Sea surface small target tracking method fused with sea wave dynamic prediction

Abstract

The invention discloses a sea surface small target tracking method fused with sea wave dynamic prediction. The sea surface small target tracking method comprises the following steps: selecting an Elfouhail sea wave spectrum model to construct a motion model of a sea surface small target; parameters of the motion model are set, and sea surface small target track data sets of the motion model in different sea conditions are obtained through simulation; preprocessing the data; inputting the preprocessed data into the constructed DA-RNN time sequence prediction network, and predicting the sea wave height at the next moment; and inputting the predicted sea wave height into an EKF filtering algorithm module to estimate the target position. The sea wave height is predicted by using a deep learning network, and the predicted sea wave height is input into a traditional filtering algorithm, so that sea surface small target tracking fused with sea wave dynamic prediction is realized.

Description

Technical Field

[0001] The invention belongs to the field of marine engineering, relates to small target tracking on the sea surface, and specifically relates to a small target tracking method on the sea surface integrating dynamic prediction of sea waves. Background Art

[0002] Small targets on the sea surface, such as unmanned boats and ships, have developed rapidly in recent years, but they also bring huge challenges to maritime traffic control and maritime law enforcement. The sea surface often encounters complex sea conditions such as strong winds and waves, and the motion state of these small targets on the sea surface is greatly affected by the ups and downs of the waves. How to obtain a stable and accurate target motion state in a complex environment with ups and downs of the sea is a major problem that needs to be solved in the field of small target tracking on the sea surface.

[0003] In the past few decades, many tracking filter algorithms have been proposed, such as Kalman filter (KF), extended Kalman filter (EKF), unscented Kalman filter (UKF) and particle filter (PF). These algorithms all need to use predefined specific dynamic models to simulate the motion state of the target, such as uniform velocity model (CV), uniform acceleration model (CA), random motion model (RM), uniform turning model (CT), etc. The tracking performance of the target depends largely on these predefined dynamic models. When the motion state of the target matches the predetermined dynamic model, the tracking system can obtain ideal tracking performance. Otherwise, the tracking performance is poor and even loss of tracking may occur. Moreover, in practice, due to the uncertainty of the target motion state in maneuvering scenarios, a single model may not accurately describe the actual motion state of the target. To address this problem, Yaakov Bar-Shalom et al. proposed a maneuvering target tracking algorithm based on multiple models (MM). In the multiple model tracking algorithm, each model represents a possible motion mode, and the tracking results of different models are weighted combined to improve the tracking effect. The IMM algorithm is an extension of the MM algorithm and introduces a probability transition matrix. Through the probability transition matrix, the system can adaptively adjust the weights of each model according to the observed data between different time steps, which makes the model more flexible in tracking maneuvering targets.

[0004] However, due to the influence of the ups and downs of the sea waves, the motion state of small targets on the sea surface has great uncertainty. Moreover, the ups and downs of the waves themselves are highly random and difficult to describe using a specific mathematical model. Therefore, it is difficult to achieve stable tracking of small targets on the sea surface through a specific state model or the weighting of different models. Summary of the invention

[0005] Purpose of the invention: In order to overcome the deficiencies in the prior art, a method for tracking small targets on the sea surface that integrates dynamic prediction of waves is provided, which uses a deep learning network to predict the wave height, and inputs the predicted wave height into a traditional filtering algorithm, thereby realizing tracking of small targets on the sea surface that integrates dynamic prediction of waves.

[0006] Technical solution: To achieve the above-mentioned purpose, the present invention provides a method for tracking small targets on the sea surface integrating dynamic prediction of sea waves, comprising the following steps:

[0007] S1: Select the Elfouhaily wave spectrum model to construct the motion model of small targets on the sea surface;

[0008] S2: Setting the parameters of the motion model, and obtaining the sea surface small target track data set of the motion model under different sea conditions through simulation;

[0009] S3: preprocessing the data of step S2;

[0010] S4: Input the preprocessed data into the constructed DA-RNN time series prediction network to predict the wave height at the next moment;

[0011] S5: Input the predicted wave height into the EKF filter algorithm module to estimate the target position.

[0012] Furthermore, in step S2, the Elfouhaily wave spectrum model is used to simulate the wave fluctuations under sea conditions of levels II, IV, V, and VI, and the CV model and the CT model are used to simulate the movement state of small targets on the sea surface, and the trajectory points of the small targets on the sea surface moving with the fluctuations of the waves are collected to establish a data set of target trajectories.

[0013] Furthermore, in step S2, a trajectory generator is designed based on the state space model to simulate segments of different maneuvering target trajectories, and the generated trajectories are regarded as samples in the data set to ensure that they can match the actual tracking scenario.

