Group Target Tracking Method Based on DBSCAN Clustering

Abstract

A group target tracking method based on DBSCAN clustering, which relates to the field of group tracking technology. The present invention is to solve the problem that the clustering results of existing group target tracking methods lack stability, resulting in low tracking accuracy. The present invention includes: obtaining a spatial target data set; using the DBSCAN algorithm on the spatial target data set to perform initial clustering on the spatial targets to obtain spatial target groups; obtaining the contours of each spatial target group, and using the contours of each spatial target group to obtain the group centers of each spatial target group; obtaining the position information of the center of each spatial target group, and using a filter to perform position tracking on the center of each spatial target group; obtaining the measurement values of each target in the spatial target group, and obtaining the boundary point set of each spatial target group according to the measurement values of each target; determining whether the contour of each spatial target group has mutated, and if there is no mutation, using a filter to track the contour of each spatial target group. The present invention is used for tracking group targets.

Description

Technical Field

[0001] The present invention relates to the technical field of group tracking, and particularly to a group target tracking method based on DBSCAN clustering. Background Art

[0002] Spatial target groups have characteristics such as a small airspace distribution range, high density, consistent movement directions, and low relative movement speeds between targets. Therefore, the targets will remain in a "grouped" state for a long time. Group target tracking has important application values in fields such as ground or sea target monitoring, multi-target formation movement, crowd or herd tracking, etc. However, since group targets are often in a densely distributed state, it is difficult to track the targets within the group separately. Therefore, tracking the group as a whole has become an important research topic in this field.

[0003] Currently, the tracking of spatial target groups mainly adopts spatial target group segmentation methods. The spatial target group segmentation methods mainly include the distance segmentation method and the improved cyclic threshold method, both of which achieve clustering by setting distance thresholds. Their principles are simple and easy to implement. However, the high density of spatial targets easily causes fluctuations in the number of members in the segmented groups at each moment, resulting in a lack of stability in the clustering results, and further reducing the tracking accuracy of group targets. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem that the clustering results of existing group target tracking methods lack stability, resulting in low tracking accuracy, and a group target tracking method based on DBSCAN clustering is proposed.

[0005] The specific process of the group target tracking method based on DBSCAN clustering is as follows:

[0006] Step 1: Obtain a spatial target data set;

[0007] The spatial target data set includes: the position coordinates of each spatial target and the speed in the coordinate direction;

[0008] Step 2: Use the DBSCAN algorithm to perform initial clustering on the spatial targets using the spatial target data set to obtain spatial target groups, and obtain the contours of each spatial target group. Use the contours of each spatial target group to obtain the group centers of each spatial target group;

[0009] Step 3: Model the spatial target group contours obtained in Step 2 using a star-convex random hypersurface to obtain a spatial target group contour model;

[0010] Step 4: Obtain the group center position information of each spatial target group obtained in Step 2, and use a filter to track the positions of each spatial target group center;

[0011] Step 5. Track the contour models of each spatial target group obtained in Step 3:

[0012] Step 5-1. Obtain the measurement values of each target in the spatial target group, and obtain the boundary point set of each spatial target group according to the measurement values of each target;

[0013] Step 5-2. Determine whether the contour of each spatial target group has mutated according to the boundary point set of each spatial target group. If a mutation occurs, re-execute Step 5-1. If no mutation occurs, use a filter to track the contour of each spatial target group.

[0014] The beneficial effects of the present invention are as follows:

[0015] The present invention first performs grouping processing on dense group targets through the DBSCAN clustering method, then tracks the group center and group contour of each group respectively, and finally obtains the track tracking results of each group. The DBSCAN clustering method adopted by the present invention is based on density clustering to group spatial targets, is applicable to group targets of various shapes, and takes the fluctuations within the group into account in the calculation of the group center. Compared with traditional algorithms, the present invention can be more adaptable to the influence of clutter, improve the stability of the grouping result, and improve the accuracy of group target tracking. The present invention utilizes the measurement information of group targets, extracts the boundary of the true shape from the measurement as the shape prior, adaptively models the group target contour using the center contour distance, constructs a shape parameter mutation detection quantity, and enables the shape parameter estimation to converge rapidly when a shape mutation occurs. Description of the Drawings

