Sea surface target classification method based on multi-motion feature fusion

Abstract

A sea surface target classification method based on multi-motion feature fusion is characterized by comprising the steps that a double-station high-frequency radar is adopted to invert a sea surface vector flow field and a wind field, and quality control and high-precision interpolation are carried out on the flow field and the wind field; simulating a drifting target track by adopting a kinetic equation, and extracting an artificial power target track by adopting radar detection, including target detection, false alarm suppression, track initiation, tracking filtering and track termination; extracting motion features of the target from the track data, including a speed feature, an acceleration feature, a course feature, a steering rate feature and a track curvature feature, and constructing a motion feature data set; and designing the architecture of the multi-channel deep convolutional neural network, training the network weight, and identifying an artificial power target and a drifting target.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and in particular relates to a target recognition technology based on motion feature fusion. Background Art

[0002] Surface target identification has significant application value. Target propulsion type is highly correlated with target threat level. For example, powered targets may be ships or periscopes, posing a significant threat. Non-powered targets may be buoys, icebergs, or debris, posing a relatively low threat level. Accurately identifying surface target propulsion type effectively supports target threat assessment, enhances situational awareness, and reduces radar resource consumption.

[0003] Differences in propulsion types lead to variations in target motion characteristics such as speed, acceleration, heading, rate of turn, and track curvature. A master's thesis from Xidian University, "Radar Target Recognition Based on Multi-Feature Fusion," describes a radar target recognition method. This method first extracts multidimensional features from the track, then uses individual features to identify the target separately. Finally, it implements decision-level fusion recognition using DS evidence theory. However, this method suffers from performance issues when resolving conflicts in multi-feature recognition results and high ambiguity in the features of multiple target classes.

[0004] Deep convolutional neural networks (CNNs) use convolution operations to extract semantic features from data and employ fully connected networks to achieve nonlinear mapping from features to target types. They are end-to-end, high-quality feature extraction and target classification algorithms. Deep CNNs are widely used in image target recognition and radar one-dimensional target recognition. They also have great potential for classifying sea surface targets based on motion characteristics. The Journal of Wuhan University of Technology (Transportation Science and Engineering) published a paper titled "Research on Ship Trajectory Classification Based on Deep Convolutional Neural Networks," which describes a ship classification method based on deep CNNs.

[0005] Existing technologies for classifying artificial powered and drifting targets on the sea surface have two shortcomings: 1. The network needs to simultaneously process multiple motion features with different dimensions and large numerical differences, and the traditional deep CNN architecture cannot meet this requirement; 2. Data-driven deep CNNs rely on large amounts of training data to ensure convergence of recognition performance, making it difficult and costly to obtain the tracks of drifting targets on the sea surface. Summary of the Invention

[0006] In the field of classification of artificial powered targets and drifting targets on the sea surface, the existing rule-based classification methods cannot meet the requirements of recognition accuracy and robustness. In order to solve the technical problems of difficulty in acquiring drift track data and unsuitable network architecture in the classification method based on deep CNN, the present invention adopts a technical solution based on multi-motion feature fusion. The present invention first uses high-frequency radar to invert vector flow field and wind field for quality control and interpolation; then uses dynamic equations to realize drift target track simulation based on high-precision flow field and wind field; then uses target detection radar to realize artificial powered target track extraction; then constructs a motion feature data set based on the target track; finally, uses multi-channel deep CNN to realize multi-motion feature fusion to complete the classification of artificial powered targets and drifting targets. Under the premise of greatly reducing the difficulty of acquiring drift target tracks, the present invention has the technical effect of greatly improving target classification accuracy and robustness, and supporting radar type recognition and threat judgment of sea surface targets. The present invention makes full use of the combination of motion features such as speed, acceleration, heading, turning rate, and track curvature to realize the classification of artificial powered targets and drifting targets, and greatly improves the classification accuracy and robustness; in order to address the problems of different dimensions of different motion features, large numerical differences, difficulty in determining joint recognition weights, and difficulty in characterizing feature relationships, a multi-channel deep CNN is designed to realize motion feature extraction, fusion and mining, thereby improving target recognition performance; the ocean surface flow field and wind field inverted by high-frequency radar are used to achieve high-precision simulation of the drifting target track, meet the deep CNN's demand for training data volume, significantly reduce data accumulation costs, and improve the engineering practicality of the algorithm.

