Sea surface extended target two-stage detection method based on DK-TBD

Abstract

The invention discloses a sea surface extended target two-stage detection method based on DK-TBD, and relates to the technical field of sea surface target detection. The objective of the invention is to solve the problem of low detection accuracy of an existing sea surface target detection method. The method comprises the steps that a radar data sequence set is preprocessed, and a set composed of all real nodes in each frame of radar data sequence is obtained; generating candidate tracks by using all real node sets qk'and qk '+ 1 in two adjacent frames of radar data sequences, obtaining a candidate track set, and obtaining a credible track set by using the candidate track set; and correcting the credible trajectory by adopting a re-tracking method so as to obtain a sea surface target detection result. The method is used for sea surface target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a two-stage detection method for sea surface extended targets based on DK-TBD. Background Art

[0002] The detection of targets at sea has always been a hot topic in the research of marine radar detection systems. Unlike optical detection systems, radar detection systems are easily affected by complex scattering from the sea surface, resulting in low signal-to-noise ratio of echo signals and often accompanied by a large amount of clutter.

[0003] Traditional radar systems use signal processing algorithms to obtain target information with high credibility. One of the more classic methods is the constant false alarm rate detection (CFAR) algorithm. The CFAR method uses a specific method to count the energy information near the echo at each spatial position and adaptively set the threshold to distinguish whether the echo comes from the target. Due to the different resolutions of the radar, the detected targets also have different manifestations. In the range-azimuth coordinates, point targets are usually isolated and scattered, while extended targets are clustered into point clusters. If the radar echo data frame in this coordinate form is regarded as a 2-D single-channel digital image, the extended target will appear as an irregular closed polygon with a certain area. On top of this, combined with some simple digital image processing techniques, such as edge detection and masking, the spatial position information and geometric shape information of the extended target can be obtained. Scholars borrowed the design of the optical detection system and simplified this information into a regular rectangular bounding box, and the target state became a combination of the center position and width and height of the bounding box. Then, for the candidate target states obtained in multiple consecutive data frames, the traditional radar system uses the principle method based on the joint probability data interconnection filter (JPDAF) or Kalman filter (KF) to associate or recursively associate the candidate target states in different data frames to obtain the target's motion trajectory. This series of processing is called detection first and then tracking (DBT). In the DBT process, the extraction process of candidate targets has an almost decisive influence on further data association. Candidate targets with high false alarm rates will cause data association errors, and reducing their false alarm rates may cause real targets to be missed, resulting in low target tracking accuracy.

[0004] In order to solve the problems existing in the process of DBT method to detect sea surface targets, researchers have developed a one-time detection and tracking method called track before detection (TBD). TBD is a method to infer the detection results from the tracking trajectory. Its core is the integration and accumulation of effective information of the target and the threshold screening of reliable trajectories. Similar to the DBT method, the TBD method also has two links: candidate target generation in a single data frame and target association between multiple data frames. The difference is that the TBD method accumulates test statistics with high discrimination between target echo and clutter (such as square amplitude and envelope likelihood ratio, etc.), and designs appropriate thresholds to distinguish between true trajectories and false alarm trajectories, so as to obtain detection and tracking results of multiple data frames at the same time. Due to the process of information accumulation between multiple data frames, the sea surface target detection method based on TBD can better deal with the problem of weak target echo energy caused by low signal-to-noise ratio factors, and has certain robustness against possible missed detection and errors that may occur during data association. The researchers integrated TBD with some classic basic theories (such as Hough transform, particle filter and dynamic programming DP, etc.) and improved them, and proposed a series of effective methods. However, the performance of these basic methods has a certain degree of decline when migrating from simple scenes to complex environmental conditions. Among them, the DP-TBD method has attracted the attention of researchers because of its strong adaptability to complex environments and the lack of a preset target motion model. L.Gao, X.Li, M.Wang, H.Chen, Z.Sun and G.Cui,"Joint Intra-Frame and Inter-Frame Integration Methodfor High Speed ​​Weak Target Detection,"2022 7th International Conference onSignal and Image Processing(ICSIP),Suzhou,China,2022,pp.290-294,doi:10.1109 / ICSIP55141.2022.9886589. In order to solve the problem of information accumulation error caused by distance walking in the TBD process, the peak energy after Radon-Fourier transform (RFT) is used to improve the test statistic of DP-TBD.Haichao Jiang, W. Yi, Guolong Cui, L. Kong and Xiaobo Yang,"Radar detection and tracking of targets in the presence of clutter edge via DP-TBD," 2015 IEEE Radar Conference (RadarCon), Arlington, VA, USA, 2015, pp. 0339-0343, doi: 10.1109 / RADAR.2015.7131021. The influence of heterogeneous Rayleigh distributed clutter background on the use of signal amplitude or log-likelihood ratio as the test statistic in DP-TBD is studied, and an improved scheme of adding a data preprocessing module for clutter suppression before the TBD framework is proposed. L.Wang,G.Zhouand T.Kirubarajan,"Track-Before-Detect Method Based on Preprocessing withPseudo-Spectrum,"2020IEEE Radar Conference(RadarConf20),Florence,Italy,2020,pp.1-6,doi:10.1109 / RadarConf2043947.2020.9266669. Considering the loss of algorithm performance caused by data preprocessing, the radar data frame is regarded as an image and a pseudo spectrum is constructed for each candidate target. The algorithm performance is improved by accumulating the peak values ​​of the pseudo spectrum.E.Grossi, M.Lops, and L.Venturino.A novel dynamic programming algorithm for track-before-detect in radar systems.IEEE Trans.SignalProcess.,vol.61,no.10,pp.2608-2619,May 2013.,Aprile,E.Grossi,M.Lops,andL.Venturino.Track-before-detect for sea clutter rejection:Tests with realdata.IEEE Trans.Aerosp.Electron.Syst.,vol.52,no.3,pp.1035-1045,Jun.2016.and W.Yi,L.Kong,and J.Yang.Thresholding process based on dynamic programmingtrackbefore-detect algorithm.IEICE Trans.Commun.,vol.E 96.B, no.1, pp.291-300, 2013, a low-threshold CFAR combined with image segmentation was used to generate candidate targets, and then the DP-TBD algorithm was used to generate tracking trajectories. However, the above-mentioned DP-TBD improvement method mainly studies the adaptability of the data preprocessing method to the clutter environment and the discrimination of the test statistic on the clutter and target echo. The DP-TBD process mainly includes three parts: candidate target (also called node) generation, adjacent target search association and threshold discrimination to form trajectory. Under low signal-to-noise ratio conditions, there is usually a difference between the state of the candidate extended target obtained by preprocessing and the actual state, such as the center point position offset and the bounding box scale reduction. In the next stage, the target may fail to associate due to missed detection. These factors cause the final accumulated test statistic to be lower than the reasonable threshold, some target detection fails, and the successfully detected target has a certain error. Therefore, the current sea surface target detection method still has the problem of low detection accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of low detection accuracy in existing sea surface target detection methods, and propose a two-stage sea surface extended target detection method based on DK-TBD.