[0016] Figure 1 is the flow chart of the present invention;

[0017] Figure 2 is the position and shape diagram of the group target in the embodiment;

[0018] Figure 3 is the initial clustering result in the embodiment;

[0019] Figure 4 is the comparison diagram of the true shape and estimated shape of the group target (1-5 s) in the embodiment;

[0020] Figure 5 is the comparison diagram of the true shape and estimated shape of the group target (17-25 s) in the embodiment;

[0021] Figure 6 is the OSPA distance curve diagram of the target state estimation in the embodiment;

[0022] Figure 7 is the quasi-Jaccard distance curve diagram of the target contour estimation in the embodiment. Detailed Embodiment

[0023] Embodiment 1: The specific process of the group target tracking method based on DBSCAN clustering is as follows:

[0024] Step 1: Obtain the spatial target dataset;

[0025] The spatial target dataset includes: the initial states of each spatial target; the initial states include: position coordinates, velocities in each direction;

[0026] Step 2: Use the DBSCAN algorithm to perform initial clustering on the spatial targets using the dataset in the spatial target data, obtain the spatial target groups, and obtain the contour of each spatial target group. Use the contour of each spatial target group to obtain the group center of each spatial target group;

[0027] The group center of the spatial target group is the geometric center of the spatial target group contour;

[0028] Step 3: Use the star-convex random hypersurface to model the spatial target group contour obtained in Step 2 to obtain the spatial target group contour model;

[0029] Step 4: Obtain the group center position information of each spatial target group obtained in Step 2, and use a filter to perform position tracking on each spatial target group center;

[0030] Step 5: Track the contour model of each spatial target group obtained in Step 3.

[0031] Embodiment 2: The use of the DBSCAN algorithm to perform initial clustering on the spatial targets using the dataset in the spatial target dataset in Step 2 to obtain the spatial target groups includes the following steps:

[0032] Step 2-1: Start from any unvisited spatial target and use this spatial target as the center. Use the preset Eps (radius) and MinPts (minimum number of samples) to determine whether the spatial target is a core object. If it is a core object, create a new class C and include the spatial targets in its neighborhood in class C; if the spatial target is not a core object, re-visit the unvisited spatial targets until a core object spatial target is found, create a new class C, and include the spatial targets in its neighborhood in class C;

[0033] Step 2-2: Look for other unvisited spatial targets in C. If the spatial target is a core object, include all the spaces in its neighborhood in class C;

[0034] Step 2-3: Repeat Step 2-2 until no new spatial targets are added to class C;

[0035] Step 2-4: Repeat Steps 2-1 to 2-3 until all spatial targets are classified to obtain the spatial target groups;

[0036] The definitions involved in this embodiment are as follows:

[0037] Definition 1 (core object): If there are more than a certain number (≥MinPts) of samples within a circle with a radius of Eps for a sample p, then the sample p is called a core object.

[0038] Definition 2 (neighborhood ε): The points within the neighborhood are defined as N Eps (p) = {q ∈ D, dist(p, q) ≤ Eps}, where D represents all objects in the dataset, and dist(p, q) represents the distance between sample p and sample q.

[0039] The distance between the sample p and the sample q is obtained in the following ways: Euclidean distance, Minkowski distance, Manhattan distance, Chebyshev distance, etc.;

[0040] Definition 3 (directly density-reachable): In the dataset D, if an object q is within the Eps neighborhood of another object p and the object p is a core object, then it is said that the object q is directly density-reachable from the object p.

[0041] Definition 4 (density-reachable): In the dataset D, there exist p1, p2, p3, …, p n , 1 ≤ i ≤ n, for p i ∈ D, and p i is directly density-reachable from p i-1 , then the point p n is density-reachable from p1.

[0042] Definition 5 (density-connected): If there exists an object o such that both the objects p and q are density-reachable from o, then it is said that the objects p and q are density-connected.