[0007] The sea surface vector flow field is inverted using a dual-station high-frequency radar: radial flow is extracted, vector flow field is synthesized, and an M*N two-dimensional vector flow matrix is output.

[0008] A dual-station high-frequency radar is used to invert the sea surface vector wind field: invert wind direction and wind speed, and output an M*N two-dimensional vector wind matrix.

[0009] The above M and N are the number of resolution units in longitude and latitude respectively.

[0010] Perform quality control on the flow field: Utilize the continuity of ocean surface flow in space and perform sliding window processing on the latitude and longitude coordinate system. The window size is set to 5 resolution units * 5 resolution units and the step is set to 1 resolution unit. Let the sliding window two-dimensional velocity matrix before quality control be R v0 , the two-dimensional flow matrix is R θ0 , calculate the mean flow velocity within the window Flow direction mean Flow rate standard deviation σ v , flow direction standard deviation σ θ ; If the flow rate of a resolution unit v i Dissatisfied Or flow to θ iDissatisfied Then remove the vector flow of the resolution unit; let the sliding window two-dimensional velocity matrix after quality control be R v1 , the two-dimensional flow matrix is R θ1 .

[0011] The same method is used to control the quality of the wind field. Let the two-dimensional wind speed matrix be R w1 , the two-dimensional wind direction matrix is R d1 .

[0012] High-precision interpolation of the flow field: Because the radar beam does not illuminate some resolution units, or some resolution units are eliminated by quality control, there are blank units in the velocity matrix; calculate the velocity matrix R v1 The mean use Instead of R v1 The blank cells in the matrix R are obtained v2 ; R v2 Singular value decomposition, let R v2 =UΛS T , where S T represents the transpose of the S matrix, Λ=diag[λ1,λ2,λ i ,…,λ5] is the singular value matrix, diag() represents the diagonal matrix, λ i represents singular values, U = [u1, ..., u5] and S = [s1, ..., s5] are EOF matrices, EOF is the empirical orthogonal function, u i and s i is the eigenvector, i ranges from 1 to 5; take the first q eigenvectors of U and S to reconstruct the matrix, let where q satisfies The minimum value of the constraint; use R v3 The corresponding unit in replaces R v1 The blank cells in the flow velocity are obtained by high-precision interpolation matrix R. v4 .

[0013] Using the same method, high-precision interpolation is performed on the matrix of flow direction, wind speed, and wind direction to obtain the high-precision interpolation matrix R of flow direction. θ4 , wind speed high-precision interpolation matrix R w4 , wind direction high-precision interpolation matrix R d4 .

[0014] Use dynamic equations to simulate the drift target trajectory: Based on the high-precision flow field, the dynamic equation is set as Where L(0) is the initial latitude and longitude of the target, V w (t) is the wind-induced floating speed, V c (t) is the ocean surface velocity, Vk (t) is the floating speed generated by the Coriolis force, t is time, L(t) is the latitude and longitude of the target at time t; from the high-precision interpolation matrix R of wind speed w4 , wind direction high-precision interpolation matrix R d4 The wind speed and direction at 10m above the ocean surface are calculated and expressed as v0 and θ respectively. k (t) Orthogonal decomposition in longitude and latitude gives The initial longitude and latitude of the target are randomly set to L(0), and the longitude and latitude of the target at time t is obtained to be L(t), which is the drifting target track.

[0015] Radar detection is used to extract the track of artificial powered targets, including target detection, false alarm suppression, track initiation, tracking filtering, and track termination.

[0016] The following motion characteristics of the target are extracted from the track data:

[0017] Speed characteristics: In the target track sequence, a sliding window is used to select samples and calculate the speed characteristic vector; the window length is set to 20 track points and the step is 1 track point; the speed characteristic vector and time characteristic vector are constructed using the data in the window, and the speed vector and time vector of the artificial power target are set to be V r =[v r1 ,v r2 ,v ri ,…,v r20 ]、T r =[t r1 ,t r2 ,t ri ,…,t r20 ],v ri , t ri They represent the speed value and time value of the artificial power target respectively. The speed vector and time vector of the drifting target are V p =[v p1 ,v p2 ,v pi ,…,v p20 ]、T p =[t p1 ,t p2 ,t pi ,…,t p20 ],v pi , t pi They represent the speed and time values of the drifting target respectively.

