates that the k-th observation and the m-th historical tracking trajectory can be correlated; when r km = 0 indicates that the k-th observation and the m-th historical tracking trajectory are not correlated; q is the number of observations; w is the number of historical tracking trajectories.

[0045] Step 203: By confirming the distance between the current observation's state information and the trajectory's state information in the previous frame, obtain the association probability between each observation in the current data frame and each trajectory in the previous frame. The association probability is calculated using the joint probability data interconnection concept, as follows:

[0046]

[0047]

[0048]

[0049]

[0050] Among them, P km G refers to the probability that the k-th observation is associated with the m-th trajectory being tracked. The m-th trajectory being tracked is the m-th historical track. km A is a quantity proportional to the Gaussian likelihood function, which can be called the effective likelihood function. It is used to characterize the correlation between the k-th observation and the m-th trajectory being tracked. mA is the sum of the effective likelihood functions of all observations of the m-th trajectory being tracked. k The sum of the effective likelihood functions of all observed targets, where B is a small nonnegative constant, the value of which is set according to the clutter density, v km (t) represents the distance between the state information of the k-th observation at time t and the state information of the m-th trajectory being tracked, S m (t) represents the area of ​​the m-th trajectory being tracked at time t, where T is the number of observations, m k The number of trajectories;

[0051] Step 204: Calculate the equivalent state observation of the trajectory at the current moment by weighting the observations corresponding to each trajectory in the previous frame of data according to the association probability:

[0052]

[0053] in, This is the equivalent state observation of the m-th trajectory being tracked at the current moment. The m-th trajectory being tracked is the m-th historical track. P km X is the correlation probability between the k-th observation and the m-th trajectory being tracked. k It is the kth observation.

[0054] In step 201, the current observation state information and the trajectory state information in the previous frame of data are respectively represented by vector a. k and b m Formal representation:

[0055] a k =[x k y k S k ]

[0056]

[0057] Among them, a k Let x be the state vector of the k-th observation in the current data frame. k y k and S k These represent the x-axis coordinates, y-axis coordinates, and area size of the k-th observation in the current data frame, respectively. m Let x be the state vector of the m-th historical tracking trajectory. m y m and S m These represent the x-axis coordinates, y-axis coordinates, and area dimensions of the m-th historical tracking trajectory, respectively. and These are the x-axis velocity and y-axis velocity of the m-th historical tracking trajectory, respectively.

[0058] In step 202, when calculating the distance between the currently observed state information and the trajectory state information in the previous frame of data, the position dimension or area dimension is not overly represented in the distance measurement of the state vector by calculating the corrected difference vector used to match the target. The formula for calculating the corrected difference vector used to match the target is as follows:

[0059]

[0060] Where d is the difference vector of state variables between time t+1 and time t, d' is the corrected difference vector used to match the target, and S t and S t+1 These are the area values ​​of the target at time t+1 and time t, respectively.

[0061] Step 3, which involves estimating the state of the tracked trajectory based on the state information of the matched trajectory, specifically includes the following steps:

[0062] Step 301: Use a Bayesian filtering method to estimate the position and velocity state. Specifically, use a Kalman filter based on a uniformly accelerated motion model to estimate the position and velocity state, and provide the corresponding variance of the estimate. Input the trajectory's state observation vector into the Bayesian filter to obtain the state estimation vector of the tracked trajectory. This yields the state information of the tracked target and the corresponding variance value. The state vector of the tracked trajectory is in the following form:

[0063]

[0064] Among them, X u (t) is the state vector of the u-th trajectory being tracked at time t, x u (t), y u (t) refers to the x-axis position and y-axis position of the u-th trajectory being tracked at time t, respectively. and Let x and y be the x-axis velocities and y-axis velocities of the u-th trajectory being tracked at time t, respectively. and Let x and x be the x-axis accelerations of the u-th trajectory being tracked at time t, respectively.

[0065] Step 302: Calculate the average area of ​​the tracked trajectory using a sliding window method.

[0066]

[0067] Where S(t) is the area size of the target estimated by the sliding window method, s(t-τ) is the area of ​​the target observed at time t-τ, and M is the width of the sliding window, with the value of M ranging from 4 to 6.