[0006] A two-stage detection method for sea surface extended targets based on DK-TBD is as follows:

[0007] Step 1: Radar data sequence set X 1:N ={x1,x2,...,x N} to perform preprocessing and obtain the set Q consisting of all real nodes in each frame of radar data sequence 1:N ={q1,q2,...,q k ,...,q N};

[0008] Among them, x N is the Nth frame radar data sequence, q N is the real node set in the Nth frame radar data sequence, k∈[1,N], N is the total number of frames in the radar data sequence, k is the frame number of the radar data sequence, q k is the real node set in the k-th frame radar data sequence;

[0009] Step 2: Use the set q of all real nodes in two adjacent frames of radar data sequence k' and q k'+1 Generate candidate trajectories and obtain the candidate trajectory set T = {t1,...,t j' ,...,t s}, and use the candidate trajectory set to obtain the credible trajectory set

[0010] Where k'=1,...,N-1,q k' is the real node set in the k'th frame radar data sequence, q k'+1 is the real node set in the k'+1th frame radar data sequence, j'∈[1,s], t j' is the j'th candidate trajectory, j' is the candidate trajectory number, s is the total number of candidate trajectories, is the s'th credible trajectory, s' is the total number of credible trajectories;

[0011] Step 3: Use the re-tracking method to correct the credible trajectory to obtain the sea surface target detection result.

[0012] The beneficial effects of the present invention are:

[0013] The method DK-TBD (DP-KCF-TBD) proposed in the present invention combines rough target detection and re-tracking. The present invention first generates candidate targets (real nodes) for each data frame using constant false alarm detection and drawing segmentation methods, uses the KCF method to generate virtual candidate associated targets (virtual nodes) of isolated targets in the next frame of radar data sequence, and then fuses the nodes that overlap in space, and then connects the nodes of continuous data frames to generate candidate trajectories. The present invention screens the connected candidate trajectories based on length and cumulative detection statistics to obtain trusted trajectories, and continuously predicts the trusted trajectories again to correct the trajectory error (i.e., re-tracking) to obtain the final sea surface target detection result. The present invention reduces the situation of missed detection and improves the accuracy of sea surface target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the present invention;

[0015] Figure 2 It is a schematic diagram of the main steps of the present invention;

[0016] Figure 3 (a) Estimated trajectory of simulated target motion under Rayleigh clutter background;

[0017] Figure 3 (b) Estimated trajectory of simulated target motion under K clutter background;

[0018] Figure 4(a) shows the PSNR-P of different methods under the condition of Rayleigh distribution clutter, where the RCS fluctuation of the target follows the Swerling model of type 0. d Curve comparison chart;

[0019] Figure 4(b) shows the PSNR-P of different methods under the condition of Rayleigh distribution clutter, where the RCS fluctuation of the target follows the Swerling model of type 1. d Curve comparison chart;

[0020] Figure 4(c) shows the PSNR-P of different methods under the condition of Rayleigh distribution clutter, where the RCS fluctuation of the target follows the Swerling model of type 3. d Curve comparison chart;

[0021] Figure 4(d) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the Rayleigh distribution clutter background;

[0022] Figure 4(e) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the Rayleigh distribution clutter background;

[0023] Figure 4(f) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the Rayleigh distribution clutter background;

[0024] Figure 4(g) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the Rayleigh distribution clutter background;

[0025] Figure 4(h) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the Rayleigh distribution clutter background;

[0026] Figure 4(i) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the Rayleigh distribution clutter background;

[0027] Figure 5(a) shows the PSNR-P of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the background of K-distributed clutter. d Curve comparison chart;

[0028] Figure 5(b) shows the PSNR-P of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the background of K-distributed clutter. d Curve comparison chart;

[0029] Figure 5(c) shows the PSNR-P of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the background of K-distributed clutter. d Curve comparison chart;

[0030] Figure 5(d) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the K-distributed clutter background;

[0031] Figure 5(e) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the K-distributed clutter background;