[0014] Furthermore, in step S4, the DA-RNN time series prediction network is constructed based on a two-stage attention mechanism. The DA-RNN time series prediction network includes an encoder and a decoder. The application of the two-stage attention mechanism includes:

[0015] A novel attention mechanism is used in the encoder to adaptively select relevant driving sequences;

[0016] In the decoder, a temporal attention mechanism is used to automatically select relevant encoder hidden states at all time steps.

[0017] Furthermore, the operation mode of the DA-RNN time series prediction network includes:

[0018] Given n driving sequences, that is, X=(x 1 ,x 2 ,...,x n )=(x 1 ,x 2 ,...,x T )∈ n×T , T is the length of the window size, represents a driving sequence of length T, n is the width of the driving sequence, The vector representing the n driving input sequences at time t, i.e., the influencing factors of the prediction of the wave height; the existing method only uses the sequence of the past period of time to predict the sequence state at the next moment, and the prediction accuracy is poor; the present invention considers other factors that affect the prediction performance of the wave height, including the wave height, running direction and running speed of the previous 10 moments.

[0019] The constructed prediction network needs to learn from (y 1 ,y 2 ,...,y T-1 )∈,(x 1 ,x 2 ,...,x T )∈ to The mapping relationship is

[0020]

[0021] Furthermore, the operation of the attention mechanism in the encoder includes:

[0022] The core of the encoder is the LSTM unit, which encodes the input sequence into a feature representation; the encoder can learn a t to h t The nonlinear activation function f 1 :

[0023] h t =f 1 (h t-1 ,x t )(2)

[0024] The LSTM unit has a hidden state and a memory cell at each time step. Access to the memory cell is controlled by three S-shaped gates: the forget gate f t , input gate i t and output gate o t ;

[0025] The unit status is updated as follows:

[0026]

[0027] Among them, [h t-1 ;x t ]∈ m+n is the previous hidden state h t-1 With the current input x t The series connection of W f ,W i ,W o ,W s ∈ m×(m+n) , b f ,b f ,b f ,b f ∈ m are the parameters that need to be learned; σ and are the Logistic activation function and element multiplication respectively;

[0028] As the length of the input sequence increases, the performance of the encoder-decoder network will decrease, so the present invention uses an encoder based on the input attention mechanism to adaptively select related sequences; assuming that the kth input driving sequence is given You can refer to the previous hidden state h in the encoder LSTM unit t-1 and the cell state s t-1 , construct an attention mechanism by determining the attention model:

[0029]

[0030] Among them, v e ∈ T , W e ∈ T×2m , U e ∈ T×T is the parameter that needs to be learned, ignoring the bias term in the equation; Use the Softmax function to ensure that the sum of all attention weights is 1;

[0031]

[0032] That is, the kth input feature, the final attention weight at time t, and then the weight is used to extract the driving sequence:

[0033]

[0034] The driving sequence at this time becomes x t , x t Substitute into update formula (3), and the hidden state is updated as follows:

[0035] h t =f(h t-1 ,xt )(7)

[0036] Among them, f is the nonlinear activation function learned by the encoder. At this time, x t Update, replacing x in the equation t , at this time, the adaptive selection drive sequence is realized.

[0037] Furthermore, the operation of the temporal attention mechanism in the decoder includes:

[0038] The decoder uses an LSTM network to predict the output Similar to the encoder, the decoder has the same problem that the network performance will decrease as the sequence length increases. Therefore, the present invention uses a temporal attention mechanism to select hidden states in the decoder; based on the hidden state m of the previous time step of the decoder t-1 ∈ p And the hidden state s' of the LSTM unit at the previous moment t-1 ∈ p , calculate the weights of the encoder hidden state:

[0039]

[0040] Among them, [m t-1 ;s' t-1 ]∈ 2p is the concatenation of the previous hidden state and the cell state of the LSTM cell, v d ∈ m ,W d ∈ m×2p ,U d ∈ m×m are the parameters that need to be learned. Use the Softmax function to ensure that the sum of all attention weights is 1:

[0041]

[0042] Attention Weight Indicates the importance of the i-th encoder state to the prediction; the context vector c is added through the attention mechanism t Represented as all encoder hidden states}h 1 ,h 2 ,,h T},

[0043]

[0044] After obtaining the weighted vector, update the target sequence:

[0045] y t-1 =w T [yt-1 ;c t-1 ]+b(11)

[0046] Among them, [y t-1 ;c t-1 ]∈ m+1 is the decoder input y t-1 With the calculated context vector c t-1 The series connection of parameters w∈ m+1 and b∈ will be jointly mapped to the size of the decoder input; then, the newly calculated y t-1 Update to decode status

