A method for extracting a ship channel model based on adaptive trajectory optimization

Abstract

The present invention discloses a method for extracting a ship channel model based on adaptive trajectory optimization. An adaptive composite kernel function integrating trajectory motion information is designed to optimize and recalculate the position and motion characteristics of trajectory points, making the trajectory approach the corresponding cluster center, so as to obtain a more discriminative trajectory representation. The recalculated trajectory is further optimized by constructing an energy function with differential smoothing and sparse penalty terms to ensure good smoothness characteristics of the trajectory and avoid deviating too much from the original trajectory. Finally, the optimized trajectory is combined with the dynamic time warping distance and the DBSCAN density clustering method to realize the extraction of the ship trajectory model. The present invention can make the route features more obvious while eliminating the ambiguity between trajectories and improving the discrimination between clusters, effectively improving the channel clustering effect.

Description

Technical Field

[0001] The present invention relates to a method for extracting a ship channel model, belonging to the field of data mining. Background Art

[0002] Detecting abnormal targets at sea is an effective means to improve the automation and intelligence levels of maritime safety supervision, which can ensure the safety of ship navigation, timely monitor maritime accidents and illegal acts, optimize traffic efficiency, and ensure the safety and economy of maritime navigation. In the research of abnormal target detection, it often relies on knowledge products of target activity rules formed by mining a large amount of historical ship trajectory data for multi-dimensional compliance matching, judging target intentions, discovering abnormal behaviors, and giving timely warnings for highly threatening targets such as suspicious and camouflaged ones. Knowledge products of target activity rules include data products such as channels, typical activity rules, and target characteristics. By extracting, modeling, clustering, and detecting anomalies of ship navigation trajectories, the description and analysis of ship behaviors can be realized and real-time anomaly detection can be achieved, which can give a warning at the initial stage of ship abnormal behaviors, timely discover potential safety hazards, and ensure the safe navigation of ships.

[0003] Ship trajectory clustering analysis is a commonly used means for anomaly detection. The form of a normal trajectory model is the most representative actual trajectory cluster in the dataset. The main clustering algorithms can be divided into clustering algorithms based on the partitioning idea and density-based clustering algorithms. The DBSCAN spatial density clustering algorithm has been widely used in ship abnormal trajectory detection because it can effectively process noise data and discover trajectory clusters of various shapes, and at the same time combines the DTW (Dynamic Time Warping) distance that can solve the matching problem of discrete time series with different lengths and rhythms to calculate the similarity between trajectories.

[0004] The trajectory records the position and speed information of the target over time and contains important information for clustering. Directly clustering the trajectory often yields poor results for two reasons: First, due to the uncertainty of the ship's movement characteristics, the variance of ship trajectories in the same type of trajectory cluster is large and the distinctiveness is low, which easily results in a low trajectory similarity metric value and affects the subsequent clustering effect. Second, ship trajectories with the same shape and similar positions may belong to different trajectory clusters due to potential semantic features (such as moving towards each other), resulting in the trajectory position information being blurred at the cluster boundary, which easily causes the performance of the clustering algorithm to decline. To address the above problems, the present invention uses the ship AIS (Automatic Identification System) trajectory data in the water area and studies a method for extracting a ship channel model based on adaptive trajectory optimization. By combining the multiple structural information of the trajectory and the local motion characteristics between trajectories, an adaptive composite kernel function is designed to optimize the calculation of the position and speed of the trajectory points, and an optimization process with differential smoothing and sparse regularization is adopted to ensure the smoothness of the trajectory and good distinctiveness between trajectory clusters, eliminate the ambiguity between trajectories, improve the distinctiveness between clusters, and improve the clustering accuracy. Summary of the Invention

[0005] Aiming at the problems that the existing ship channel extraction results are easily affected by the fuzzy boundary between trajectory clusters and the large variance and low distinguishability within the clusters, the present invention proposes a method for extracting a ship channel model based on adaptive trajectory optimization, aiming to improve the distinctiveness between trajectory clusters and ensure the smoothness of the trajectory, so as to improve the accuracy and efficiency of ship trajectory clustering, and provide guarantee for the accurate mining of the target activity law knowledge and the behavior evaluation of combat readiness equipment.