[0032] Figure 5(f) is a comparison of PSNR-RMSE curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the K-distributed clutter background;

[0033] Figure 5(g) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the K-distributed clutter background;

[0034] Figure 5(h) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the K-distributed clutter background;

[0035] Figure 5(i) is a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the K-distributed clutter background;

[0036] Figure 6(a) shows the effect of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the Rayleigh distribution clutter background;

[0037] Figure 6(b) shows the effect of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the Rayleigh distribution clutter background;

[0038] Figure 6(c) shows the effect of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the background of Rayleigh distribution clutter;

[0039] Figure 7 (a) The RCS fluctuation of the target follows the Swerling model of type 0. PSNR-P of the DK-TBD method with and without the re-tracking step d Curve comparison chart;

[0040] Figure 7(b) The RCS fluctuation of the target follows the Swerling model of type 1. PSNR-P of the DK-TBD method with and without the re-tracking step d Curve comparison chart;

[0041] Figure 7(c) The RCS fluctuation of the target follows the Swerling model of type 3. PSNR-P of the DK-TBD method with and without the re-tracking step d Curve comparison chart;

[0042] Figure 7(d) Comparison of PSNR-MIoU curves of the DK-TBD method with and without the re-tracking step under the condition that the RCS fluctuation of the target follows the Swerling model of type 0;

[0043] Figure 7(e) Comparison of PSNR-MIoU curves of the DK-TBD method with and without the re-tracking step under the condition that the RCS fluctuation of the target follows the Swerling model of type 1;

[0044] Figure 7(f) Comparison of PSNR-MIoU curves of the DK-TBD method with and without the re-tracking step under the condition that the RCS fluctuation of the target follows the Swerling model of type 3. DETAILED DESCRIPTION

[0045] Specific implementation method 1: Figure 1-Figure 2 As shown, the specific process of the two-stage detection method of sea surface extended targets based on DK-TBD in this embodiment is as follows:

[0046] Step 1: For a set of N radar data sequences X containing multiple extended target echoes 1:N ={x1,x2,...,x N} to perform preprocessing and obtain the set Q consisting of all real nodes in each frame of radar data sequence 1:N ={q1,q2,...,q N}, specifically:

[0047] The N-frame radar data sequence set X containing multiple extended target echoes 1:N ={x1,x2,...,x N}Preprocessing is performed, and the candidate targets generated by the preprocessing are called real nodes. The preprocessing specifically includes: a unit average constant false alarm detection method (CA-CFAR) and drawing segmentation; the threshold of the constant false alarm detection is γ1;

[0048] The set of m real nodes in the k-th frame radar data sequence is:

[0049] q k ={n k,1 ,n k,2 ,...,n k,i ,...,n k,m}

[0050] n k,i = {b k,i ,z k,i}={[x k,i ,y k,i ,w k,i ,h k,i ],z k,i}

[0051] b k,i =[x k,i ,y k,i ,w k,i ,h k,i ]

[0052]

[0053] Among them, k∈[1,N], N is the total number of radar data sequences, k is the label of the radar data sequence, q k is the set of all real nodes in the k-th frame radar data sequence, i∈[1,m], i is the real node number, m is the total number of real nodes, n k,i is the i-th real node in the k-th frame radar data sequence, b k,i is the real node n k,i The node space attribute, z k,i is the real node n k,i The average echo power within the node area Ω, x k,i Yes k,i Center point horizontal coordinate, y k,i Yes k,i The vertical coordinate of the center point, w k,i Yes k,i Width, h k,i Yes k,i High, g k,i (x, y) is the echo intensity at position (x, y) in the k-th frame radar data sequence, (x, y) is any point within the node area range Ω, a k,i is the peak amplitude of the extended target corresponding to the ith real node in the kth frame radar data sequence, c k,i is the amplitude of the independent and identically distributed clutter corresponding to the extended target of the ith real node in the kth frame radar data sequence, and h(x,y) is the point spread function of the target echo energy dispersed in the preset area;

[0054] In order to minimize missed detection of target echoes, the CFAR method used in generating real nodes usually sets a lower threshold γ1, which results in the real node set q k There are many false alarms caused by clutter or noise in the image. The extended target is the bounding rectangle of the candidate target.

[0055] Step 2: Use the set of all real nodes q in any adjacent frame radar data sequence k' and q k'+1 Generate candidate trajectories and obtain the candidate trajectory set T = {t1,...,t j' ,…,t s}, and use the candidate trajectory set to obtain the credible trajectory set Specifically:

[0056] Step 21: Get all real node sets q in any adjacent frame radar data sequence k' and q k'+1 ;

[0057] Where, k'=1,…,N-1;

[0058] Step 22: Initialize the radar sequence frame number k'=1;

[0059] Step 23: Initialize the set q consisting of all real nodes in the k'+1 frame radar data sequence k'+1 The node number i'=1;

[0060] Step 24: According to the preset distance threshold d nds In q k'+1 Search for node n k',i' If the neighboring node is found, execute step 25; if the neighboring node is not found, execute step 26;

[0061] Among them, n k',i' is the real node set q k' The i'th real node in ;

[0062] Step 25: Get the neighboring node set d k'+1,i' , the set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect, add the trajectory obtained after the connection as a candidate trajectory to the candidate trajectory set, and then execute step 27, specifically:

[0063] Step 251: Get n k',i' The set of neighboring nodes d k'+1,i' , and then get the neighboring node set d k'+1,i' The detection statistic for each neighboring node in :

[0064] λ(n k'+1，j )=z k'+1，j

[0065] Where j is d k'+1,i' The label of the neighboring node in n k'+1，j Yes k'+1,i' The jth neighboring node in z k'+1，j is the neighboring node n k'+1，j Average power in the area;

[0066] Step 252: Set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect and add the connected trajectory as a candidate estimate to the candidate trajectory set, specifically:

[0067]

[0068] Among them, → is a connection relationship.