[0068] Step 4, the process of correcting the parameters of the clustering algorithm based on the state estimation results, specifically includes the following steps:

[0069] Step 401: Calculate the variance value of the position estimate from the variance matrix obtained from the state estimation. If the variance value exceeds a preset threshold, use this variance value to correct the parameters of the clustering algorithm. This step is only performed when the tracking variance of the uniformly accelerated motion model has converged.

[0070] D'=D-δ|P xy |

[0071] Where D' is the similarity threshold of the modified clustering algorithm, i.e., the distance threshold parameter of the modified target recognition algorithm, and D is the similarity threshold of the unmodified clustering algorithm, P xy To estimate the positional variance matrix in the variance matrix, δ is a correction coefficient;

[0072] Step 402: Re-perform target identification and filtering to avoid large errors in target identification due to improper parameter settings of the clustering algorithm.

[0073] In step 5, the tracking results of multiple radars are fused using a distributed fusion structure. Local tracks from each radar are fed into the data center for data fusion, specifically, the state estimation of local tracks is fused, and the current state information of the input track is correlated with the state information of historical tracking tracks in the data center.

[0074]

[0075] Where b is the weighting exponent in the fuzzy mean algorithm, representing the weight of each trajectory among the various fuzzy classes, and d um For the correlation distance, υ um Let be the membership degree between the u-th trajectory and the m-th historical tracking trajectory in the data center, and c be the total number of trajectories in the data center.

[0076] The formula for calculating the association distance is as follows:

[0077]

[0078] Where Δ1 and Δ2 are the data resolutions of the data center and the radar, respectively, and x1 and x2 are the current status information of the data center and the radar, respectively.

[0079] In step 5, the expressions for the state information x and variance matrix p obtained by data fusion of tracking results from multiple radars are as follows:

[0080]

[0081]

[0082] Where P1 and P2 are the variance matrices of the data center and radar, respectively.

[0083] Compared with the prior art, the present invention has the following advantages:

[0084] The proposed method for tracking maritime targets based on multiple marine radars in this invention improves the clustering algorithm parameters by introducing the variance of the tracking filter parameter estimation, thereby alleviating the difficulty of setting the target recognition algorithm parameters and increasing the system robustness.

[0085] Second, the maritime target tracking method based on multiple marine radars proposed in this invention avoids errors caused by erroneous observations of individual radars by using a distributed tracking system of multiple radars, thus improving the algorithm's adaptability to complex environments. Attached Figure Description

[0086] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0087] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0088] Example

[0089] A method for tracking maritime targets based on multiple marine radars, comprising the following steps:

[0090] Step 1: Use clustering algorithms to identify targets in the point cloud data acquired by radar, and obtain the location and size information of the identified targets;

[0091] Step 2: Based on the joint probabilistic data interconnection method, match the current observation with the existing historical tracking trajectory to obtain the status information of the matched trajectory;

[0092] Step 3: Estimate the state of the track being tracked based on the state information of the matched trajectory, and obtain the state estimation results, including the target's state information and the corresponding estimated variance value;

[0093] Step 4: Adjust the parameters of the clustering algorithm based on the state estimation results;

[0094] Step 5: Fusion of tracking results from multiple radars, that is, merging the state estimation results of each radar individually.

[0095] In step 1, the process of using a clustering algorithm to identify targets from the point cloud data acquired by the radar specifically includes the following steps:

[0096] Step 101: Transform the point cloud data acquired by the radar in the polar coordinate system into the rectangular coordinate system using coordinate transformation; the transformation formula is:

[0097] x = r n cosθ n +X n

[0098] y = r n sinθ n +Y n

[0099] Where, (θ n ,r n Let (x, y) be the position of the nth point cloud data in the polar coordinate system, and (x, y) be the position of the nth point cloud data in the world coordinate system. n ,Y n () represents the radar's position in the world coordinate system when observing the nth point cloud data.

[0100] Step 102: While processing the coordinate system transformation, compress the point cloud data. Compress the point cloud data using the average value along the equal angle and equal distance directions observed by the radar, and store it in a `radar_data` matrix. The size of the `radar_data` matrix is ​​5 times the number of compressed points in that frame. The first four rows of the `radar_data` matrix store the point cloud data in polar coordinates and the world coordinate system. Calculate the measurement error matrix R in the Cartesian coordinate system.