[0006] To solve the above technical problems, the present invention provides a method for extracting a ship channel model based on adaptive trajectory optimization, including the following steps:

[0007] 1) Downsample the original trajectory data to reduce the computational complexity of subsequent trajectory optimization and clustering algorithms;

[0008] 2) Design an adaptive composite kernel function that fuses trajectory motion information to optimize and recalculate the position and motion characteristics of the trajectory points, so that the trajectory approaches the corresponding cluster center, thereby obtaining a more distinctive trajectory representation;

[0009] 3) Further optimize the recalculated trajectory by constructing an energy function with differential smoothing and sparse penalty terms to ensure good smoothness characteristics of the trajectory and avoid deviating too much from the original trajectory;

[0010] 4) Finally, combine the optimized trajectory and use the dynamic time warping distance and DBSCAN density clustering method to realize the extraction of the ship trajectory model.

[0011] Advantages achieved by the present invention: The present invention solves the problems that in the same type of trajectory clusters, the trajectory variance is large and the distinguishability is low, resulting in a low trajectory similarity measurement value and blurred trajectory position information at the cluster boundary, thereby affecting the clustering effect. It eliminates the ambiguity between trajectories, improves the inter-cluster distinguishability and trajectory smoothness, and better detects abnormal ship behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] The implementation schematic diagram of the present invention is as shown in Figure 1 and the specific steps of the embodiment include:

[0014] 1) Downsample the original trajectory data to reduce the computational complexity of subsequent trajectory optimization and clustering algorithms. The preferred implementation process includes:

[0015] Assume that the ship trajectory set corresponds to the movement processes of M target voyages. Any trajectory composed of trajectory points can be represented as T i Total = [p i,1 , …, p i,Total Τ , where p i,t ∈ R D represents the D-dimensional coordinate position information of the i-th voyage at time t. To improve the extraction efficiency of the ship channel model, dimensionality reduction processing is performed on the ship trajectory. As shown in the following formula, one trajectory point is extracted from every M sample points in T i Total (t), and the remaining trajectory points are discarded to generate the output trajectory sequence T i = [p i,1 , …, p i,T Τ .

[0016] T i (t) = T i Total (tM) (1)

[0017] For any trajectory point p i,t , its corresponding target transient motion speed can be estimated by v i,t = (p i,t+1 - p i,t ) / Δt. At this time, the speed corresponding to the trajectory T i can be expressed as V i = ΔT i = [v i,1 , …, v i,T Τ .

[0018] 2) Design an adaptive composite kernel function that integrates trajectory motion information to optimize and recalculate the position and motion characteristics of trajectory points, so that the trajectory is close to the corresponding cluster center, thereby obtaining a more discriminative trajectory representation. The preferred implementation process includes:

[0019] 21) Assume that there are K different cluster centers in the trajectory set, trajectory T i It belongs to one of the trajectory clusters. The trajectory position is recalculated by an adaptive composite kernel function to make it closer to the corresponding cluster center, thereby obtaining a more discriminative trajectory representation.

[0020] For any trajectory point p i,t ∈T i , first find out its neighborhood trajectory point set Each neighborhood trajectory point p n Satisfy||p n -p i,t ||2≤r, where r is the neighborhood radius. The trajectories of these neighborhood trajectory points may belong to different trajectory clusters. Then the trajectory point p is calculated by non-parametric estimation method. i,t Optimize the calculated position and corresponding speed:

[0021]

[0022]

[0023] In the formula, the kernel function K(p i,t ,p n ) is crucial for the rationality of the estimation. A composite kernel function is designed to fully integrate the trajectory structure information from multiple aspects. The composite kernel function consists of three parts and is defined as follows: n' ∈T i , p n ∈T j :

[0024] K(p n' ,p n )=(1-ζ)·min(K p (p n' ,p n )K s (p n' ,p n ),th ps )+ζ·min(K a (p n' ,p n ),th a )(4)

[0025]

[0026]

[0027]