[0018] Acceleration characteristics: Use the velocity characteristic vector to find the partial derivative of the time characteristic vector to calculate the acceleration characteristic vector. Suppose the acceleration vector of the artificial power target is The acceleration vector of the drifting target is

[0019] Extract the heading characteristics of the target: When using the sliding window to calculate the velocity feature vector, the heading feature vector is constructed using the samples in the window. Let the heading vectors of the artificial power target and the drifting target be θ respectively. r =[θ r1 ,θ r2 ,θ ri ,…,θ r20 ]、θ p =[θ p1 ,θ p2 ,θ pi ,…,θ p20 ],θ pi ,θ pi They represent the heading values of artificial powered targets and drifting targets respectively.

[0020] Turn rate characteristics: Use the heading characteristic vector to find the partial derivative of the time characteristic vector to calculate the turning rate characteristic vector. Let the turning rate vector of the artificial power target be The turning rate vector of the drifting target is

[0021] Track curvature characteristics: When using a sliding window to calculate the velocity feature vector, the longitude and latitude vectors of the track are constructed using the samples in the window, and the second and first derivatives of the longitude vector are calculated using the latitude vector. r ″、y r ′ represents the trajectory curvature vector of the artificial power target. Using the same method, the track curvature vector K of the drifting target is calculated p .

[0022] Constructing motion feature dataset: Splicing multiple motion features of artificial dynamic targets, assuming K r is 1*18 dimensional, A r and Z r is 1*19 dimensions, V r and θ r It is 1*20 dimensions, and the dimensions are inconsistent; select A r 、Z r 、V r ,θ r The first 1*18 dimensions of the eigenvector A′ are constructed r , Z′ r , V′ r ,θ′ r ; Vertically concatenate the characteristic vectors of velocity, acceleration, heading, turning rate, and track curvature to construct a 5*18 two-dimensional matrix Using the same method, construct the two-dimensional matrix M of multiple motion features of the drifting target p ; Sliding window processing of multiple tracks, constructing multiple M r and Mp The datasets are merged into the target motion feature dataset M.

[0023] Different motion features have different dimensions and ranges of values. The element M in the dataset M i The values of each row are different; using traditional CNN to process M i ,The eigenvectors with large values play a dominant role in the recognition results, ,while eigenvectors with small values are not fully utilized, and the ,recognition performance is degraded.

[0024] Design the architecture of a multi-channel deep convolutional neural network: use independent convolution kernels to perform the multi-channel convolution on the elements M in the dataset M. i Convolution is performed on each row of , and the deep features formed by convolution are spliced and fused in the fully connected layer to identify the target type; Figure 2 As shown, the multi-channel deep CNN has 8 layers, including input layer, convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, fully connected layer 1, fully connected layer 2, and output layer; the size of the input layer is 5*18, and the two-dimensional feature vector M is connected. i ; The output layer has a vector size of 2*1 Among them, p1 and p2 are the confidence of artificial power targets and drifting targets respectively, and the sum of the two is 1; the loss function adopts the cross entropy loss function.

[0025] Training network weights: Collect 100 tracks of artificial powered targets and drifting targets, and extract 2000 two-dimensional motion feature matrices M for each. i , forming a dataset M of 4000 samples; randomly select 3500 samples as the training dataset and the remaining 500 samples as the validation dataset; input the two-dimensional feature matrix in the training dataset into the multi-channel deep CNN network one by one, calculate the weights, and calculate the loss function using the validation dataset; perform iterative training until the loss function converges and is lower than the preset threshold, completing the training and solidifying the network weights.

[0026] Identify artificial powered targets and drifting targets: Load the architecture and network weights of the multi-channel deep convolutional neural network into the detection radar processor, use a sliding window with a window length of 20 points and a step of 1 point to extract motion features, and splice them to obtain a two-dimensional feature vector; input the two-dimensional feature vector into the multi-channel deep convolutional neural network, and output the recognition result. The identification results are sent to the radar system to determine the threat and generate the situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a process flow chart. Figure 2 It is a diagram of the architecture of a multi-channel deep convolutional neural network. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0029] The process of the present invention is as follows Figure 1 As shown, a dual-station high-frequency radar is used to invert the sea surface vector flow field and wind field, and the flow field and wind field are quality controlled and interpolated with high precision; the dynamic equation is used to simulate the track of drifting targets, and radar detection is used to extract the track of artificial powered targets; the motion characteristics of the target are extracted from the track data, and a motion feature dataset is constructed; the architecture of a multi-channel deep convolutional neural network is designed, and the network weights are trained to identify artificial powered targets and drifting targets.