[0047] m t =f j (m t-1 ,y t-1 )(12) Where m t Update via:

[0048]

[0049] Among them, W′ 4 ,W′ i ,W′ o ,W′ s ∈ p×(p+1) , b′ f ,b′ i ,b′ o ,b′ s ∈ p is the parameter to be learned; σ and represent the Logistic activation function and element multiplication respectively; In general, for the nonlinear autoregressive model (NARX), the purpose of the present invention is to use the designed network to infinitely approach the function F to obtain an estimate of the current output,

[0050]

[0051] Among them, [m T ;c T ]∈ p+m is the concatenation of the decoder hidden state and the context vector; parameter W y ∈ p×(p+m) and b w ∈ p which are jointly mapped to the size of the decoder hidden state;

[0052] Finally, the weight v y ∈ p , deviation b v ∈ generates the prediction result.

[0053] Furthermore, in step S4, the DA-RNN time series prediction network is trained according to the root mean square error loss, and the root mean square error loss of the network training is defined as:

[0054]

[0055] Among them, N is the number of training samples; in the process of error back propagation, the stochastic gradient descent method is used to train the LSTM network to minimize the root mean square error loss.

[0056] Furthermore, in step S5, the target performs uniform linear motion or uniform turning motion in the motion scene, and the target state for estimating the target position includes position information and speed information, and the motion equation is as follows:

[0057] x t =Fx t-1 +u t +w k (16)

[0058] Among them, the state quantity x t Defined as [d x,t ,v x,t ,d y,t ,v y,t ,d z,t ],[d x,t ,d y,t ,d z,t ] T is the target position, [v x,t ,v y,t ] T is the corresponding speed; F is the state transfer matrix, u t is the wave rise and fall, defined as u t =[0,0,0,0,d z,t ] T , wave height d z,t It is necessary to extract the wave space based on the wave spectrum simulation; k is the process noise caused by environmental factors, w k ~N(0,Q k ), where Q k is the covariance matrix;

[0059] The observation equation is as follows:

[0060] z t =h(x t )+v k (17)

[0061] Among them, z t represents the observation vector, defined as [r t ,θ t, ρ t ], containing the distance value r t , azimuth angle θ t and radial velocity ρ t , h() is the nonlinear observation function, v k is the observation noise.

[0062] In the present invention, different sea conditions are simulated according to the Elfouhaily wave spectrum, and a database of the motion trajectory of small targets on the sea surface is established. Then, the wave height is predicted using an LSTM network based on a dual-stage attention mechanism, and the predicted height is input into a state transfer model, and the extended Kalman filter method is used for tracking.

[0063] In the present invention, the problem of predicting the wave height is characterized as a time series prediction problem. A recurrent neural network based on a two-stage attention mechanism is used to predict the wave height in real time and compensate it to the motion state of the target. The extended Kalman filter algorithm is used to track the target in three-dimensional space. The motion models are a uniform linear motion model (CV) and a uniform turning motion model (CT).

[0064] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0065] 1. Taking the fluctuation of sea waves into consideration in target tracking alleviates the problem of low target tracking accuracy caused by the fluctuation of sea waves. Compared with the existing filtering method without sea wave prediction, the method of the present invention can effectively improve the tracking accuracy of small targets on the sea surface and achieve stable and accurate tracking of small targets on the sea surface.

[0066] 2. The DA-RNN time series prediction network is used to predict wave height. The network combines the influence of heading, wind speed, wind direction and other factors on wave height to obtain more accurate wave height. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 1 is a flow chart for implementing the method of the present invention;

[0068] Figure 2 It is a partial trajectory diagram of the small target moving on the sea surface in the present invention;

[0069] Figure 3 3D simulation diagram of 8 tracks used for testing in the present invention;

[0070] Figure 4 It is a network structure diagram of the DA-RNN time series prediction network in the present invention;

[0071] Figure 5 The effect diagram of single vector time prediction and DA-RNN network prediction in the present invention;

[0072] Figure 6 This is a target trajectory tracking result diagram of the nonlinear model with and without wave prediction methods in the present invention. DETAILED DESCRIPTION

[0073] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0074] The present invention provides a method for tracking small targets on the sea surface by integrating dynamic prediction of sea waves. Figure 1 As shown, it includes the following steps:

[0075] S1: Select the Elfouhaily wave spectrum model to construct the motion model of small targets on the sea surface;

[0076] S2: Set the parameters of the motion model and obtain the sea surface small target track data set of the motion model under different sea conditions through simulation:

[0077] The Elfouhaily wave spectrum model is used to simulate the dynamic fluctuation of waves, simulating the wave fluctuations under sea conditions of II, IV, V and VI respectively.