[0043] Specific Embodiment 3: Tracking the contour of each spatial target group obtained in Step 3 in Step 5 specifically includes the following steps:

[0044] Step 5-1: Obtain the measurement values of each target in the spatial target group, and obtain the boundary point set of each target group according to the measurement values of each target;

[0045] Step 5-1-1: Obtain the target measurement values of a certain spatial target in each spatial target group at k moments

[0046]

[0047]

[0048] e(φ k,l ) = [cosφ k,lsinφ k,l T

[0049] where s k,l is the scaling factor, whose value follows a Gaussian distribution, R(φ k,l ) is the Fourier series expansion, φ k,l represents the azimuth angle of the l-th target measurement value of the group target at time k relative to the group centroid m k , N F is the preset Fourier order, is the shape parameter at time k, v k,l is the measurement noise, z k,l is the l-th target measurement value at time k, N z,k is the total number of target measurement values at time k, e(φ k,l ) is an intermediate variable;

[0050] Step 5-1-2: Divide the target measurement values obtained in Step 5-1-1 into N subsets according to the angle θ (θ = 2π / N). Taking the Euclidean distance from the measurement value to the center of the spatial target group as the criterion, select the measurement point with the largest distance in the subset as the target boundary point, and obtain the target group boundary point set:

[0051]

[0052]

[0053] where j = 1, ···, N is the label of the subset, N is the total number of subsets, is the group measurement center, i = 1, ···, N j is the label of the measurement value in the j-th subset, z k,j,i is the i-th measurement sequence value in the j-th subset, N j is the total number of target measurement values in the j-th subset, z k,j is the target measurement value of the j-th subset at time k, d(z k,j,i ) is the distance from the i-th target measurement value in the j-th subset to the group measurement center, d(z k,j ) is the maximum value of d(z k,j,i );

[0054] Step 5-2: Judge whether the contour of each spatial target group has mutated according to each target group boundary point set. If a mutation occurs, re-execute Step 5-1. If no mutation occurs, use a filter to track the contour of each spatial target group, including the following steps:

[0055] ​Step 521: Construct a sequence f(n) in the preset order using the obtained target boundary point set in Step 51, and obtain the Fourier coefficients of the radial function, i.e., the shape parameter estimation of the spatial target group;

[0056] Step 522: Construct a detection statistic using the shape parameter of the spatial target group obtained in Step 521 to determine whether the shape of the spatial target group has mutated. If a mutation occurs, re - execute Step 51; if no mutation occurs, use a filter to track the contour of the spatial target group;

[0057] The process of constructing a detection statistic using the shape parameter of the spatial target group obtained in Step 521 to determine whether the shape of the spatial target group has mutated includes the following steps:

[0058] Step 5221: Set a sliding window with a time length of H, and construct a detection statistic using the shape parameter estimation of the spatial target group obtained in Step 521

[0059]

[0060]

[0061] where, represents the shape parameter estimation of the spatial target group at time k, follows a χ 2 distribution with degrees of freedom H·n F , where n F is the dimension of the shape parameter , is the estimation error of the group target shape parameter, following a normal distribution is 's covariance, follows a χ 2 distribution (chi - square distribution), a ∈ [k - H + 1, k] is the time when the shape parameter estimation of the spatial target group is obtained, is the shape parameter estimation of the spatial target group at time a;

[0062] Step 5222: Set a detection threshold. If the detection threshold is less than the detection statistic, it means that the shape of the spatial target group at time k has mutated, and re - execute Steps 51 to 52 until the detection threshold is greater than or equal to the detection statistic. If the detection threshold is greater than or equal to the detection statistic, it means that the shape of the spatial target group has not mutated, and use a filter to track the shape contour of the spatial target group;

[0063] The detection threshold is obtained in the following way:

[0064]

[0065] Among them, p FA represents the false alarm probability, and H·n F is the degree of freedom of, and n F is the dimension of.