[0101] R = TΛT T

[0102] Where T is the transformation matrix and Λ is the radar measurement error covariance matrix;

[0103] The expression for the transformation matrix T is:

[0104]

[0105] Where (θ,r) represents the position in the polar coordinate system;

[0106] The expression for the radar measurement error covariance matrix Λ is:

[0107]

[0108] in, For radar radial observation variance, Variance of radar angle observations;

[0109] Step 103: Initialize the parameters of the target recognition algorithm based on prior environmental information;

[0110] Step 104: The clustering algorithm uses hierarchical clustering, which iteratively calculates similarity to achieve point clustering for target recognition. The similarity between point clusters is measured by the nearest distance between them. The formula for calculating the nearest distance between point clusters is:

[0111]

[0112] Among them, A i and B j Let a and b represent the set of the i-th point cluster and the set of the j-th point cluster, respectively, where a and b are any points in the i-th point cluster and the j-th point cluster, respectively.

[0113] Step 105: For each point cloud data, the last row of the radar_data matrix stores the category to which the point cloud data belongs. For point clusters belonging to the same observation, the position of the centroid is used as the observation position, that is, the average X and Y coordinates of all point cloud data in a group are used as the position of the point cluster. By traversing the radar_data matrix and accumulating the point cloud data of each category, the centroid of the point cloud data of each category is calculated to obtain the position of the observation, and the size of the observation is calculated. The formula for calculating the size is:

[0114]

[0115] Among them, S k (t) represents the area at time t for the k-th observation, δ θ δ is the angular resolution of the radar data. R C represents the radial resolution of radar data. k (t) represents the radial distance of the l-th point cloud data belonging to the k-th observation.

[0116] In step 2, the process of matching the current observation with existing historical tracking trajectories based on the joint probabilistic data interconnection method specifically includes the following steps:

[0117] The process of matching current observations with existing historical tracking trajectories based on the joint probabilistic data interconnection method includes the following steps:

[0118] Step 201: Calculate the distance between the current observed state information and the state information of each trajectory in the previous frame of data. The distance v between the current observed state information and the state information of each trajectory in the previous frame of data is... km The calculation formula is:

[0119]

[0120] Among them, [x k ,y k[x] represents the position of the k-th observation. m ,y m [S] represents the predicted position of the m-th historical tracking trajectory, which is the trajectory in the previous frame of data. k S is the area of ​​the kth observation. m S is the predicted area of ​​the m-th historical tracking trajectory. m (t) represents the area of ​​the m-th historical tracking trajectory at time t, λ is the weight constant, and N is the transformation matrix;

[0121] Step 202: Determine if the distance is less than the nearest distance threshold parameter. If so, there is a correlation probability between the observation and the tracking trajectory. Record the result using an acknowledgment matrix. The predicted state of the trajectory in the previous frame is obtained by second-order extrapolation.

[0122] X m (t|t-1)=Φ m (t-1)X m (t-1|t-1)

[0123] Among them, X m (t|t-1) represents the prediction of the state at time t for the m-th historical tracking trajectory based on the state estimate at time t-1, Φ m (t-1) is the state transition matrix of the m-th historical tracking trajectory at time t-1, using a second-order linear model, X. m (t-1|t-1) represents the state estimate of the m-th historical tracking trajectory at time t-1;

[0124] The expression for the confirmation matrix is:

[0125] R = [r] km ] q×w ,(k=1,2,...,q),(m=1,2,...,w)

[0126]

[0127] Where, r km To confirm the elements in the matrix, when r km =1 indicates that the k-th observation and the m-th historical tracking trajectory can be correlated; when r km = 0 indicates that the k-th observation and the m-th historical tracking trajectory are not correlated; q is the number of observations; w is the number of historical tracking trajectories.