[0078] The CV model and the CT model are used to simulate the motion state of small targets on the sea surface, and the trajectory points of small targets moving with the ups and downs of the waves are collected to establish a target trajectory data set.

[0079] Table 1 shows the parameter settings for each level of sea conditions during the simulation process, where u 10 is the wind speed at 10m above the sea surface, X 0 is the width of the sea area, and based on this, the uniform linear motion and uniform turning motion trajectory of the target on the sea surface are simulated.

[0080] Table 1 Parameter settings for various sea conditions

[0081]

[0082] The dataset contains trajectory samples under four different sea conditions, and the corresponding parameters of these samples are different, such as wind speed, wind direction, heading and azimuth. Since it is difficult to obtain enough real data of the trajectory of maneuvering targets in actual sea surface surveillance scenarios, the present invention designs a trajectory generator based on the state space model, which can simulate fragments of different maneuvering target trajectories. These generated trajectories are regarded as samples in the dataset to ensure that they can match the real tracking scenario. All parameters of trajectory generation are set according to the real scenario to ensure its representativeness and accuracy.

[0083] Through the trajectory generator, 720,000 samples were generated, and a dataset covering the motion trajectories of common maneuvering targets such as unmanned surface boats and ships was constructed. This dataset provides rich data support for further research and application.

[0084] S3: Preprocess the data of step S2; each piece of data includes the wave height, running direction, and running speed of the previous 10 moments, and each 10 pieces of data form a group of input data to predict the wave height at the next moment. Finally, the mean normalization method is used to process the data.

[0085] S4: Input the preprocessed data into the constructed DA-RNN time series prediction network to predict the wave height at the next moment:

[0086] The DA-RNN time series prediction network is built based on a two-stage attention mechanism. The DA-RNN time series prediction network includes an encoder and a decoder. The applications of the two-stage attention mechanism include:

[0087] A novel attention mechanism is used in the encoder to adaptively select relevant driving sequences;

[0088] In the decoder, a temporal attention mechanism is used to automatically select relevant encoder hidden states at all time steps.

[0089] The operation of the DA-RNN time series prediction network includes:

[0090] Given n driving sequences, that is, X=(x 1 ,x 2 ,...,x n )=(x 1 ,x 2 ,...,x T )∈ n×T , T is the length of the window size, represents a driving sequence of length T, n is the width of the driving sequence, The vector representing the n driving input sequences at time t, i.e., the influencing factors of the prediction of the wave height; the existing method only uses the sequence of the past period of time to predict the sequence state at the next moment, and the prediction accuracy is poor; the present invention considers other factors that affect the prediction performance of the wave height, including the wave height, running direction and running speed of the previous 10 moments.

[0091] The constructed prediction network needs to learn from (y 1 ,y 2 ,...,y T-1 )∈,(x 1 ,x 2 ,...,x T )∈ to The mapping relationship is

[0092]

[0093] The operation of the attention mechanism in the encoder includes:

[0094] The core of the encoder is the LSTM unit, which encodes the input sequence into a feature representation; the encoder can learn a t to h t The nonlinear activation function f 1 :

[0095] h t =f 1 (h t-1 ,x t )(2)

[0096] The LSTM unit has a hidden state and a memory cell at each time step. Access to the memory cell is controlled by three S-shaped gates: the forget gate f t , input gate i t and output gate o t ;

[0097] The unit status is updated as follows:

[0098]

[0099] Among them, [h t-1 ;x t ]∈ m+n is the previous hidden state h t-1 With the current input x t The series connection of W f ,W i ,W o ,W s ∈ m×(m+n) , b f ,b f ,b f ,b f ∈ m are the parameters that need to be learned; σ and are the Logistic activation function and element multiplication respectively;

[0100] As the length of the input sequence increases, the performance of the encoder-decoder network will decrease, so the present invention uses an encoder based on the input attention mechanism to adaptively select related sequences; assuming that the kth input driving sequence is given You can refer to the previous hidden state h in the encoder LSTM unit t-1 and the cell state s t-1 , construct an attention mechanism by determining the attention model:

[0101]

[0102] Among them, v e ∈ T , W e ∈ T×2m , U e ∈ T×T is the parameter that needs to be learned, ignoring the bias term in the equation; Use the Softmax function to ensure that the sum of all attention weights is 1;

[0103]

[0104] That is, the kth input feature, the final attention weight at time t, and then the weight is used to extract the driving sequence:

[0105]

[0106] The driving sequence at this time becomes x t , x t Substitute into update formula (3), the hidden state is updated as:

[0107] h t =f(h t-1 ,x t )(7)

[0108] Among them, f is the nonlinear activation function learned by the encoder. At this time, x t Update, replacing x in the equation t , at this time, the adaptive selection drive sequence is realized.