[0028] wherein, K p (p n' , p n ) and K s (p n' , p n ) respectively represent kernel functions for measuring the position and velocity consistency between trajectory points. Points with similar positions / velocities are closer in the corresponding kernel space. K a (p n' , p n ) is the kernel function for measuring Figure 1 consistency. a i = [p i,1 ; p i,T records the start and end state information of the trajectory T i . The parameter ζ is used to adjust the weights of the trajectory motion and direction information. th ps and th a are the cut-off values used to suppress the influence of outliers. The parameters α, β, and γ are the self-adjustment coefficients of the kernel functions for the position, velocity, and intention of the trajectory points respectively. The shrunk trajectory estimated by formula (2) is expressed as The velocity estimated by formula (3) is expressed as where i = 1,..., M.

[0029] When using the designed composite kernel function, the kernel functions at adjacent trajectory points are anisotropic. At this time, the trajectory points shift towards the directions of adjacent trajectories within the same cluster, and the ambiguity between trajectories can be effectively eliminated, and the distinguishability between clusters can be improved.

[0030] 22) For the parameters in the algorithm: the neighborhood radius r, the kernel function parameters α, β, γ, an adaptive strategy is adopted to configure these parameters. Specifically, given the trajectory set The histogram of trajectory points is calculated by the following formula to represent the density map Μ = [H(x1,..., x D )]:

[0031] H(x1,..., x D ) = |{p i,t ∈ [x1, x1 + 1) ×.. × [x D , x D + 1)}| (8)

[0032] wherein, |.| represents calculating the cardinality of the set. The smoothed density map is obtained through a Gaussian filter Thus, the calculation method of r is set as where C is a constant. Given r i,t , set α = β = r i,t / 3, γ = 2r i,t / 3. This setting can effectively reduce the influence of non-clustered trajectory points in the neighborhood.

[0033] 3) Further optimize the recalculated trajectory by constructing an energy function with differential smoothing and sparse penalty terms to ensure good smoothness characteristics of the trajectory and avoid deviating too much from the original trajectory. The preferred implementation process includes:

[0034] For the trajectory calculated by the above method, while shifting the trajectory points in the direction of adjacent trajectories within the same cluster, it is necessary to ensure the retention of the relative position information of the trajectory points to avoid the shrunk trajectory deviating too much from the original trajectory. To address this problem, the trajectory is further optimized by constructing differential smoothing and sparse regularization terms, and the penalty term of the optimization function is constructed using the speed estimate value to obtain the optimized target trajectory:

[0035]

[0036] In the formula, the first term is the data fidelity term, which can ensure that the finally optimized trajectory is close to the trajectory obtained by adaptive multi-core estimation ; the second term is the regular term for multi-trajectory speed guidance and sparse constraint, which can ensure that the speed of the optimized trajectory is close to the speed estimate value , where W i = diag(w i,1 , …, w i,T ), and its elements can be calculated by . For Equation (9), the two sub-objective functions of Equation (10) and (11) are constructed for step-by-step optimization and solution:

[0037]

[0038]

[0039] In Equation (10), this optimization problem belongs to a regularized least squares problem, and its closed-form solution is:

[0040]

[0041] In the formula, I is the identity matrix. Combining with the gradient descent algorithm to solve the sub-objective function shown in Equation (11), the final trajectory optimization result

[0042] 4) Combine the optimized trajectory and use the dynamic time warping distance and DBSCAN density clustering method to achieve the extraction of the ship trajectory model. The preferred implementation process includes:

[0043] The trajectory optimized based on Equation (9) The dynamic time warping distance is used to calculate the similarity metric value between trajectories, and combined with the DBSCAN density clustering algorithm for trajectory clustering to obtain the final clustering result of the ship channel model. The formula for the dynamic time warping distance is as follows:

[0044] D(T i ,T j )=dist(p i,1 ,p j,1 )+min{D(R(T i ),R(T j )),D(R(T i ),T j ),D(T i ,R(T j ))}(13)

[0045] In the formula, dist(p i,1 ,p j,1 )=||p i,1 -p j,1 ||2 represents the Euclidean distance between the trajectory points p i,1 and p j,1 . R(T i ) represents the remaining trajectories except the currently calculated trajectory point p i,1 .