[0069] Step 26: Generate virtual nodes using kernel correlation filtering method and compare the virtual nodes with n k',i' Connect and add the virtual node to q k'+1 Update q k'+1 , and then use the updated k'+1 The overlapping virtual nodes are merged and updated to obtain candidate trajectories, and the candidate trajectories are added to the candidate trajectory set, and then step 27 is executed, which is specifically:

[0070] Step 261: n k',i' Generate the size w for the center to be tracked δ ×h δ The initial target template δ k',i' (n k',i' ), and set the Gaussian form flag Using δ k',i' (n k',i' ) and L k',i' Get the weight parameter matrix learned by KCF

[0071]

[0072] in, is the weight parameter matrix learned by KCF, w δ is the width of the initial target template, h δ is the height of the initial target template, is the initial template δ k',i' (n k',i' )'s autocorrelation (kernel) matrix, L k',i' is the target’s initial response, is the two-dimensional Fourier transform symbol, is the two-dimensional inverse Fourier transform, is a positive hyperparameter used to adjust the effect of regularization, is the kernel function, τ = 0.5 is the proportional coefficient of the linear combination, yes The normalized result is yes The normalized result of ⊙ is element-wise multiplication, σ g is the standard deviation of the isotropic Gaussian kernel, is a Gaussian kernel function, It is a kernel function in the form of dot-product, μ and η are hyperparameters;

[0073] Step 262: Obtain the target template δ in the k'+1th frame radar sequence k',i' Monitoring areas of the same size Using the weight matrix get The corresponding response matrix Specifically:

[0074]

[0075] in, is the initial template δ k',i' And monitoring area The cross-correlation (kernel) matrix of

[0076] Step 263: Utilization The position where the maximum value appears in [x k'+1,j* ,y k'+1,j* ]Generate a virtual node:

[0077]

[0078] in, is the coordinate of the center point of the j*th virtual node generated in the k'+1th frame, is the width of the j*th virtual node generated in the k'+1th frame, is the height of the j*th virtual node generated in the k'+1th frame, is the real node n k',i' Predict the region in the k+1th frame The average power, Is a node The node status;

[0079] Step 264: Generate the virtual node With the real node n k',i' Connect and add the virtual node To the real node set q of the k'+1th frame k'+1 In k'+1 Update to the following format:

[0080] q k'+1 ={n k'+1,1 ,…,n k'+1,m' ,v k'+1,1 ,…,v k'+1,n}

[0081] Where n is the number of virtual nodes generated, m' is q k'+1 Total number of nodes;

[0082] Step 265: Update the q k'+1The virtual nodes that overlap in space are merged to obtain the merged virtual nodes, and the virtual nodes that need to be merged are replaced with the merged virtual nodes, so as to obtain candidate trajectories, and the candidate trajectories are added to the candidate trajectory set;

[0083] The updated q k'+1 There are spatially overlapping virtual nodes in the fusion:

[0084]

[0085] Among them, Ψ(··) is the node fusion function, z k'+1,f Yes k'+1,f The average power of the area, b k'+1,f is the node status of the virtual node after fusion, v k'+1,f is the virtual node after fusion. is the horizontal coordinate of the centers of the two virtual nodes that need to be fused. is the ordinate of the centers of the two virtual nodes that need to be fused. is the width of the two virtual nodes that need to be merged, is the height of the two virtual nodes that need to be merged.

[0086] Step 27: determine whether i' is equal to m'. If i'=m', execute step 28; if i'<m', set i'=i'+1 and return to step 24;

[0087] Where m' is q k'+1 Total number of nodes;

[0088] Step 28: Determine whether k' is greater than the minimum trajectory length threshold l of the target motion fn , if k'<l fn , then let k'=k'+1 and return to step 2 and 3; if l fn <k'≤N-1, execute step 29;

[0089] Step 29: Filter the candidate trajectory set to obtain the filtered candidate trajectory set; if k'<N-1, set k'=k'+1 and return to step 23; if k'=N-1, use the filtered candidate trajectory set as the set of credible trajectories

[0090] The candidate trajectory set is screened as follows:

[0091] Get candidate trajectory t j' The cumulative test statistic λ t (t j' ), and delete the candidate trajectories corresponding to the cumulative test statistics below the preset threshold γ2 to obtain a set of credible trajectories

[0092] Candidate trajectory t j' The cumulative test statistic λ t (t j' ) is obtained by:

[0093]

[0094] in, is the s'th credible trajectory, s' is the total number of credible trajectories, is the l'th credible trajectory, is the first candidate trajectory in the j'th candidate trajectory. nodes, yes The test statistic of .