[0066] In this embodiment, the random hypersurface uses a radial function to describe the irregular shape contour. When the prior shape is unknown, the random hypersurface assumes that the prior shape of the target is a circle and approximates the true shape contour through a recursive filtering method. However, the actual situation is much more complex than a circle, and the size of the circle is unknown. When the difference between the true target shape and the assumed circle is large, the estimation of the target shape parameters is inaccurate. The present invention uses the random hypersurface as the basic model of the group diffusion form. On this basis, considering the use of group target measurement information, the boundary of the true shape is extracted from the measurement as the shape prior, and the central contour distance is used to study the adaptive modeling of the group target contour, and a more reasonable prior value is assigned to the radial function. Then, a shape parameter mutation detection quantity is constructed to enable the shape parameter estimation to converge rapidly when a shape mutation occurs.

[0067] Example: The following simulation experiment is used to verify the beneficial effects of the present invention:

[0068] A simulation scenario is set in a two-dimensional plane, and the observation area is [-500,500]m × [-500,500]m. It is assumed that all group targets move in a uniform straight line independently, and the state transition equation is

[0069] x k = F k x k-1 + w k (3-1)

[0070] Among them

[0071]

[0072]

[0073] F k represents the group target state transition matrix, including two parts: the motion state transition and the shape parameter transition. w k is the process noise with a mean of 0 and a covariance of Q k , and σ = 0.2 is the process noise standard deviation; the sampling period T = 1s, 02 is the second-order zero matrix, and I nF is the n F -order identity matrix. In order to describe the detailed information of the group target diffusion form, n F is taken as 9. x kDenote the state vector of the cluster center at time k, which includes 4D motion parameters and 9D shape parameters. The number of measurements generated by each cluster target follows a Poisson distribution with a mean of 50. The measurement noise is zero-mean Gaussian white noise with a covariance of Σ v = diag{0.2 2 , 0.2 2}. The number of clutter follows a Poisson distribution with a mean of 20. The scaling factor s in the star-convex RHM adaptive shape modeling algorithm k,l is assumed to follow a Gaussian distribution with a mean of 0.7 and a variance of 0.08. The pruning threshold T of the exponential mixture term of the GLMB filter is set to 1e-5, the merging threshold U is set to 4, and the maximum number of exponential terms is J max = 100, is an intermediate variable, and the tracking flow chart is as shown in Figure 1 .

[0074] In the experiment, 24 point targets are simulated, which are located in two different clusters respectively. Each cluster consists of 12 point targets. The initial positions of the cluster centers of the two clusters are respectively:

[0075] x1 = [-450m; 300m; 20m / s; -10m / s]

[0076] x2 = [-450m; -300m; 20m / s; 10m / s]

[0077] Assume that the lateral length of the shape of the cluster target is 8m from 1s to 19s and 16m from 20s to 40s, and the longitudinal width of the cluster target remains unchanged at 4m, that is, the contour of the cluster target changes suddenly at the 20th second of tracking. The true shape of the simulated cluster target is as shown in Figure 2 , where the black dots are the true targets and the gray cross frames are the true diffusion forms of the cluster targets.

[0078] Perform initial clustering on the 24 point targets. The clustering results are as shown in Figure 3 . It can be seen that the DBSCAN clustering algorithm can correctly group the targets for subsequent tracking of the cluster center.

[0079] The shape prior of the circular prior random hypersurface algorithm is set to a circle with a radius of 10. The present invention extracts the shape boundary from the measurements of the cluster target as the prior shape, performs a Fourier transform on the extracted shape boundary point set to obtain the Fourier coefficients of the radial function, and takes the adaptive contour parameter θ = 200.

[0080] In the simulation experiment settings, take the sliding time window length H = 3, and the false alarm probability p FA= 0.05. To compare the performance of the centroid and shape tracking and estimation of the group targets between the circular prior random hypersurface algorithm and the star-convex random hypersurface adaptive shape algorithm described in the third specific embodiment of the present invention, 100 Monte Carlo simulations were conducted for this experiment.