[0046] The above has introduced in detail a method for extracting a ship channel model based on adaptive trajectory optimization proposed by the present invention, which can be widely applied to the fields of maritime ship trajectory clustering analysis and abnormal target discovery.

Claims

1. A method for extracting a ship channel model based on adaptive trajectory optimization, characterized in that: 1) Downsample the original trajectory data to reduce the computational complexity of subsequent trajectory optimization and clustering algorithms; 2) Design an adaptive composite kernel function that fuses trajectory motion information to optimize and recalculate the position and motion characteristics of trajectory points, making the trajectory closer to the corresponding cluster center to obtain a more discriminative trajectory representation; 3) Further optimize the recalculated trajectory by constructing an energy function with differential smoothing and sparse penalty terms to ensure good trajectory smoothing characteristics and avoid deviating too much from the original trajectory; 4) Combine the optimized trajectory and use the dynamic time warping distance and density clustering method to achieve the extraction of the ship trajectory model.

2. The method for extracting a ship channel model based on adaptive trajectory optimization according to claim 1, wherein: The step (2) further includes: (21) Suppose there are K different cluster centers in the trajectory set, and the trajectory T i then belongs to one of the trajectory clusters. For any trajectory point p i,t ∈T i , first find out its neighborhood trajectory point set Each neighborhood trajectory point p n satisfies ||p n - p i,t ||2 ≤ r, where r is the neighborhood radius. Then, calculate the optimized position and corresponding velocity of the trajectory point p i,t by a non-parametric estimation method: Design a composite kernel function \(K(p i,t ,p n ) that consists of three parts and is defined as follows: For \(p n' \in T i , p n \in T j K(p n' ,p n ) = (1 - ζ)·min(K p (p n' ,p n ),th s (p n' ,p n ),th ps ) + ζ·min(K a (p n' ,p n ),th a ) (4) where K p (p n' , p n ) and K s (p n' , p n ) respectively represent kernel functions for measuring the position and velocity consistency between trajectory points. Points with similar positions / velocities are closer in the corresponding kernel space; K a (p n' , p n ) is the kernel function for measuring intention consistency, a i = [p i,1 ; p i,T records the start and end state information of trajectory T i ; the parameter ζ is used to adjust the weights of trajectory motion and direction information, th ps and th a are cut-off values used to suppress the influence of outliers; the parameters α, β, and γ are the self-adjustment coefficients of the kernel functions for trajectory point position, velocity, and intention respectively; the shrunk trajectory estimated by formula (2) is expressed as The velocity estimated by formula (3) is expressed as where i = 1, …, M; 22) The neighborhood radius r, kernel function parameters α, β, γ are configured using an adaptive strategy for such parameters; given a trajectory set The density map Μ = [H(x1,…,x D )] of the trajectory points is calculated by the following formula to represent the trajectory points H(x1,…,x D )=|{p i,t ∈[x1,x1+1)×..×[x D ,x D +1)}| (8) Wherein, |.| represents calculating the cardinality of a set, and the smoothed density map is obtained through a Gaussian filter. Therefore, the calculation method of r is set as where C is a constant; given r i,t , set α = β = r i,t / 3, γ = 2r i,t / 3.

3. A method for extracting a ship channel model based on adaptive trajectory optimization according to claim 2, characterized in that: The step (3) further includes: Further optimize the trajectory by constructing differential smoothing and sparse regularization terms, and use the speed estimate value to construct the penalty term of the optimization function to obtain the optimized target trajectory: In the formula, the first term is the data fidelity term, which can ensure that the finally optimized trajectory is similar to the trajectory obtained by adaptive multi-core estimation ; the second term is the regular term of multi-trajectory speed guidance and sparse constraint, which can ensure that the speed of the optimized trajectory is similar to the speed estimation value , where W i = diag(w i,1 , …, w i,T ), and its elements can be calculated by ; for Equation (9), it is gradually optimized and solved by constructing two sub-objective functions of Equation (10) and (11): In Equation (10), this optimization problem belongs to a regularized least squares problem, and its closed-form solution is: Where, I is the identity matrix; combining with the gradient descent algorithm to solve the sub-objective function shown in Equation (11), the final trajectory optimization result is obtained

Images