[0095] Step 3: Re-track the credible trajectory in the credible trajectory set, correct the credible trajectory, and obtain the sea surface target detection result, specifically:

[0096] Step 31: Take the virtual nodes in each trusted trajectory as real nodes, and connect all real nodes to obtain the trusted trajectory expressed as follows:

[0097]

[0098] Among them, n b It is a credible trajectory The bth real node in the , b is the credible trajectory The number of the real node in b+1 It is a credible trajectory The b+1th real node in, L is the credible trajectory The real number of nodes in

[0099] Step 3.2: Using the KCF method to exploit trusted trajectories The first real node in generates the initial template δ' and the corresponding virtual node v'1, and then uses the initial template δ' and the corresponding virtual node v'1 to obtain the weight matrix, and obtains the corresponding predicted virtual nodes for other real nodes, thereby forming a virtual trajectory

[0100]

[0101] Leveraging Trusted Trajectory The first real node in generates the initial template δ' and the corresponding virtual node v'1, and then uses the initial template δ' and the corresponding virtual node v'1 to obtain the weight matrix. For other real nodes, the steps of obtaining the corresponding predicted virtual nodes are the same as steps 261 to 263.

[0102] Step 33: When the trusted track and virtual tracks When there is an overlapping part in , the center point of the smallest circumscribed rectangular area of ​​the overlapping node area is taken as the center point of the real node in the corrected credible trajectory, so as to obtain the corrected credible trajectory. Each node in the corrected credible trajectory is the target detection result.

[0103]

[0104] Among them, Ψ(n b ,v b ) is the bth real node in the corrected credible trajectory.

[0105] The invention proposed by the present invention is named DK-TBD algorithm.

[0106] The pseudo code flow of the present invention is shown in Table 1;

[0107] Table 1

[0108]

[0109] Example:

[0110] In order to verify the beneficial effects of the present invention, this embodiment carried out a simulation experiment:

[0111] This embodiment compares the DK-TBD proposed by the present invention with some TBD methods in the simulated data of Rayleigh distribution and K distribution clutter background, including DP-MSA-TBD, DP-WSA-TBD and MKCF-TBD. Among them, DP-MSA-TBD represents the classical DP-TBD method using the mean square amplitude as the test statistic, DP-WSA-TBD represents the improved DP-TBD method using the weighted square amplitude as the test statistic, and MKCF-TBD is a TBD method based entirely on KCF. All algorithms in this embodiment are implemented using Python and tested on a computer with i7-12700KF CPU (clocked at 3.60-GHz) and 32-GB RAM.

[0112] This embodiment uses the following indicators to evaluate the pre-detection tracking results of various algorithms.

[0113] (1) Discovery probability P d : When the intersection-and-union ratio of the detected extended target and the corresponding true value (i.e. the percentage of the overlapping area to the sum of the two total areas) exceeds a certain threshold, the target is considered to be found, then P d It represents the ratio of the detected part of the multi-target true trajectory to the total length of all target trajectories, as follows:

[0114]

[0115] Wherein, m represents the total number of extended targets in all data frames, and k represents the total number of targets found therein.

[0116] (2) Root mean square error (RMSE): The average distance between the center points of the multi-frame extended target estimate and the corresponding true value, as follows:

[0117]

[0118] (3) Average Intersection-over-Union (MIoU): The average intersection-over-union (MIoU) of the multi-frame extended target estimate and the corresponding true value, as follows:

[0119]

[0120] Among them, E and G represent the bounding boxes of the extended target estimate and the true value, respectively, and S represents the calculated area.

[0121] Then the motion state of the target in the Rayleigh or K distribution clutter background is simulated. Table 2 lists the motion parameters of each target.

[0122] Table 2

[0123]

[0124] Set up four extended targets with different motion models: targets A and B move according to the uniform motion model (CV), target C moves according to the accelerated motion model CA, and target D moves according to the continuous turning motion model (CT). During the motion, targets A and D first approach each other and then move away from each other after meeting. The RCS fluctuations of the targets follow the Swerling models of types 0, 1, and 3. Set the number of data frames for simulation to 50 frames, the resolution of the data frame image to 300×300, and assume that one pixel represents 1m, the time interval between two frames is 1s, and the units of initial velocity, acceleration, and turning rate are m / s, m / s respectively. 2 The motion estimation of the simulated target under the Rayleigh clutter background is shown in Figure 3(a), and the motion estimation of the simulated target under the K clutter background is shown in Figure 3(b).

[0125] For the CV motion model, it is assumed that the motion state of the target can be represented by a 4-dimensional vector:

[0126]

[0127] Among them, x k and k Respectively represent the position of the target in the horizontal and vertical directions, and Indicates the horizontal and vertical speed of the target.

[0128] The state transfer matrix of the CV motion model can be expressed as follows:

[0129]

[0130] Where I2 represents the 2×2 unit matrix, Δt represents the time interval between two frames, and the symbol represents the Kronecker product.

[0131] The CV motion model can be described as follows:

[0132]

[0133] in, is the state vector of zero-mean Gaussian noise.

[0134] For the CA motion model, it is assumed that the motion state of the target can be represented by a 6-dimensional vector:

[0135]

[0136] in, and Indicates the horizontal and vertical acceleration of the target.

[0137] The state transfer matrix of the CA motion model can be expressed as follows:

[0138]

[0139] The CA motion model can be described as follows:

[0140]

[0141] in, is the state vector of zero-mean Gaussian noise.

[0142] For the CT motion model, the motion state vector of the target is only related to its position, which is:

[0143]

[0144] The state transfer matrix of the CT motion model can be expressed as follows:

[0145]

[0146] Where R represents the turning radius and ω represents the angular velocity of the target.

[0147] The CT motion model can be described as follows:

[0148]

[0149] Among them, the state vector It consists of two zero-mean Gaussian noises and one zero noise.

[0150] According to the parameters given in Table 2, the four targets in the simulation setting move according to the CV model, CA model and CT model described above. For each simulation frame, the center positions and bounding box attributes of the four targets are recorded and set to the true values ​​for evaluating the algorithm performance.