[0109] The operation of the temporal attention mechanism in the decoder includes:

[0110] The decoder uses an LSTM network to predict the output Similar to the encoder, the decoder has the same problem that the network performance will decrease as the sequence length increases. Therefore, the present invention uses a temporal attention mechanism to select hidden states in the decoder; based on the hidden state m of the previous time step of the decoder t-1 ∈ p And the hidden state s' of the LSTM unit at the previous moment t-1 ∈ p , calculate the weights of the encoder hidden state:

[0111]

[0112] Among them, [m t-1 ;s' t-1 ]∈ 2pis the concatenation of the previous hidden state and the cell state of the LSTM cell, v d ∈ m ,W d ∈ m×2p ,U d ∈ m×m are the parameters that need to be learned. Use the Softmax function to ensure that the sum of all attention weights is 1:

[0113]

[0114] Attention Weight Indicates the importance of the i-th encoder state to the prediction; the context vector c is added through the attention mechanism t Denoted as all encoder hidden states {h 1 ,h 2 ,,h T},

[0115]

[0116] After obtaining the weighted vector, update the target sequence:

[0117] y t-1 =w T [y t-1 ;c t-1 ]+b(11)

[0118] Among them, [y t-1 ;c t-1 ]∈ m+1 is the decoder input y t-1 With the calculated context vector c t-1 The series connection of parameters w∈ m+1 and b∈ will be jointly mapped to the size of the decoder input; then, the newly calculated y t-1 Update to decode status

[0119] m t =f j (m t-1 ,y t-1 )(12) Where m t Update via:

[0120]

[0121] Among them, W′ f ,W′ i ,W′ o ,W′ s ∈ p×(p+1) , b′ f ,b′ i,b′ o ,b′ s ∈p is a parameter to be learned; σ and represent the Logistic activation function and element multiplication respectively; In general, for the nonlinear autoregressive model (NARX), the purpose of the present invention is to use the designed network to infinitely approach the function F to obtain an estimate of the current output,

[0122]

[0123] Among them, [m T ;c T ]∈ p+m is the concatenation of the decoder hidden state and the context vector; parameter W y ∈ p×(p+m) and b w ∈ p which are jointly mapped to the size of the decoder hidden state;

[0124] Finally, the weight v y ∈ p , deviation b v ∈ generates the prediction result.

[0125] In the present invention, the DA-RNN time series prediction network is trained according to the root mean square error loss, and the root mean square error loss of network training is defined as:

[0126]

[0127] Among them, N is the number of training samples; in the process of error back propagation, the stochastic gradient descent method is used to train the LSTM network to minimize the root mean square error loss.

[0128] S5: Input the predicted wave height into the EKF filter algorithm module to estimate the target position:

[0129] The target performs uniform linear motion or uniform turning motion in the motion scene. The target state for estimating the target position includes position information and velocity information. The motion equation is as follows:

[0130] x t =Fx t-1 +u t +w k (16)

[0131] Among them, the state quantity x t Defined as [d x,t ,v x,t ,d y,t ,v y,t ,d z,t ],[d x,t ,d y,t,d z,t ] T is the target position, [v x,t ,v y,t ] T is the corresponding speed; F is the state transfer matrix, u t is the wave rise and fall, defined as u t =[0,0,0,0,d z,t ] T , wave height d z,t It is necessary to extract the wave space based on the wave spectrum simulation; k is the process noise caused by environmental factors, w k ~N(0,Q k ), where Q k is the covariance matrix;

[0132] The observation equation is as follows:

[0133] z t =h(x t )+v k (17)

[0134] Among them, z t represents the observation vector, defined as [r t ,θ t , ρ t ], containing the distance value r t , azimuth angle θ t and radial velocity ρ t , h() is the nonlinear observation function, v k is the observation noise.

[0135] In view of the wave height, the present invention improves the existing EKF filtering algorithm, adds the wave fluctuation to the original state equation, and uses the wave spectrum model to simulate the real motion trajectory of the target; in the prediction process, d z,t The input is the wave height predicted in step S4.