[0030] The dual-station high-frequency radar inverts the sea surface vector flow field: extracts radial flow, synthesizes the vector flow field, and outputs an M*N two-dimensional vector flow matrix. The dual-station high-frequency radar inverts the sea surface vector wind field: inverts wind direction and wind speed, and outputs an M*N two-dimensional vector wind matrix; where M and N are the number of resolution units in longitude and latitude, respectively.

[0031] Perform quality control on the flow and wind fields: Utilize the continuity of ocean surface flow in space and perform sliding window processing on the latitude and longitude coordinate system. The window size is set to 5 resolution units * 5 resolution units and the step is set to 1 resolution unit. Let the sliding window two-dimensional velocity matrix before quality control be R v0 , the two-dimensional flow matrix is R θ0 , calculate the mean flow velocity within the window Flow direction mean Flow rate standard deviation σ v , flow direction standard deviation σ θ ; If the flow rate of a resolution unit v i Dissatisfied Or flow to θ i Dissatisfied Then remove the vector flow of the resolution unit; let the sliding window two-dimensional velocity matrix after quality control be R v1 , the two-dimensional flow matrix is R θ1 ; Use the same method to control the quality of the wind field, and set the two-dimensional wind speed matrix as R w1 , the two-dimensional wind direction matrix is R d1 .

[0032] High-precision interpolation of flow and wind fields: Calculation of velocity matrix R v1 The mean use Instead of R v1 The blank cells in the matrix R are obtained v2 ; R v2 Singular value decomposition, let R v2 =UΛS T , where S T represents the transpose of the S matrix, Λ=diag[λ1,λ2,λi ,…,λ5] is the singular value matrix, diag() represents the diagonal matrix, λ i represents singular values, U = [u1, ..., u5] and S = [s1, ..., s5] are EOF matrices, EOF is the empirical orthogonal function, u i and s i is the eigenvector, i ranges from 1 to 5; take the first q eigenvectors of U and S to reconstruct the matrix, let where q satisfies The minimum value of the constraint; use R v3 The corresponding unit in replaces R v1 The blank cells in the flow velocity are obtained by high-precision interpolation matrix R. v4 ; Using the same method, high-precision interpolation is performed on the matrix of flow direction, wind speed, and wind direction to obtain the high-precision interpolation matrix R of flow direction θ4 , wind speed high-precision interpolation matrix R w4 , wind direction high-precision interpolation matrix R d4 .

[0033] Use dynamic equations to simulate the drift target trajectory: Based on the high-precision flow field, the dynamic equation is set as Where L(0) is the initial latitude and longitude of the target, V w (t) is the wind-induced floating speed, V c (t) is the ocean surface velocity, V k (t) is the floating speed generated by the Coriolis force, t is time, L(t) is the latitude and longitude of the target at time t; from the high-precision interpolation matrix R of wind speed w4 , wind direction high-precision interpolation matrix R d4 The wind speed and direction at 10m above the ocean surface are calculated and expressed as v0 and θ respectively. k (t) Orthogonal decomposition in longitude and latitude gives The initial longitude and latitude of the target are randomly set to L(0), and the longitude and latitude of the target at time t is obtained to be L(t), which is the drifting target track.

[0034] Radar detection is used to extract the track of artificial powered targets: including target detection, false alarm suppression, track initiation, tracking filtering, and track termination.

[0035] Extract the speed characteristics of the target: In the target track sequence, use the sliding window to select samples and calculate the speed characteristic vector; set the window length to 20 track points and the step to 1 track point; use the data in the window to construct the speed characteristic vector and time characteristic vector, and set the speed vector and time vector of the artificial power target to be V r =[v r1 ,v r2 ,v ri ,…,vr20 ]、T r =[t r1 ,t r2 ,t ri ,…,t r20 ],v ri , t ri They represent the speed value and time value of the artificial power target respectively. The speed vector and time vector of the drifting target are V p =[v p1 ,v p2 ,v pi ,…,v p20 ]、T p =[t p1 ,t p2 ,t pi ,…,t p20 ],v pi , t pi They represent the speed and time values of the drifting target respectively.