[0128] Step 203: Obtain the association probability between each observation in the current data frame and each trajectory in the previous data frame by confirming the distance between the current observation's state information and the current observation's state information and the trajectory's state information in the previous frame of data;

[0129] Step 204: Calculate the equivalent state observation of the trajectory at the current moment by weighting the observations corresponding to each trajectory in the previous frame of data according to the association probability:

[0130]

[0131] in, This is the equivalent state observation of the m-th trajectory being tracked at the current moment. The m-th trajectory being tracked is the m-th historical track. P km X is the correlation probability between the k-th observation and the m-th trajectory being tracked. k It is the kth observation;

[0132] Tracking in progress refers to both historical tracking and trajectories that existed before the current moment. Here, the concept of joint probability data interconnection is used to calculate the correlation probability, and the method is as follows:

[0133]

[0134]

[0135]

[0136]

[0137] Among them, P km G refers to the probability that the k-th observation is associated with the m-th trajectory being tracked. The m-th trajectory being tracked is the m-th historical track. km A is a quantity proportional to the Gaussian likelihood function, which can be called the effective likelihood function. It is used to characterize the correlation between the k-th observation and the m-th trajectory being tracked. m A is the sum of the effective likelihood functions of all observations of the m-th trajectory being tracked. k The sum of the effective likelihood functions of all observed targets is given by B, a small non-negative constant. This value is generally set according to the clutter density; setting it to 0 can sometimes yield good results. km (t) represents the distance between the state information of the k-th observation at time t and the state information of the m-th trajectory being tracked, S m (t) represents the area of ​​the m-th trajectory being tracked at time t, where T is the number of observations, m k The number of trajectories.

[0138] The current observation state information and the trajectory state information in the previous frame of data are represented by vector a. k and b m Formal representation:

[0139] a k =[x k yk S k ]

[0140]

[0141] Among them, a k Let x be the state vector of the k-th observation in the current data frame. k y k and S k These represent the x-axis coordinates, y-axis coordinates, and area size of the k-th observation in the current data frame, respectively. m Let x be the state vector of the m-th historical tracking trajectory. m y m and S m These represent the x-axis coordinates, y-axis coordinates, and area dimensions of the m-th historical tracking trajectory, respectively. and These are the x-axis velocity and y-axis velocity of the m-th historical tracking trajectory, respectively.

[0142] When calculating the distance for target matching, it is necessary to specify x. i y i and S i The physical quantities being measured are different. By calculating the corrected difference vector used for matching targets, the position or area dimension is prevented from having an excessively large weight in the distance measurement of the state vector. The formula for calculating the corrected difference vector used for matching targets is as follows:

[0143]

[0144] Where d is the difference vector of state variables between time t+1 and time t, d' is the corrected difference vector used for matching the target, and S t and S t+1 These are the area values ​​of the target at time t+1 and time t, respectively.

[0145] In step 3, the process of estimating the state of the track being tracked based on the state information of the matched trajectory specifically includes the following steps:

[0146] Step 301: Use a Kalman filter based on a uniformly accelerated motion model to estimate the position and velocity, and obtain the variance of the corresponding estimate. The state vector of the tracked trajectory fed into the Bayesian filter is in the form of:

[0147]

[0148] Among them, X u (t) is the state vector of the u-th trajectory being tracked at time t, x u (t), y u(t) refers to the x-axis position and y-axis position of the u-th trajectory being tracked at time t, respectively. and Let x and y be the x-axis velocities and y-axis velocities of the u-th trajectory being tracked at time t, respectively. and Let x and x be the x-axis accelerations of the u-th trajectory being tracked at time t, respectively.

[0149] Step 302: Calculate the average area of ​​the tracked trajectory using a sliding window method.

[0150]

[0151] Where S(t) is the area size of the target estimated by the sliding window method, s(t-τ) is the target area observed at time t-τ, and M is the sliding window width, with the value of M ranging from 4 to 6.

[0152] In step 4, the process of correcting the parameters of the clustering algorithm based on the state estimation results specifically includes the following steps:

[0153] Step 401: Calculate the variance value of the position estimate from the variance matrix obtained from the state estimation. If the variance value exceeds a preset threshold, use this variance value to correct the parameters of the target recognition algorithm. This step is only performed when the tracking variance of the system has converged.

[0154] D'=D-δ|P xy |

[0155] Where D' is the similarity threshold of the modified clustering algorithm, i.e., the distance threshold parameter of the modified target recognition algorithm, and D is the similarity threshold of the unmodified clustering algorithm, P xy To estimate the positional variance matrix in the variance matrix, δ is a correction coefficient;

[0156] Step 402: Re-perform target recognition and filtering to avoid large target recognition errors due to improper parameter settings.