[0081] Figure 4 and Figure 5 The solid line and the dashed line in [figure] respectively represent the simulation results of the star-convex random hypersurface adaptive modeling and the circular prior contour algorithm. It can be clearly seen that the algorithm proposed in the present invention has obvious advantages in the initial estimation of the target shape when the target appears. This is because the circular prior random hypersurface algorithm sets the shape of the group target as a circle with a fixed size. Once the prior is unknown, the initial tracking performance is unreliable. The star-convex random hypersurface adaptive contour algorithm proposed in the present invention extracts the shape contour of the group target according to the measurements of the group target, which is more in line with the actual situation compared with the circular prior, and the Fourier coefficients of the radial function are relatively more accurate, providing a more reasonable initial value for the subsequent tracking and estimation of the filter, which is conducive to the rapid convergence of the shape filtering estimation of the group target. This is of great significance for quickly identifying and attacking enemy targets on the actual battlefield. In addition, when the shape of the group target undergoes a sudden change, the star-convex random hypersurface algorithm can still quickly respond.

[0082] Figure 6 and Figure 7 are the estimation results of the OSPA distance and shape of the group targets by the circular prior random hypersurface and the star-convex random hypersurface adaptive shape algorithms under 100 Monte Carlo simulation experiments. As can be seen from Figure 6 it, the OSPA distances of the two algorithms for the group targets are generally the same, indicating that there is no obvious difference in the performance of the two algorithms in the estimation of the position and quantity of the group targets. However, the present invention is significantly superior to the traditional algorithm at the initial moment and has a faster convergence speed. Figure 7 It reflects that the star-convex random hypersurface adaptive contour algorithm proposed in the present invention has better performance in the shape estimation of the group target compared with the circular prior random hypersurface. Its pseudo-Jaccard distance is much smaller than that of the circular prior random hypersurface algorithm. At 20 s, the contour of the group target undergoes a sudden change. The dashed-line adaptive algorithm is closer to the true shape of the target after the mutation than the solid-line circular prior. The pseudo-Jaccard distance increases rapidly at the mutation moment, and the shape estimation performance deteriorates rapidly. However, it can still be found that the present invention can obtain the estimation after the target shape mutation in a short time. Compared with the RHM algorithm of the circular prior contour, its shape estimation is more robust and the subsequent shape estimation performance is also better.

Claims

1. A group target tracking method based on DBSCAN clustering, characterized in that The specific process of the method is as follows: Step 1: Obtain a spatial target data set; The spatial target data set includes: the position coordinates of each spatial target and the velocity in the coordinate direction; Step 2: Use the DBSCAN algorithm on the spatial target data set to initially cluster the spatial targets to obtain spatial target clusters, and obtain the contour of each spatial target cluster. Use the contour of each spatial target cluster to obtain the cluster center of each spatial target cluster; Step 3: Use a star-convex random hypersurface to model the contour of the spatial target clusters obtained in Step 2 to obtain a spatial target cluster contour model; Step 4: Obtain the position information of the cluster center of each spatial target cluster obtained in Step 2, and use a filter to track the position of each spatial target cluster center; Step 5: Track the contour model of each spatial target cluster obtained in Step 3: Step 5-1: Obtain the measurement value of each target in the spatial target cluster, and obtain the boundary point set of each spatial target cluster according to the measurement value of each target; Step 5-2: Determine whether the contour of each spatial target cluster has mutated according to the boundary point set of each spatial target cluster. If a mutation occurs, re-execute Step 5-1. If no mutation occurs, use a filter to track the contour of each spatial target cluster, including the following steps: Step 5-2-1: Use the target boundary point set obtained in Step 5-1 to construct a sequence f(n) in a preset order, and use f(n) to obtain the Fourier coefficients of the radial function, that is, the shape parameter estimation of the spatial target cluster; Step 5-2-2: Use the shape parameters of the spatial target cluster obtained in Step 5-2-1 to construct a detection statistic to determine whether the shape of the spatial target cluster has mutated. If a mutation occurs, re-execute Step 5-1. If no mutation occurs, use a filter to track the contour of the spatial target cluster. Specifically: Step Five Two Two One: Set a sliding window with a time length of H, and construct a detection statistic using the shape parameter estimation of the spatial target group obtained in Step Five Two One Among them, represents the estimation of the shape parameter of the space target group at time k, follows a χ 2 distribution with degrees of freedom H·n F , where n F is the shape parameter dimension, is the estimation error of the group target shape parameter, following a normal distribution is covariance, follows a χ 2 distribution, a ∈ [k - H + 1, k] is the time to obtain the estimation of the space target group shape parameter, is the estimation of the space target group shape parameter at time a; Step 5-2-2-2: Set a detection threshold. If the detection threshold is less than the detection statistic, the shape of the spatial target cluster at time k has mutated, and re-execute Steps 5-1 to 5-2 until the detection threshold is greater than or equal to the detection statistic. If the detection threshold is greater than or equal to the detection statistic, the shape of the spatial target cluster has not mutated, and use a filter to track the shape contour of the spatial target cluster.