[0036] Extract the acceleration characteristics of the target: use the velocity characteristic vector to find the partial derivative of the time characteristic vector to calculate the acceleration characteristic vector. Suppose the acceleration vector of the artificial power target is The acceleration vector of the drifting target is

[0037] Extract the heading characteristics of the target: When using the sliding window to calculate the velocity feature vector, the heading feature vector is constructed using the samples in the window. Let the heading vectors of the artificial power target and the drifting target be θ respectively. r =[θ r1 ,θ r2 ,θ ri ,…,θ r20 ]、θ p =[θ p1 ,θ p2 ,θ pi ,…,θ p20 ],θ pi ,θ pi They represent the heading values of artificial powered targets and drifting targets respectively.

[0038] Extract the turning rate characteristics of the target: use the heading feature vector to find the partial derivative of the time feature vector to calculate the turning rate feature vector. Let the turning rate vector of the artificial power target be The turning rate vector of the drifting target is

[0039] Extract the curvature characteristics of the target track: When using the sliding window to calculate the velocity feature vector, use the samples in the window to construct the longitude and latitude vectors of the track, and use the latitude vector to calculate the second-order derivative and first-order derivative of the longitude vector, respectively. r ″、yr ′ represents the trajectory curvature vector of the artificial power target. Using the same method, the track curvature vector K of the drifting target is calculated p .

[0040] Constructing motion feature dataset: Splicing multiple motion features of artificial dynamic targets, assuming K r is 1*18 dimensional, A r and Z r is 1*19 dimensions, V r and θ r It is 1*20 dimensions, and the dimensions are inconsistent; select A r 、Z r 、V r ,θ r The first 1*18 dimensions of the eigenvector A′ are constructed r , Z′ r , V′ r ,θ′ r ; Vertically concatenate the characteristic vectors of velocity, acceleration, heading, turning rate, and track curvature to construct a 5*18 two-dimensional matrix Using the same method, construct the two-dimensional matrix M of multiple motion features of the drifting target p ; Sliding window processing of multiple tracks, constructing multiple M r and M p The datasets are merged into the target motion feature dataset M.

[0041] Design the architecture of a multi-channel deep convolutional neural network: use independent convolution kernels to perform the multi-channel convolution on the elements M in the dataset M. i Convolution is performed on each row of , and the deep features formed by convolution are spliced and fused in the fully connected layer to identify the target type; Figure 2 As shown, the multi-channel deep CNN has 8 layers, including input layer, convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, fully connected layer 1, fully connected layer 2, and output layer; the size of the input layer is 5*18, and the two-dimensional feature vector M is connected. i ; The output layer has a vector size of 2*1 Among them, p1 and p2 are the confidence of artificial power targets and drifting targets respectively, and the sum of the two is 1; the loss function adopts the cross entropy loss function.

[0042] Training network weights: Collect 100 tracks of artificial powered targets and drifting targets, and extract 2000 two-dimensional motion feature matrices M for each. i, forming a dataset M of 4000 samples; randomly select 3500 samples as the training dataset and the remaining 500 samples as the validation dataset; input the two-dimensional feature matrix in the training dataset into the multi-channel deep CNN network one by one, calculate the weights, and calculate the loss function using the validation dataset; perform iterative training until the loss function converges and is lower than the preset threshold, completing the training and solidifying the network weights.

[0043] Identify artificial powered targets and drifting targets: Load the architecture and network weights of the multi-channel deep convolutional neural network into the detection radar processor, use a sliding window with a window length of 20 points and a step of 1 point to extract motion features, and splice them to obtain a two-dimensional feature vector; input the two-dimensional feature vector into the multi-channel deep convolutional neural network, and output the recognition result. The identification results are sent to the radar system to determine the threat and generate the situation.

[0044] The above are embodiments of the present invention and do not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention are included in the protection scope of the present invention.

Claims

1. A sea surface target classification method based on multi-motion feature fusion, characterized in that: include: A dual-station high-frequency radar is used to invert the sea surface vector flow field and wind field, and the flow field and wind field are quality controlled and interpolated with high precision; dynamic equations are used to simulate the track of drifting targets, and radar detection is used to extract the track of artificially powered targets, including target detection, false alarm suppression, track initiation, tracking filtering, and track termination; the motion characteristics of the target are extracted from the track data, including speed characteristics, acceleration characteristics, heading characteristics, turning rate characteristics, and track curvature characteristics, and a motion feature dataset is constructed; the architecture of a multi-channel deep convolutional neural network is designed, and the network weights are trained to identify artificially powered targets and drifting targets.