[0157] In step 5, the process of fusing the tracking results from multiple radars is specifically as follows:

[0158] The tracking results from multiple radars are fused using a distributed fusion structure. The local tracks sent from each radar to the data center participate in the data fusion, which essentially involves fusing the state estimates of the local tracks. In this system architecture, the radars essentially perform preliminary filtering and noise removal functions, only needing to transmit the target's current state. Specifically, the current state information of the input tracks is correlated with the state information in the data center.

[0159]

[0160] Where b is the weighting exponent in the fuzzy mean algorithm, representing the weight of each trajectory among the various fuzzy classes, and d um For the correlation distance, υ um Let be the membership degree between the u-th trajectory and the m-th trajectory in the data center, and c be the total number of trajectories in the data center.

[0161] The formula for calculating the association distance is:

[0162]

[0163] Where Δ1 and Δ2 are the data resolutions of the data center and the radar, respectively, and x1 and x2 are the current status information of the data center and the radar, respectively.

[0164] The fused state information x and variance matrix p are:

[0165]

[0166]

[0167] Where P1 and P2 are the variance matrices of the data center and radar, respectively.

[0168] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for tracking maritime targets based on multiple marine radars, characterized in that, The method includes the following steps: Step 1: Use clustering algorithms to identify targets in the point cloud data acquired by radar, and obtain the location and size information of the identified targets; Step 2: Based on the joint probabilistic data interconnection method, match the current observation with the existing historical tracking trajectory to obtain the status information of the matched trajectory; Step 3: Estimate the state of the track being tracked based on the state information of the matched trajectory, and obtain the state estimation results, including the target's state information and the corresponding estimated variance value; Step 4: Adjust the parameters of the clustering algorithm based on the state estimation results; Step 5: Fusion of tracking results from multiple radars, that is, fusion of the state estimation results from each radar individually; Step 4, the process of correcting the parameters of the clustering algorithm based on the state estimation results, specifically includes the following steps: Step 401: Calculate the variance value of the position estimate from the variance matrix obtained from the state estimation. If the variance value exceeds a preset threshold, use this variance value to correct the parameters of the clustering algorithm. This step is only performed when the tracking variance of the uniformly accelerated motion model has converged. D'=D-δ|P xy | Where D' is the similarity threshold of the modified clustering algorithm, i.e., the distance threshold parameter of the modified target recognition algorithm, and D is the similarity threshold of the unmodified clustering algorithm, P xy To estimate the positional variance matrix in the variance matrix, δ is a correction coefficient; Step 402: Re-perform target identification and filtering to avoid large errors in target identification due to improper parameter settings of the clustering algorithm.

2. The method for tracking maritime targets based on multiple marine radars according to claim 1, characterized in that, In step 1, the clustering algorithm identifies targets by determining the minimum distance between two categories. The process of using the clustering algorithm to identify targets from the point cloud data acquired by radar specifically includes the following steps: Step 101: Transform the point cloud data acquired by the radar in the polar coordinate system into the world coordinate system using coordinate transformation. The formula for coordinate transformation is: x=r n cosθ n +X n y=r n sinθ n +Y n Where, (θ n ,r n Let (x, y) be the position of the nth point cloud data in the polar coordinate system, and (x, y) be the position of the nth point cloud data in the world coordinate system. n ,Y n () represents the radar's position in the world coordinate system when observing the nth point cloud data. Step 102: While processing the coordinate system transformation, compress the point cloud data. Compress the point cloud data using the average value along the equal angle and equal distance directions observed by the radar, and store it in a radar_data matrix. The first four rows of the radar_data matrix store the point cloud data in polar coordinates and the world coordinate system. Calculate the measurement error matrix R in the world coordinate system. The formula for calculating the measurement error matrix R is: R=TΛT T Where T is the transformation matrix and Λ is the radar measurement error covariance matrix; The expression for the transformation matrix T is: Where (θ,r) represents the position in the polar coordinate system; The expression for the radar measurement error covariance matrix Λ is: in, For radar radial observation variance, Variance of radar angle observations; Step 103: Initialize the parameters of the target recognition algorithm based on the prior environmental information. The prior environmental information is the nearest distance threshold parameter set according to the density of the target. When the nearest distance between two point clusters is lower than the nearest distance threshold parameter, they are considered to belong to the same observation. Step 104: Point clustering is achieved by iteratively calculating the similarity between point clusters to realize target recognition. A point cluster is a set of point cloud data that is considered to be of the same class during the clustering process. The similarity between point clusters is measured by the nearest distance between point clusters. The formula for calculating the nearest distance between point clusters is: Among them, A i and B j Let a and b represent the set of the i-th point cluster and the set of the j-th point cluster, respectively, where a and b are any points in the i-th point cluster and the j-th point cluster, respectively. Step 105: For each point cloud data, the last row of the radar_data matrix stores the category to which the point cloud data belongs. For point clusters belonging to the same observation, the position of the centroid is used as the observation position, that is, the average X and Y coordinates of all point cloud data in a group are used as the position of the point cluster. By traversing the radar_data matrix and accumulating the point cloud data of each category, the centroid of the point cloud data of each category is calculated to obtain the position of the observation, and the size of the observation is calculated. The formula for calculating the size is: Among them, S k (t) represents the area at time t for the k-th observation, δ θ δ is the angular resolution of the radar data. R C represents the radial resolution of radar data. i (t) represents the radial distance of the i-th point cloud data belonging to the k-th observation.