2. The method for group target tracking based on DBSCAN clustering according to claim 1, wherein: The initial clustering of spatial targets using the DBSCAN algorithm on the data set in the spatial target data in Step 2 to obtain spatial target clusters includes the following steps: Step 2-1: Take any unvisited spatial target as the center, and use a preset radius Eps and the minimum number of samples MinPts to determine whether the spatial target is a core object. If it is a core object, create a new class C and include the spatial targets in the neighborhood of the core object into class C; Step 2-2: Find other unvisited spatial targets in C and determine whether the unvisited spatial targets are core objects. If they are core objects, include all the spatial targets in their neighborhoods into class C; Step 2-3: Repeat Step 2-2 until no new spatial targets are added to class C; Step 2-4: Repeat Steps 2-1 to 2-3 until all spatial targets are classified to obtain spatial target clusters.

3. The method for group target tracking based on DBSCAN clustering according to claim 2, characterized in that: The nuclear object neighborhood is obtained through the following formula: N Eps (p) = {q ∈ D, dist(p, q) ≤ Eps}; Among them, N Eps (p) is the neighborhood of the spatial target p, D is the data set containing all spatial targets, and dist(p, q) is the distance between the spatial target p and the spatial target q.

4. The method for tracking group targets based on DBSCAN clustering according to claim 2 or 3, characterized in that: In step 5-1, obtaining the measurement values of each target in the spatial target group, and obtaining the boundary point set of each spatial target group according to the measurement values of each target, includes the following steps: Step Five One: Obtain the measurement values of a certain space target in each space target group at k moments where z k,l is the l-th target measurement value at time k, and N z,k is the total number of target measurement values at time k; Step 5-1-2: Divide the measurement values of all spatial targets in each spatial target group obtained in step 5-1-1 into N subsets according to the angle θ, and use the spatial target with the largest Euclidean distance from the measurement value to the center of the spatial target group in each subset as the boundary point of the target group, and obtain the boundary point set of the target group.

5. The group target tracking method based on DBSCAN clustering according to claim 4, wherein: e(φ k,l ) = [cosφ k,l sinφ k,l T ​ where s k,l is the scaling factor, R(φ k,l ) is the Fourier series expansion, φ k,l represents the azimuth angle of the l-th target measurement value in the spatial target group at time k relative to the group centroid m k , e(φ k,l ) is an intermediate variable, v k,l is the measurement noise, is the shape parameter at time k.

6. The group target tracking method based on DBSCAN clustering according to claim 5, wherein: where N F is a preset Fourier order.

7. The method for tracking group targets based on DBSCAN clustering according to claim 6, wherein: The boundary point set of the target group in step 5-1-2 is specifically: where \(j = 1,\cdots,N\) is the label of the subset, and \(N\) is the total number of subsets. is the group measurement center, and \(i = 1,\cdots,N\) j is the label of the measurement value in the \(j\)-th subset, \(z\) k,j,i is the \(i\)-th measurement sequence value in the \(j\)-th subset, \(N\) j is the total number of target measurement values in the \(j\)-th subset, \(z\) k,j is the target measurement value of the \(j\)-th subset at time \(k\), \(d(z\) k,j,i ) is the distance between the \(i\)-th target measurement value in the \(j\)-th subset and the group measurement center, \(d(z\) k,j ) is \(d(z\) k,j,i ) is the maximum value in \(d(z\) 8. The method for tracking group targets based on DBSCAN clustering according to claim 7, wherein: The detection threshold is obtained by the following method: Among them, p FA represents the false alarm probability, and H·n F is the degree of freedom of F and n is the dimension of

Images