2. The sea surface target classification method based on multi-motion feature fusion according to claim 1 is characterized in that: The dual-station high-frequency radar inverts the sea surface vector flow field, including: extracting radial flow, synthesizing the vector flow field, and outputting an M*N two-dimensional vector flow matrix; the dual-station high-frequency radar inverts the sea surface vector wind field: inverting wind direction and wind speed, and outputting an M*N two-dimensional vector wind matrix; where M and N are the number of resolution units in longitude and latitude, respectively.

3. The sea surface target classification method based on multi-motion feature fusion according to claim 1 is characterized in that: The quality control of the flow field and wind field includes: utilizing the continuity of the ocean surface flow in space, performing sliding window processing on the latitude and longitude coordinate system, setting the window size to 5 resolution units * 5 resolution units, and setting the step to 1 resolution unit; assuming that the sliding window two-dimensional velocity matrix before quality control is R v0 , the two-dimensional flow matrix is R θ0 , calculate the mean flow velocity within the window Flow direction mean Flow rate standard deviation σ v , flow direction standard deviation σ θ ; If the flow rate of a resolution unit v i Dissatisfied Or flow to θ i Dissatisfied Then remove the vector flow of the resolution unit; let the sliding window two-dimensional velocity matrix after quality control be R v1 , the two-dimensional flow matrix is R θ1 ; Use the same method to control the quality of the wind field, and set the two-dimensional wind speed matrix as R w1 , the two-dimensional wind direction matrix is R d1 .

4. The sea surface target classification method based on multi-motion feature fusion according to claim 3 is characterized in that: The high-precision interpolation of the flow field and wind field includes: calculating the velocity matrix R v1 The mean use Instead of R v1 The blank cells in the matrix R are obtained v2 ; R v2 Singular value decomposition, let R v2 =UΛS T , where S T represents the transpose of the S matrix, Λ=diag[λ1,λ2,λ i ,…,λ5] is the singular value matrix, diag() represents the diagonal matrix, λ i represents singular values, U = [u1, ..., u5] and S = [s1, ..., s5] are EOF matrices, EOF is the empirical orthogonal function, u i and s i is the eigenvector, i ranges from 1 to 5; take the first q eigenvectors of U and S to reconstruct the matrix, let where q satisfies The minimum value of the constraint; use R v3 The corresponding unit in replaces R v1 The blank cells in the flow velocity are obtained by high-precision interpolation matrix R. v4 ; Using the same method, high-precision interpolation is performed on the matrix of flow direction, wind speed, and wind direction to obtain the high-precision interpolation matrix R of flow direction θ4 , wind speed high-precision interpolation matrix R w4 , wind direction high-precision interpolation matrix R d4 .

5. The sea surface target classification method based on multi-motion feature fusion according to claim 4 is characterized in that: The method of simulating the drift target track by using the dynamic equation includes: based on the high-precision flow field, assuming the dynamic equation is Where L(0) is the initial latitude and longitude of the target, V w (t) is the wind-induced floating speed, V c (t) is the ocean surface velocity, V k (t) is the floating speed generated by the Coriolis force, t is time, L(t) is the latitude and longitude of the target at time t; from the high-precision interpolation matrix R of wind speed w4 , wind direction high-precision interpolation matrix R d4 The wind speed and direction at 10m above the ocean surface are calculated and expressed as v0 and θ respectively. k (t) Orthogonal decomposition in longitude and latitude gives The initial longitude and latitude of the target are randomly set to L(0), and the longitude and latitude of the target at time t is obtained to be L(t), which is the drifting target track.