3. The method for tracking maritime targets based on multiple marine radars according to claim 2, characterized in that, Step 2, which involves matching the current observation with existing historical tracking trajectories based on the joint probabilistic data interconnection method, specifically includes the following steps: Step 201: Calculate the distance between the current observed state information and the state information of each trajectory in the previous frame of data. Each trajectory in the previous frame of data is the historical tracking trajectory. The distance v between the current observed state information and the state information of each trajectory in the previous frame of data is... km The calculation formula is: Among them, [x k ,y k [x] represents the position of the k-th observation. m ,y m [S] represents the predicted location of the m-th historical tracking trajectory. k S is the area of ​​the kth observation. m S is the predicted area of ​​the m-th historical tracking trajectory. m (t) represents the area of ​​the m-th historical tracking trajectory at time t, λ is the weight constant, and N is the transformation matrix; Step 202: Determine if the distance is less than the nearest distance threshold parameter. If so, there is a correlation probability between the observation and the tracking trajectory. Record the result using an acknowledgment matrix. The predicted state of the trajectory in the previous frame is obtained by second-order extrapolation. X m (t|t-1)=Φ m (t-1)X m (t-1|t-1) Among them, X m (t|t-1) represents the prediction of the state at time t for the m-th historical tracking trajectory based on the state estimate at time t-1, Φ m (t-1) is the state transition matrix of the m-th historical tracking trajectory at time t-1, using a second-order linear model, X. m (t-1|t-1) represents the state estimate of the m-th historical tracking trajectory at time t-1; The expression for the confirmation matrix is: R = [r] km ] q×w Where k = 1, 2, ..., q, m = 1, 2, ..., w Where, r km To confirm the elements in the matrix, when r km =1 indicates that the k-th observation and the m-th historical tracking trajectory can be correlated; when r km = 0 indicates that the k-th observation and the m-th historical tracking trajectory are not correlated; q is the number of observations; w is the number of historical tracking trajectories. Step 203: By confirming the distance between the current observation's state information and the trajectory's state information in the previous frame, obtain the association probability between each observation in the current data frame and each trajectory in the previous frame. The association probability is calculated using the joint probability data interconnection concept, as follows: Among them, P km G refers to the probability that the k-th observation is associated with the m-th trajectory being tracked. The m-th trajectory being tracked is the m-th historical track. km A is a quantity proportional to the Gaussian likelihood function, called the effective likelihood function, used to characterize the correlation between the k-th observation and the m-th trajectory being tracked. m A is the sum of the effective likelihood functions of all observations of the m-th trajectory being tracked. k The sum of the effective likelihood functions of all observed targets, where B is a small nonnegative constant, the value of which is set according to the clutter density, v km (t) represents the distance between the state information of the k-th observation at time t and the state information of the m-th trajectory being tracked, S m (t) represents the area of ​​the m-th trajectory being tracked at time t, where T is the number of observations, m k The number of trajectories; Step 204: Calculate the equivalent state observation of the trajectory at the current moment by weighting the observations corresponding to each trajectory in the previous frame of data according to the association probability: in, This is the equivalent state observation of the m-th trajectory being tracked at the current moment. The m-th trajectory being tracked is the m-th historical track. P km X is the correlation probability between the k-th observation and the m-th trajectory being tracked. k It is the kth observation.