[0151] (2) Extended target RCS fluctuation model settings:

[0152] In the simulation, Swerling 0, 1 and 3 models are used to simulate the fluctuation of the target, and a two-dimensional elliptical Gaussian function is used to model the diffusion of target energy. The target modeling includes:

[0153] g k (x,y|x k )=a k ·h k (x,y|x k )+c k

[0154] Among them, g k (x,y|x k ) represents the extended target x k The value of the pixel point in the area is also the energy intensity of the point, a k represents the peak amplitude, which changes continuously according to the preset Swerling model. k Represents independent background clutter, which changes continuously according to the preset clutter distribution.

[0155] h k (x,y|x k ) is of the following form:

[0156]

[0157] Among them, x k and k Represents the extended target x k The coordinates of the center position, variables x and y represent the current pixel and x k The relative coordinates of the center point, σ x and σ y They represent the standard deviation of the two axes of the ellipse Gaussian. θ represents the clockwise rotation angle of the target, which can be obtained by the target's speed.

[0158] For the peak amplitude a k, whose value remains unchanged in the Swerling-0 heave model, varies according to the Rayleigh distribution in the Swerling-1 heave model, and varies according to the 4-degree-of-freedom χ distribution in the Swerling-3 heave model. 2 Distribution changes.

[0159] The probability density function (PDF) of the Rayleigh distribution in the Swerling-1 model is:

[0160]

[0161] The 4-DOF χ 2 The PDF of the distribution is:

[0162]

[0163] Among them, σ t represents the average power of the target,

[0164] Assume that the average power of clutter is constant at σ c , the average peak signal-to-noise ratio between the extended target and the background clutter can be described as:

[0165]

[0166] Then the average power of the target σ t The average power of the clutter σ c And the PSNR of the two are:

[0167] σ t =σ c 10 PSNR / 10

[0168] Therefore, for different Swerling types, the peak amplitude of the extended target in the kth frame can be expressed as follows: For the Swerling-0 type, the peak amplitude of the extended target in all data frames is k equal

[0169] Table 3 lists the main hyperparameter settings of multiple TBD methods. In order to ensure the fairness of the algorithm comparison as much as possible, the false alarm rate P of the detected trajectory of all methods is controlled by controlling the threshold γ2 of the cumulative test statistic and other hyperparameters. fa It is maintained at around 0.01 to 0.05.

[0170] Table 3

[0171]

[0172] This embodiment tests the detection performance of the present invention and multiple comparison methods under different clutter backgrounds, and conducts evaluation and comparison, wherein the index evaluation results of each method adopt the average value of 100 Monte Carlo experiment evaluation results. The simulation experiment obtains multiple groups of evaluation curves by changing the clutter distribution, PSNR and Swerling type and other conditions, including PSNR-P d Curve, PSNR-RMSE curve and PSNR-MIoU curve.

[0173] Figure 4(a)-Figure 4(i) It shows the comparison of detection and tracking performance of various algorithms under Rayleigh distribution clutter background. d In the figure (Figure 4(a), Figure 4(b), and Figure 4(c), the black dashed line represents the P d Under the three types of conditions of Swerling-0, 1, and 3, DK-TBD can achieve higher P when PSNR is low (PSNR≤5). d . DK-TBD P d The curve is closer to the black dotted line, and may be higher, which reflects the improvement of the virtual node design in DK-TBD on missed detection during CFAR preprocessing. d Lower than DK-TBD's P d This is because more targets are lost in the tracking process of the baseline method, which also proves the role of virtual nodes. In the PSNR-RMSE graph (Figure 4(d), Figure 4(e), Figure 4(f)), the multi-target tracking trajectory of DK-TBD shows a lower RMSE curve. Thanks to the two designs of virtual nodes and re-tracking, DK-TBD can also achieve higher accuracy when PSNR is low. In the PSNR-MIoU graph (Figure 4(g), Figure 4(h), Figure 4(i)), although DK-TBD has a higher MIoU than the baseline DP-MSA-TBD method (whose estimation of target size depends entirely on CFAR preprocessing), it is not much different from MKCF-TBD at low SNR. This is because the multi-candidate box fusion design in MKCF-TBD greatly improves the size of the detected extended target, but at high SNR, the overlap of the multi-candidate boxes generated by MKCF-TBD becomes higher, and the scale adaptability adjustment of the target becomes worse, thus becoming lower.

[0174] Figure 5(a)-Figure 5(i) The comparison of detection and tracking performance of various algorithms under K-distributed clutter background. K-distributed clutter is more complex, and more false targets are detected during CFAR preprocessing. d , RMSE and MIoU indicators can still achieve good results.

[0175] Figure 6(a)-Figure 6(c) Analyzed the cumulative number of frames l fn Impact on algorithm performance, the discovery probability P of DK-TBD d Not with l fn increases with the increase of l, or even decreases, because the larger l fn This causes more false alarm nodes to be mixed into the credible trajectory, affecting the effective connection of the real target node. Especially in the case of low signal-to-noise ratio, the difference between the test statistics of the false alarm node and the real node is small, and it is easier to miss the target due to wrong association, P d To maintain a high P d , DK-TBD must set a lower threshold γ2, but this will cause the P fa In order to balance P d With P fa , select Set cumulative frame number l fn =3.

[0176] Figure 7(a)-Figure 7(f) The effectiveness of the re-tracking part of DK-TBD is analyzed. As can be seen from the figure, under all Swerling types, the DK-TBD method with the re-tracking part has a P d Both MIoU and RMSE are higher than those of DK-TBD without re-tracking, and the DK-TBD method with the re-tracking part has a lower RMSE, which shows that the re-tracking part can effectively improve the performance of the algorithm.