6. The sea surface target classification method based on multi-motion feature fusion according to claim 1 is characterized in that: The method for extracting the speed characteristics of the target includes: selecting samples in the target track sequence using a sliding window to calculate the speed characteristic vector; setting the window length to 20 track points and the step to 1 track point; constructing the speed characteristic vector and time characteristic vector using the data in the window, and setting the speed vector and time vector of the artificial power target to be V respectively. r =[v r1 ,v r2 ,v ri ,…,v r20 ]、T r =[t r1 ,t r2 ,t ri ,…,t r20 ],v ri , t ri They represent the speed value and time value of the artificial power target respectively. The speed vector and time vector of the drifting target are V p =[v p1 ,v p2 ,v pi ,…,v p20 ]、T p =[t p1 ,t p2 ,t pi ,…,t p20 ],v pi , t pi Respectively represent the speed value and time value of the drifting target; the said extraction of the acceleration characteristics of the target includes: using the speed characteristic vector to find the partial derivative of the time characteristic vector to calculate the acceleration characteristic vector, assuming that the acceleration vector of the artificial power target is The acceleration vector of the drifting target is The extraction of the heading feature of the target includes: when calculating the velocity feature vector using a sliding window, constructing the heading feature vector using samples in the window, assuming that the heading vectors of the artificial power target and the drifting target are θ r =[θ r1 ,θ r2 ,θ ri ,…,θ r20 ]、θ p =[θ p1 ,θ p2 ,θ pi ,…,θ p20 ],θ pi ,θ pi Respectively represent the heading values of the artificial power target and the drifting target; the extraction of the turning rate feature of the target includes: using the heading feature vector to find the partial derivative of the time feature vector to calculate the turning rate feature vector, and assuming that the turning rate vector of the artificial power target is The turning rate vector of the drifting target is The method of extracting the curvature characteristics of the target track includes: when calculating the velocity characteristic vector using a sliding window, constructing the longitude vector and latitude vector of the track using the samples in the window, and using the latitude vector to calculate the second-order derivative and the first-order derivative of the longitude vector, respectively. r ″、y r ′ represents the trajectory curvature vector of the artificial power target. Using the same method, the track curvature vector K of the drifting target is calculated p .

7. The sea surface target classification method based on multi-motion feature fusion according to claim 6 is characterized in that: The construction of the motion feature dataset includes: splicing multiple motion features of the artificial power target, assuming K r is 1*18 dimension, A r and Z r is 1*19 dimensions, V r and θ r It is 1*20 dimensions, and the dimensions are inconsistent; select A r 、Z r 、V r ,θ r The first 1*18 dimensions of the eigenvector A′ are constructed r , Z′ r , V′ r ,θ′ r ; Vertically concatenate the characteristic vectors of velocity, acceleration, heading, turning rate, and track curvature to construct a 5*18 two-dimensional matrix Using the same method, construct the two-dimensional matrix M of multiple motion features of the drifting target p ; Sliding window processing of multiple tracks, constructing multiple M r and M p The datasets are merged into the target motion feature dataset M.

8. The sea surface target classification method based on multi-motion feature fusion according to claim 7 is characterized in that: The method includes: designing a multi-channel deep convolutional neural network architecture: using independent convolution kernels to perform the convolution on the elements M in the dataset M respectively. i Convolution is performed on each row of the convolution layer, and the deep features formed by the convolution are spliced and fused in the fully connected layer to identify the target type; as shown in Figure 2, the multi-channel deep CNN has 8 layers, including input layer, convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, fully connected layer 1, fully connected layer 2, and output layer; the size of the input layer is 5*18, and the two-dimensional feature vector M is connected. i ; The output layer has a vector size of 2*1 Among them, p1 and p2 are the confidence of artificial power targets and drifting targets respectively, and the sum of the two is 1; the loss function adopts the cross entropy loss function.

9. The method for sea surface target classification based on multi-motion feature fusion according to claim 8, characterized in that: The training network weights include: collecting 100 tracks of artificial power targets and drifting targets, extracting 2000 two-dimensional motion feature matrices M respectively i , forming a dataset M of 4000 samples; randomly select 3500 samples as the training dataset and the remaining 500 samples as the validation dataset; input the two-dimensional feature matrix in the training dataset into the multi-channel deep CNN network one by one, calculate the weights, and calculate the loss function using the validation dataset; perform iterative training until the loss function converges and is lower than the preset threshold, completing the training and solidifying the network weights.

10. The sea surface target classification method based on multi-motion feature fusion according to claim 1, characterized in that: The method includes: identifying artificial powered targets and drifting targets: loading the architecture and network weights of a multi-channel deep convolutional neural network into the processor of the detection radar, extracting motion features using a sliding window with a window length of 20 points and a step of 1 point, and splicing to obtain a two-dimensional feature vector; inputting the two-dimensional feature vector into the multi-channel deep convolutional neural network and outputting the recognition result. The identification results are sent to the radar system to determine the threat and generate the situation.

Images