4. The method for tracking maritime targets based on multiple marine radars according to claim 3, characterized in that, In step 201, the current observation state information and the trajectory state information in the previous frame of data are respectively represented by vector a. k and b m Formal representation: to k =[x k and k S k ] b m =[x m y m S m ] Among them, a k Let x be the state vector of the k-th observation in the current data frame. k y k and S k These represent the x-axis coordinates, y-axis coordinates, and area size of the k-th observation in the current data frame, respectively. m Let x be the state vector of the m-th historical tracking trajectory. m y m and S m These represent the x-axis coordinates, y-axis coordinates, and area dimensions of the m-th historical tracking trajectory, respectively.

5. A method for tracking maritime targets based on multiple marine radars according to claim 4, characterized in that, In step 201, when calculating the distance between the currently observed state information and the trajectory state information in the previous frame of data, the position dimension or area dimension is not overly represented in the distance measurement of the state vector by calculating the corrected difference vector used to match the target. The formula for calculating the corrected difference vector used to match the target is as follows: Where d is the difference vector between the state vectors at time t+1 and time t, d' is the corrected difference vector used to match the target, and S t and S t+1 These are the area values ​​of the target at time t+1 and time t, respectively.

6. The method for tracking maritime targets based on multiple marine radars according to claim 1, characterized in that, Step 3, which involves estimating the state of the tracked trajectory based on the state information of the matched trajectory, specifically includes the following steps: Step 301: Use a Bayesian filtering method to estimate the position and velocity state. Specifically, use a Kalman filter based on a uniformly accelerated motion model to estimate the position and velocity state, and provide the corresponding variance of the estimate. Input the trajectory's state observation vector into the Bayesian filter to obtain the state estimation vector of the tracked trajectory. This yields the state information of the tracked target and the corresponding variance value. The state vector of the tracked trajectory is in the following form: Among them, X u (t) is the state vector of the u-th trajectory being tracked at time t, x u (t), y u (t) refers to the x-axis position and y-axis position of the u-th trajectory being tracked at time t, respectively. and Let x and y be the x-axis velocities and y-axis velocities of the u-th trajectory being tracked at time t, respectively. and These are the x-axis acceleration and y-axis acceleration of the u-th trajectory being tracked at time t, respectively. Step 302: Calculate the average area of ​​the tracked trajectory using a sliding window method. Where S(t) is the area size of the target estimated by the sliding window method, s(t-τ) is the area of ​​the target observed at time t-τ, and M is the width of the sliding window, with the value of M ranging from 4 to 6.

7. A method for tracking maritime targets based on multiple marine radars according to claim 1, characterized in that, In step 5, the tracking results of multiple radars are fused using a distributed fusion structure. Local tracks from each radar are fed into the data center for data fusion, specifically, the state estimation of local tracks is fused, and the current state information of the input track is correlated with the state information of historical tracking tracks in the data center. Where b is the weighting exponent in the fuzzy mean algorithm, representing the weight of each trajectory among the various fuzzy classes, and d um For the correlation distance, υ um Let be the membership degree between the u-th trajectory and the m-th historical tracking trajectory in the data center, and c be the total number of trajectories in the data center.

8. A method for tracking maritime targets based on multiple marine radars according to claim 7, characterized in that, The formula for calculating the association distance is as follows: Where Δ1 and Δ2 are the data resolutions of the data center and the radar, respectively, and x1 and x2 are the current status information of the data center and the radar, respectively.

9. A method for tracking maritime targets based on multiple marine radars according to claim 8, characterized in that, In step 5, the expressions for the state information x and variance matrix p obtained by data fusion of tracking results from multiple radars are as follows: Where P1 and P2 are the variance matrices of the data center and radar, respectively.

Images