[0136] Based on the above content, the present invention is further described below in conjunction with specific drawings and example data:

[0137] like Figure 2 As shown, the present invention will simulate the motion state of a small target on the sea surface and establish a data set of the target motion trajectory. It is assumed that the small target on the sea surface makes uniform linear motion or uniform turning motion on the sea surface. Since the ship itself is designed with the ability to correct the course, the present invention ignores the influence of sea surface fluctuations on the target position and only considers the influence of waves on the target in the height dimension. Taking sea condition 4 as an example, the wind speed is 10m / s, the wind direction is 45°, and the effective wave height is set between 1.25m and 2.5m. Figure 2This is a schematic diagram of part of the trajectory of the target making uniform linear motion on the waves. The red dot is the target trajectory. It can be seen that the target rises and falls with the waves.

[0138] like Figure 3 As shown, the present invention simulates a sea area of ​​1500m×1500m, the x-axis is [-750m, 750m], the y-axis is [-750m, 750m], and the z-axis is the fluctuation of the wave height. The simulation time is 150s, a total of 300 frames, 0.5s / frame, and a total of 300 simulations per round. For the uniform linear motion model (CV), the initial position of the x-axis is randomly selected from [-750,-700], the initial position of the y-axis is randomly selected from [-750,-700], the speed in the x-axis direction is randomly selected from [0,10], and the speed in the y-axis direction is randomly selected between [0,10], simulating the 2nd, 4th, 5th, and 6th level sea conditions respectively. For the uniform turning model, ω is set to 0.01 to avoid the situation where small targets circle in place. To ensure that all points fall within the simulation area, the initial position of the x-axis is randomly selected between [-750, -700], the initial position of the y-axis is randomly selected between [-750, -700], the speed of the x-axis is randomly selected between [3, 6], and the speed of the y-axis is randomly selected between [1, 3]. To prevent the motion trajectory from falling outside the simulation area, the x-axis speed is larger than the y-axis speed. The time frame is 0.5s, and the sea surface is updated every 0.5s. Each sampling takes 300 frames to obtain the target data set. Figure 3 These are the eight motion trajectories of small targets on the sea surface under four sea conditions.

[0139] like Figure 4 As shown, the present invention models the prediction problem of wave height as a time series prediction problem, and adopts a recurrent neural network (DA-RNN) based on a dual-stage attention mechanism to predict wave height. The process is divided into two stages: 1) In the encoder, we use a new attention mechanism that can adaptively select relevant driving sequences. 2) In the decoder, a time attention mechanism is used to automatically select relevant encoder hidden states at all time steps. The present invention integrates these two attention mechanisms in a recurrent neural network (RNN) based on LSTM. Therefore, this time series prediction model can adaptively select the most influential factors as input features to capture the long-term time dependency of time series prediction. Figure 4 It is a diagrammatic illustration of the model of the present invention.

[0140] like Figure 5As shown in the figure, when training the wave height prediction model, the data is divided into two groups at a ratio of 7:3, the former is the training set and the latter is the test set. In the present invention, the prediction effect of the target motion under four sea conditions will be given. Finally, considering the XYZ three-dimensional plane for tracking, 8 test motion trajectories are designed, all within the simulation area. Among them, the maneuvering turning rate of the uniform turning model is 0.01. Figure 3 The target motion trajectory under the disturbance of sea wave fluctuations is shown, which are the trajectories of the small target in uniform motion and uniform turning motion under sea conditions II, IV, V, and VI. Each trajectory lasts for 150 seconds. In this invention, these 8 motion trajectory examples are used to test the performance of the extended Kalman filter tracking algorithm integrated with the wave height prediction.

[0141] In order to verify the effectiveness of the wave height prediction network used in the present invention, it is compared with the single vector time prediction. The single vector time prediction is an LSTM network for time series prediction, which includes an LSTM layer and a linear layer. The LSTM layer is used to process sequence data, and the linear layer is used to predict the output. The single vector prediction uses the wave height of the first 12 moments as the driving sequence and the wave height of the 13th moment as the target sequence. Taking sea condition 4 as an example, Figure 5 The left picture is the prediction effect picture of the single vector time prediction network. The blue line is the real wave height, and the orange line is the predicted wave height. The input is 12 real data, and the wave height of the 13th to 300th time step is predicted. It can be clearly seen from the figure that the comparison network only predicts the trend of change, and the predicted height is far from the real height, and the prediction effect is not good. The LSTM network based on the time attention mechanism proposed in this paper is a multi-vector prediction. The driving sequence is not only the wave height of the previous period, but also the running direction, wind speed, and wind direction. The right picture is the prediction picture of the network used in this paper, which predicts the wave height of the 11th to 300th time step. Taking the level 4 sea condition as an example, it can be seen that the prediction effect is much better than the single vector prediction.