[0177] In simulation experiments, the comparison with other TBD methods shows that the detection trajectory of DK-TBD has better accuracy. In the case of low signal-to-noise ratio, the method proposed in the present invention can better discover and track the target.

Claims

1. A two-stage detection method for sea surface extended targets based on DK-TBD, characterized in that The specific process of the method is: Step 1: Radar data sequence set X 1:N ={x1,x2,...,x N }Preprocessing is performed to obtain a set Q consisting of all real nodes in each frame of radar data sequence 1:N ={q1,q2,...,q k ,...,q N }; Among them, x N is the Nth frame radar data sequence, q N is the real node set in the Nth frame radar data sequence, k∈[1,N], N is the total number of frames in the radar data sequence, k is the frame number of the radar data sequence, q k is the real node set in the k-th frame radar data sequence; Step 2: Use the set q of all real nodes in two adjacent frames of radar data sequence k' and q k'+1 Generate candidate trajectories and obtain the candidate trajectory set T = {t1,...,t j' ,...,t s }, and use the candidate trajectory set to obtain the credible trajectory set Where k'=1,...,N-1,q k' is the real node set in the k'th frame radar data sequence, q k'+1 is the real node set in the k'+1th frame radar data sequence, j'∈[1,s], t j' is the j'th candidate trajectory, j' is the candidate trajectory number, s is the total number of candidate trajectories, is the s'th credible trajectory, s' is the total number of credible trajectories; Step 3: Use the re-tracking method to correct the credible trajectory to obtain the sea surface target detection result.

2. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 1 is characterized in that: In step 1, the radar data sequence set X 1:N ={x1,x2,...,x N The preprocessing is performed by first using the unit average constant false alarm detection method to process the radar data sequence set, and then the output of the unit average constant false alarm detection method is segmented by drawing to obtain a set of all real nodes in each frame of the radar data sequence.

3. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 2 is characterized in that: The real node set in the k-th frame radar data sequence in step 1 is specifically: q k ={n k,1 ,n k,2 ,…,n k,i ,…,n k,m } n k,i ={b k,i ,z k,i }={[x k,i ,y k,i ,w k,i ,h k,i ],z k,i } b k,i =[x k,i ,y k,i ,w k,i ,h k,i ] Among them, i∈[1,m], i is the real node label, m is the total number of real nodes, n k,i is the i-th real node in the k-th frame radar data sequence, b k,i is the real node n k,i The node space attribute, z k,i is the real node n k,i The average echo power within the node area Ω, x k,i Yes k,i Center point horizontal coordinate, y k,i Yes k,i The vertical coordinate of the center point, w k,i Yes k,i Width, h k,i Yes k,i High, g k,i (x, y) is the echo intensity at position (x, y) in the k-th frame radar data sequence, (x, y) is any point within the node area range Ω, a k,i is the peak amplitude of the extended target corresponding to the ith real node in the kth frame radar data sequence, c k,i is the amplitude of the independent and identically distributed clutter corresponding to the extended target of the ith real node in the kth frame radar data sequence, and h(x,y) is the point spread function of the target echo energy dispersed in the preset area.

4. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 3 is characterized in that: The step 2 uses all the real node sets q in the two adjacent frames of radar data sequences k' and q k'+1 Generate candidate trajectories and obtain the candidate trajectory set T = {t1,...,t j' ,...,t s }, and use the candidate trajectory set to obtain the credible trajectory set Specifically: Step 21: Obtain all real node sets q in the adjacent frame radar data sequence k' and q k'+1 ; Where, k'=1,...,N-1; Step 22: Initialize the radar sequence frame number k'=1; Step 23: Initialize the set q consisting of all real nodes in the k'+1 frame radar data sequence k'+1 The node number i'=1; Step 24: According to the preset distance threshold d nds In q k'+1 Search for node n k',i' If the neighboring node is found, execute step 25; if the neighboring node is not found, execute step 26; Among them, n k',i' is the real node set q k' The i'th real node in ; Step 25: Get the neighboring node set d k'+1,i' , the set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect, add the trajectory obtained after the connection as a candidate trajectory to the candidate trajectory set, and then execute step 27; Step 26: Generate virtual nodes using kernel correlation filtering method and compare the virtual nodes with n k',i' Connect and add the virtual node to q k'+1 Update q k'+1 , and then use the updated k'+1 The overlapping virtual nodes are merged and updated to obtain candidate trajectories, and the candidate trajectories are added to the candidate trajectory set, and then step 27 is executed; Step 27: determine whether i' is equal to m'. If i'=m', execute step 28; if i'<m', set i'=i'+1 and return to step 24; Step 28: Determine whether k' is greater than the minimum trajectory length threshold l of the target motion fn , if k'<l fn , then let k'=k'+1 and return to step 2 and 3; if l fn <k'≤N-1, execute step 29; Step 29: Filter the candidate trajectory set to obtain the filtered candidate trajectory set; if k'<N-1, set k'=k'+1 and return to step 23; if k'=N-1, use the filtered candidate trajectory set as the set of credible trajectories in, is the s'th credible trajectory, s' is the total number of credible trajectories, is the l'th credible trajectory.

5. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 4 is characterized in that: The step 25 of obtaining the neighboring node set d k'+1,i' , the set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect, add the trajectory obtained after the connection as a candidate trajectory to the candidate trajectory set, specifically: Step 251: Get n k',i' The set of neighboring nodes d k'+1,i' , and then get the neighboring node set d k'+1,i' The detection statistic for each neighboring node in : λ(n k'+1，j )=z k'+1，j Where j is d k'+1,i' The label of the neighboring node in n k'+1，j Yes k'+1,i' The jth neighboring node in z k'+1，j is the neighboring node n k'+1，j Average power in the area; Step 252: Set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect and add the connected trajectories as candidate estimates to the candidate trajectory set.

6. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 5 is characterized in that: In the step 26, the virtual node is generated by using the kernel correlation filtering method, and the virtual node is connected to n k',i' Connect and add the virtual node to q k'+1 Update q k'+1 , and then use the updated k'+1 The overlapping virtual nodes are merged and updated to obtain candidate trajectories, and the candidate trajectories are added to the candidate trajectory set. Specifically: Step 261: n k',i' Generate the size w for the center to be tracked δ ×h δ The initial target template δ k',i' (n k',i' ), and set the Gaussian form flag Using δ k',i' (n k',i' ) and L k',i' Get the weight parameter matrix learned by KCF in, is the weight parameter matrix learned by KCF, w δ is the width of the initial target template, h δ is the height of the initial target template, is the initial template δ k',i' (n k',i' )’s autocorrelation matrix, L k',i' is the target’s initial response, is the two-dimensional Fourier transform symbol, is the two-dimensional inverse Fourier transform, is a positive hyperparameter, is the kernel function, τ is the proportional coefficient of the linear combination, yes The normalized result is yes The normalized result of ⊙ is element-wise multiplication, σ g is the standard deviation of the isotropic Gaussian kernel, is a Gaussian kernel function, is a kernel function in the form of a dot product, μ and η are hyperparameters; Step 262: Obtain the target template δ in the k'+1th frame radar sequence k',i' Monitoring areas of the same size Using the weight matrix get The corresponding response matrix Specifically: in, is the initial template δ k',i' And monitoring area The cross-correlation matrix of Step 263: Utilization The position where the maximum value appears Generate virtual nodes Step 264: Generate the virtual node With the real node n k',i' Connect the virtual node Add to the real node set q of the k'+1th frame k'+1 Update q k'+1 : q k'+1 ={n k'+1,1 ,…,n k'+1,m' ,v k'+1,1 ,...,v k'+1,n } Where n is the number of virtual nodes generated, m' is q k'+1 Total number of nodes; Step 265: Update the q k'+1 The virtual nodes that overlap in space are fused to obtain the fused virtual nodes, and the virtual nodes that need to be fused are replaced with the fused virtual nodes to obtain candidate trajectories, and the candidate trajectories are added to the candidate trajectory set.

7. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 6 is characterized in that: The use of step 263 The position where the maximum value appears Generate virtual nodes Specifically: in, is the coordinate of the center point of the j*th virtual node generated in the k'+1th frame, w k'+1,j* is the width of the j*th virtual node generated in the k'+1th frame, is the height of the j*th virtual node generated in the k'+1th frame, is the real node n k',i' Predict the region in the k+1th frame The average power, Is a node The node status.

8. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 7 is characterized in that: In step 265, the updated q k'+1 There is a fusion of virtual nodes that overlap in space, specifically: Among them, Ψ(··) is the node fusion function, z k'+1,f Yes k'+1,f The average power of the area, b k'+1,f is the node status of the virtual node after fusion, v k'+1,f is the virtual node after fusion. is the horizontal coordinate of the centers of the two virtual nodes that need to be fused. is the ordinate of the centers of the two virtual nodes that need to be fused. is the width of the two virtual nodes that need to be merged, is the height of the two virtual nodes that need to be merged.

9. The two-stage detection method for sea surface extended targets based on DK-TBD according to claim 8 is characterized in that: The candidate trajectory set in step 29 is screened, and the screened candidate trajectory set is specifically: Get candidate trajectory t j' The cumulative test statistic λ t (t j' ), and delete the candidate trajectories corresponding to the cumulative test statistics that are lower than the preset threshold γ2 to obtain a screened candidate trajectory set; Candidate trajectory t j' The cumulative test statistic λ t (t j' ) is obtained by: in, is the first candidate trajectory in the j'th candidate trajectory. nodes, yes The test statistic of .

10. A two-stage detection method for sea surface extended targets based on DK-TBD according to claim 9, characterized in that: The re-tracking method in step 3 is used to correct the credible trajectory, so as to obtain the sea surface target detection result, specifically: Step 31: Take the virtual node in each trusted track as the real node, and connect all the real nodes to obtain the trusted track, as follows: Among them, n b It is a credible trajectory The bth real node in the , b is the credible trajectory The number of the real node in b+1 It is a credible trajectory The b+1th real node in, L is the credible trajectory The real number of nodes in Step 32: Utilize Trusted Trajectory The first real node in generates the initial template δ' and the corresponding virtual node v'1, and then uses the initial template δ' and the corresponding virtual node v'1 to obtain the weight matrix, and obtains the corresponding predicted virtual nodes for other real nodes, thereby forming a virtual trajectory Virtual Track for: Step 33: When the trusted track and virtual tracks When there is an overlapping part in , the center point of the smallest circumscribed rectangular area of ​​the overlapping node area is taken as the center point of the real node in the corrected credible trajectory, so as to obtain the corrected credible trajectory. Each node in the corrected credible trajectory is the target detection result. The corrected credible trajectory is as follows: Among them, Ψ(n b ,v b ) is the bth real node in the corrected credible trajectory.

Images