[0142] In order to prove the superiority of the prediction network provided by the present invention, two networks were used to predict 8 targets under different sea conditions. Table 2 is a comparison of the prediction effects. The evaluation index of time series prediction used here is the root mean square error (RMSE). It can be seen that the prediction effect of the DA-RNN network used in the present invention is better than that of the LSTM network as a whole, and the lower the sea condition level and the smaller the wave height, the better the wave prediction effect.

[0143] Table 2 Prediction results (RMSE) of 8 targets under different sea conditions

[0144]

[0145] like Figure 6As shown in Figure 1, after obtaining the predicted information of the wave height, the predicted wave height is fed back into the state transfer equation, and the EKF algorithm is used for target tracking. This experiment uses nonlinear sensors to obtain observation values. The radar sensor is used as the observation sensor device of this experiment. The radar sensor is set to be located at (0,0,0)m in the scene. The observation equation is shown in formula (14), where the observation noise v k The mean is 0 and the variance is R k = diag(100,0.000005,5) Gaussian white noise, the experiment lasts for 150s, the number of Monte Carlo iterations is 200, targets 1, 2, 3, and 4 make uniform linear motion, and targets 5, 6, 7, and 8 make uniform turning motion. Figure 6 The target trajectory tracking results of the nonlinear model with and without wave prediction are shown. The blue line is the tracking trajectory of the method without wave prediction. The wave height is set to a constant value h=1. The figure shows that the tracking has deviated from the true motion trajectory in the later stage. It can be concluded that the target trajectory curve tracked by the algorithm proposed in the present invention fits the actual motion trajectory of the target well, and the deviation between the target state estimation and the true trajectory is minimal.

[0146] Table 3 compares the tracking accuracy of 8 targets under different sea conditions and models. When all parameters are the same, the tracking accuracy with wave prediction is significantly better than that without wave prediction. Experiments show that the sea surface small target tracking algorithm designed by the present invention and integrating wave prediction has improved the tracking accuracy by about 14% compared with the existing tracking algorithm, proving that the algorithm proposed by the present invention can effectively solve the interference problem of wave fluctuations on target tracking under complex sea conditions.

[0147] Table 3 Improvement of tracking accuracy of 8 targets in target tracking algorithm integrating wave prediction

[0148]

[0149]

Claims

1. A method for tracking small targets on the sea surface integrating dynamic prediction of sea waves, characterized in that: The steps include: S1: Select the Elfouhaily wave spectrum model to construct the motion model of small targets on the sea surface; S2: Setting the parameters of the motion model, and obtaining the sea surface small target track data set of the motion model under different sea conditions through simulation; S3: preprocessing the data of step S2; S4: Input the preprocessed data into the constructed DA-RNN time series prediction network to predict the wave height at the next moment; S5: Input the predicted wave height into the EKF filter algorithm module to estimate the target position.

2. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 1, characterized in that: In step S2, the Elfouhaily wave spectrum model is used to simulate the wave fluctuations under sea conditions of levels II, IV, V, and VI, and the CV model and the CT model are used to simulate the movement state of small targets on the sea surface. The trajectory points of the small targets on the sea surface moving with the fluctuations of the waves are collected to establish a data set of target trajectories.

3. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 1, characterized in that: In step S2, a trajectory generator is designed based on the state space model to simulate segments of different maneuvering target trajectories, and the generated trajectories are regarded as samples in the data set to ensure that they can match the actual tracking scene.

4. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 1, characterized in that: The DA-RNN time series prediction network in step S4 is constructed based on a two-stage attention mechanism. The DA-RNN time series prediction network includes an encoder and a decoder. The application of the two-stage attention mechanism includes: A novel attention mechanism is used in the encoder to adaptively select relevant driving sequences; In the decoder, a temporal attention mechanism is used to automatically select relevant encoder hidden states at all time steps.

5. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 4, characterized in that: The operation mode of the DA-RNN time series prediction network includes: Given n driving sequences, that is, X=(x 1 ,x 2 ,...,x n )=(x1,x2,...,x T )∈ n×T , T is the length of the window size, represents a driving sequence of length T, n is the width of the driving sequence, represents the vector of n driving input sequences at time t, i.e., the influencing factors of the prediction of wave height; The constructed prediction network needs to learn from (y1,y2,...,y T-1 )∈,(x1,x2,...,x T )∈ to The mapping relationship is 6. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 5, characterized in that: The operation of the attention mechanism in the encoder includes: The core of the encoder is the LSTM unit, which encodes the input sequence into a feature representation; the encoder can learn a t to h t The nonlinear activation function f1 is: h t =f1(h t-1 ,x t )(2) The LSTM unit has a hidden state and a memory cell at each time step. Access to the memory cell is controlled by three S-shaped gates: the forget gate f t , input gate i t and output gate o t ; The unit status is updated as follows: Among them, [h t-1 ;x t ]∈ m+n is the previous hidden state h t-1 With the current input x t The series connection of f ,W i ,W o ,W s ∈ m×(m+n) , b f ,b f ,b f ,b f ∈ m are the parameters that need to be learned; σ and are the Logistic activation function and element multiplication respectively; An encoder based on input attention mechanism is used to adaptively select relevant sequences; assuming that given the kth input driving sequence You can refer to the previous hidden state h in the encoder LSTM unit t-1 and the cell state s t-1 , construct an attention mechanism by determining the attention model: Among them, v e ∈ T , W e ∈ T×2m , U e ∈ T×T is the parameter that needs to be learned, ignoring the bias term in the equation; Use the Softmax function to ensure that the sum of all attention weights is 1; That is, the kth input feature, the final attention weight at time t, and then the weight is used to extract the driving sequence: The driving sequence at this time becomes x t , x t Substitute into update formula (3), the hidden state is updated as: h t =f(h t-1 ,x t )(7) Where f is the nonlinear activation function learned by the encoder. t Update, replacing x in the equation t , at this time, the adaptive selection drive sequence is realized.

7. The method for tracking small targets on the sea surface integrating dynamic prediction of sea waves according to claim 6, characterized in that: The operation of the temporal attention mechanism in the decoder includes: The decoder uses an LSTM network to predict the output The decoder uses a temporal attention mechanism to select hidden states; based on the hidden state m of the decoder in the previous time step t-1 ∈ p And the hidden state s' of the LSTM unit at the previous moment t-1 ∈ p , calculate the weights of the encoder hidden state: Among them, [m t-1 ;s' t-1 ]∈ 2p is the concatenation of the previous hidden state and the cell state of the LSTM cell, v d ∈ m ,W d ∈ m×2p ,U d ∈ m×m are the parameters that need to be learned. Use the Softmax function to ensure that the sum of all attention weights is 1: Attention Weight Indicates the importance of the i-th encoder state to the prediction; the context vector c is added through the attention mechanism t Represented as all encoder hidden states {h1,h2,,h T }, After obtaining the weighted vector, update the target sequence: y t-1 =w T [y t-1 ;c t-1 ]+b (11) Among them, [y t-1 ;c t-1 ]∈ m+1 is the decoder input y t-1 With the calculated context vector c t-1 The series connection of parameters w∈ m+1 and b∈ will be jointly mapped to the size of the decoder input; then, the newly calculated y t-1 Update to decode status m t =f j (m t-1 ,y t-1 ) (12) Among them, m t Update via: Among them, W' f ,W' i ,W' o ,W' s ∈ p×(p+1) , b' f ,b' i ,b' o ,b' s ∈ p are the parameters that need to be learned; σ and represent the Logistic activation function and element multiplication respectively; use the designed network to infinitely approach the function F to obtain an estimate of the current output, Among them, [m T ;c T ]∈ p+m is the concatenation of the decoder hidden state and the context vector; parameter W y ∈ p×(p+m) and b w ∈ p which are jointly mapped to the size of the decoder hidden state; Finally, the weight v y ∈ p , deviation b v ∈ generates the prediction result.

8. The method for tracking small targets on the sea surface integrating dynamic prediction of ocean waves according to claim 1, characterized in that: In step S4, the DA-RNN time series prediction network is trained according to the root mean square error loss, and the root mean square error loss of the network training is defined as: Among them, N is the number of training samples; in the process of error back propagation, the stochastic gradient descent method is used to train the LSTM network to minimize the root mean square error loss.

9. The method for tracking small targets on the sea surface integrating dynamic prediction of sea waves according to claim 1, characterized in that: The target state for estimating the target position in step S5 includes position information and velocity information, and the motion equation is as follows: x t =Fx t-1 +u t +w k (16) Wherein, the state quantity x t Defined as [d x,t ,v x,t ,d y,t ,v y,t ,d z,t ],[d x,t ,d y,t ,d z,t ] T is the target position, [v x,t ,v y,t ] T is the corresponding speed; F is the state transfer matrix, u t is the wave rise and fall, defined as u t =[0,0,0,0,d z,t ] T , wave height d z,t It is necessary to extract the wave space based on the wave spectrum simulation; k is the process noise caused by environmental factors, w k ~N(0,Q k ), where Q k is the covariance matrix; The observation equation is as follows: z t =h(x t )+v k (17) Among them, z t represents the observation vector, defined as [r t ,θ t , ρ t ], containing the distance value r t , azimuth angle θ t and radial velocity ρ t , h() is the nonlinear observation function, v k is the observation noise.

Images